In this paper we investigate several state-of-the-art machine-learning methods for automated classification of clustered microcalcifications (MCs), aimed to assisting radiologists for more accurate diagnosis of breast cancer in a computer-aided diagnosis (CADx) scheme. The methods we consider include: support vector machine (SVM), kernel Fisher discriminant (KFD), and committee machines (ensemble averaging and AdaBoost), most of which have been developed recently in statistical learning theory. We formulate differentiation of malignant from benign MCs as a supervised learning problem, and apply these learning methods to develop the classification algorithms. As input, these methods use image features automatically extracted from clustered MCs. We test these methods using a database of 697 clinical mammograms from 386 cases, which include a wide spectrum of difficult-to-classify cases. We use receiver operating characteristic (ROC) analysis to evaluate and compare the classification performance by the different methods. In addition, we also investigate how to combine information from multiple-view mammograms of the same case so that the best decision can be made by a classifier. In our experiments, the kernel-based methods (i.e., SVM, KFD) yield the best performance, significantly outperforming a well-established CADx approach based on neural network learning.

We have developed a computer-aided detection (CAD) system for breast masses on mammograms. In this study, our purpose was to improve the performance of our mass detection system by using a new dual system approach which combines a CAD system optimized with "average" masses with another CAD system optimized with subtle masses. The latter system is trained to provide high sensitivity in detecting subtle masses. For an unknown mammogram, the two systems are used in parallel to detect suspicious objects. A feed-forward backpropagation neural network trained to merge the scores of the two linear discriminant analysis (LDA) classifiers from the two systems makes the final decision in differentiation of true masses from normal tissue. A data set of 86 patients containing 172 mammograms with biopsy-proven masses was partitioned into a training set and an independent test set. This data set is referred to as the average data set. A second data set of 214 prior mammograms was used for training the second CAD system for detection of subtle masses. When the single CAD system trained on the average data set was applied to the test set, the A_z for false positive (FP) classification was 0.81 and the FP rates were 2.1, 1.5 and 1.3 FPs/image at the case-based sensitivities of 95%, 90% and 85%, respectively. With the dual CAD system, the A_z was 0.85 and the FP rates were improved to 1.7, 1.2 and 0.8 FPs/image at the same case-based sensitivities. Our results indicate that the dual CAD system can improve the performance of mass detection on mammograms.

We present a new algorithm and preliminary results for classifying lesions into BI-RADS shape categories: round, oval, lobulated, or irregular. By classifying masses into one of these categories, computer aided detection (CAD) systems will be able to provide additional information to radiologists. Thus, such a tool could potentially be used in conjunction with a CAD system to enable greater interaction and personalization. For this classification task, we have developed a new set of features using the Beamlet transform, which is a recently developed multi-scale image analysis transform. We trained a k-Nearest Neighbor classifier using images from the Digital Database for Digital Mammography (DDSM). The method was tested on a set of 25 images of each type and we obtained a classification accuracy of 78% for classifying masses as oval or round and an accuracy of 72% for classifying masses as lobulated or round.

Mass detection algorithms generally consist of two stages. The aim of the first stage is to detect all potential masses. In the second stage, the aim is to reduce the false-positives by classifying the detected objects as masses or normal tissue. In this paper, we present a new evidence based, stage-one algorithm for the detection of spiculated masses and architectural distortions. By evidence based, we mean that we use the statistics of the physical characteristics of these abnormalities to determine the parameters of the detection algorithm. Our stage-one algorithm consists of two steps, an enhancement step followed by a filtering step. In the first step, we propose a new technique for the enhancement of spiculations in which a linear filter is applied to the Radon transform of the image. In the second step, we filter the enhanced images with a new class of linear image filters called Radial Spiculation Filters. We have invented these filters specifically for detecting spiculated masses and architectural distortions that are marked by converging lines or spiculations. These filters are highly specific narrowband filters, which are designed to match the expected structures of these abnormalities and form a new class of wavelet-type filterbanks derived from optimal theories of filtering. A key aspect of this work is that each parameter of the filter has been incorporated to capture the variation in physical characteristics of spiculated masses and architectural distortions and that the parameters of the stage-one detection algorithm are determined by the physical measurements.

We introduce a computer-assisted detection (CAD) system for the automated detection of breast masses in screening mammograms. The system targets the directional behavior of the neighborhood pixels surrounding a reference image pixel. The underlying hypothesis is that in the presence of a mass the directional properties of the breast tissue surrounding the mass should be altered. The hypothesis was tested using a database of 1,337 mammographic regions of interest (ROIs) extracted from DDSM mammograms. There were 681 ROIs containing a biopsy-proven mass centered in the ROI (340 malignant, 341 benign) and 656 ROIs depicting normal breast parenchyma. Initially, eight main directional propagations were identified and modeled given the center of the ROI as the reference pixel. Subsequently, eight novel morphological features were extracted for each direction. The features were designed to characterize the disturbance occurring in normal breast parenchyma due to the presence of a mass. Finally, the extracted features were merged using a back propagation neural network (BPANN). The network served as a non linear classifier trained to determine the presence of a mass centered at the reference image pixel. The BPANN was trained and tested using a leave-one-out sampling scheme. Its performance was evaluated with Receiver Operating Characteristics (ROC) analysis. Our CAD system showed an ROC area index of Az=0.88±0.01 for discriminating mass vs. normal ROIs. Detection performance was robust for both malignant (Az=0.88±0.01) and benign masses (Az=0.87±0.01). Thus, the proposed directional neighborhood analysis (DNA) can be applied effectively to identify suspicious masses in screening mammograms.

Breast cancer is second only to lung cancer as a tumor-related cause of death in women. Currently, the method of choice for the early detection of breast cancer is mammography. While sensitive to the detection of breast cancer, its positive predictive value (PPV) is low, resulting in biopsies that are only 15-34% likely to reveal malignancy. This paper explores the use of two novel approaches called Partial Least Squares (PLS) and Kernel-PLS (K-PLS) to the diagnosis of breast cancer. The approach is based on optimization for the partial least squares (PLS) algorithm for linear regression and the K-PLS algorithm for non-linear regression. Preliminary results show that both the PLS and K-PLS paradigms achieved comparable results with three separate support vector learning machines (SVLMs), where these SVLMs were known to have been trained to a global minimum. That is, the average performance of the three separate SVLMs were Az = 0.9167927, with an average partial Az (Az90) = 0.5684283. These results compare favorably with the K-PLS paradigm, which obtained an Az = 0.907 and partial Az = 0.6123. The PLS paradigm provided comparable results. Secondly, both the K-PLS and PLS paradigms out performed the ANN in that the Az index improved by about 14% (Az ≈ 0.907 compared to the ANN Az of ≈ 0.8). The "Press R squared" value for the PLS and K-PLS machine learning algorithms were 0.89 and 0.9, respectively, which is in good agreement with the other MOP values.

We present the results on the development of an automated system for prostate cancer diagnosis and Gleason grading. Images of representative areas of the original Hematoxylin-and-Eosin (H&E)-stained tissue retrieved from each patient, either from a tissue microarray (TMA) core or whole section, were captured and analyzed. The image sets consisted of 367 and 268 color images for the diagnosis and Gleason grading problems, respectively. In diagnosis, the goal is to classify a tissue image into tumor versus non-tumor classes. In Gleason grading, which characterizes tumor aggressiveness, the objective is to classify a tissue image as being from either a low- or high-grade tumor. Several feature sets were computed from the image. The feature sets considered were: (i) color channel histograms, (ii) fractal dimension features, (iii) fractal code features, (iv) wavelet features, and (v) color, shape and texture features computed using Aureon Biosciences' MAGIC system. The linear and quadratic Gaussian classifiers together with a greedy search feature selection algorithm were used. For cancer diagnosis, a classification accuracy of 94.5% was obtained on an independent test set. For Gleason grading, the achieved accuracy of classification into low- and high-grade classes of an independent test set was 77.6%.

Uterine cervical cancer is the second most common cancer among women worldwide. Colposcopy is a diagnostic method, whereby a physician (colposcopist) visually inspects the lower genital tract (cervix, vulva and vagina), with special emphasis on the subjective appearance of metaplastic epithelium comprising the transformation zone on the cervix. Cervical cancer precursor lesions and invasive cancer exhibit certain distinctly abnormal morphologic features. Lesion characteristics such as margin; color or opacity; blood vessel caliber, intercapillary spacing and distribution; and contour are considered by colposcopists to derive a clinical diagnosis. Clinicians and academia have suggested and shown proof of concept that automated image analysis of cervical imagery can be used for cervical cancer screening and diagnosis, having the potential to have a direct impact on improving women’s health care and reducing associated costs. STI Medical Systems is developing a Computer-Aided-Diagnosis (CAD) system for colposcopy -- ColpoCAD. At the heart of ColpoCAD is a complex multi-sensor, multi-data and multi-feature image analysis system. A functional description is presented of the envisioned ColpoCAD system, broken down into: Modality Data Management System, Image Enhancement, Feature Extraction, Reference Database, and Diagnosis and directed Biopsies. The system design and development process of the image analysis system is outlined. The system design provides a modular and open architecture built on feature based processing. The core feature set includes the visual features used by colposcopists. This feature set can be extended to include new features introduced by new instrument technologies, like fluorescence and impedance, and any other plausible feature that can be extracted from the cervical data. Preliminary results of our research on detecting the three most important features: blood vessel structures, acetowhite regions and lesion margins are shown. As this is a new and very complex field in medical image processing, the hope is that this paper can provide a framework and basis to encourage and facilitate collaboration and discussion between industry, academia, and medical practitioners.

The purpose of this study is to develop a computer-aided diagnosis (CAD) system to detect small-sized (from 2mm to 10mm) pulmonary nodules in high resolution helical CT scans. A new CAD system is proposed to locate both juxtapleural nodules and non-pleural nodules. Isotropic resampling and lung segmentation are performed as preprocessing steps. Morphological closing was utilized to smooth the lung contours to include the indented possible juxtapleural locations, thresholding and 3D component analysis were used to obtain 3D volumetric nodule candidates; furthermore, gray level and geometric features were extracted, and analyzed using linear discriminant analysis (LDA) classifier. Leave one case out method was used to evaluate the LDA. To deal with non-pleural nodules, a discrete-time cellular neural network (DTCNN) based on local shape features was developed. This scheme employed the local shape property to perform voxel classification. The shape index feature successfully captured the local shape difference between nodules and non-nodules, especially vessels. To tailor it for lung nodule detection, this DTCNN was trained using genetic algorithms (GAs) to derive the shape index variation pattern of nodules. Nonoverlapping training and testing sets were utilized in the non-pleural nodule detection. 19 clinical thoracic CT cases involving a total of 4838 sectional images were used in this work. The juxtapleural nodule detection method was able to obtain sensitivity 81.25% with an average of 8.29 FPs per case. The non-pleural nodule finding scheme attained sensitivity of 83.9% with an average 3.47 FPs/case. Combining the two subsystems together, an overall performance of 82.98% sensitivity with 11.76 FPs/case can be obtained.

A major challenge in computer-aided diagnostic (CAD) schemes for nodule detection on chest radiographs is the detection of nodules that overlap with ribs. Our purpose was to develop a technique for false-positive reduction in a CAD scheme using a rib-suppression technique based on massive training artificial neural networks (MTANNs). We developed a multiple MTANN (multi-MTANN) consisting of eight MTANNs for removing eight types of false positives. For further removal of false positives caused by ribs, we developed a rib-suppression technique using a multi-resolution MTANN consisting of three different resolution MTANNs. To suppress the contrast of ribs, the multi-resolution MTANN was trained with input chest radiographs and the teaching soft-tissue images obtained by using a dual-energy subtraction technique. Our database consisted of 91 nodules in 91 chest radiographs. With our original CAD scheme based on a difference image technique with linear discriminant analysis, a sensitivity of 82.4% (75/91 nodules) with 4.5 (410/91) false positives per image was achieved. The trained multi-MTANN was able to remove 62.7% (257/410) of false positives with a loss of one true positive. With the rib-suppression technique, the contrast of ribs in chest radiographs was suppressed substantially. Due to the effect of rib-suppression, 41.2% (63/153) of the remaining false positives were removed without a loss of any true positives. Thus, the false-positive rate of our CAD scheme was improved substantially, while a high sensitivity was maintained.

We have been developing computer-aided diagnostic (CAD) scheme to assist radiologists detect lung nodules in thoracic CT images. In order to improve the sensitivity for nodule detection, we developed a selective nodule enhancement filter for nodule which can simultaneously enhance nodules and suppress other normal anatomic structures such as blood vessels and airway walls. Therefore, as preprocessing steps, this filter is useful for improving the sensitivity of nodule detection and for reducing the number of false positives. Another new technique we employed in this study is an automated rule-based classifier. It can significantly reduce the extent of the disadvantages of existing rule-based classifiers, including manual design, poor reproducibility, poor evaluation methods such as re-substitution, and a large overtraining effect. Experimental results performed with Monte Carlo simulation and a real lung nodule CT dataset demonstrated that the automated method can completely eliminate overtraining effect in the procedure of cutoff threshold selection, and thus can minimize overall overtraining effect in the rule-based classifier.

Ground glass nodules (GGNs) have proved especially problematic in lung cancer diagnosis, as despite frequently being malignant they have extremely slow growth rates. In this work, the GGN segmentation results of a computer-based method were compared with manual segmentation performed by two dedicated chest radiologists. CT volumes of 8 patients were acquired by multi-slice CT. 21 pure or mixed GGNs were identified and independently segmented by the computer-based method and by two readers. The computer-based method is initialized by a click point, and uses a Markov random field (MRF) model for segmentation. While the intensity distribution varies for different GGNs, the intensity model used in MRF is adapted for each nodule based on initial estimates. This method was run three times for each nodule using different click points to evaluate consistency. In this work, consistency was defined by the overlap ratio (overlap volume/mean volume). The consistency of the computer-based method with different initial points, with a mean overlap ratio of 0.96±0.02 (95% confidence interval on mean), was significantly higher than the inter-observer consistency between the two radiologists, indicated by a mean overlap ratio of 0.73±0.04. The computer consistency was also significantly higher than the intra-observer consistency of two measurements from the same radiologist, indicated by an overlap ratio of 0.69±0.05 (p-value < 1E-05). The concordance of the computer with the expert interpretation demonstrated a mean overlap ratio of 0.69±0.05. As shown by our data, the consistency provided by the computer-based method is significantly higher than between observers, and the accuracy of the method is no worse than that of one physician’s accuracy with respect to another, allowing more reproducible assessment of nodule growth.

We present a novel technique that provides a case-specific confidence measure for artificial neural network (ANN) based computer-assisted diagnostic (CAD) decisions. The technique relies on the analysis of the feature space neighborhood for each query case and dynamically creates a validation set that allows estimation of a local accuracy of the decisions made by the network. Then a case-specific reliability measure is assigned to each system's response, which can be used to stratify network's predictions according to the acceptable validation error value. The study was performed using a database containing 1,337 mammographic regions of interest (ROIs) with biopsy-proven diagnosis (681 with masses, 656 with normal parenchyma). Two types of neural networks (1) a feed forward network with error back propagation (BPNN) and (2) a generalized regression neural network with RBF nodes (GRNN) were developed to detect masses based on 8 morphological features automatically extracted from each ROI. The performance of the networks was evaluated with Receiver Operating Characteristics (ROC) analysis. The study shows that as the threshold on the acceptable validation error declines, the technique rejects more CAD decisions as not reliable enough. However, the ROC performance for the reliable results steadily improves (from A_z = 0.88 to A_z = 0.98 for BPNN, from A_z = 0.86 to A_z = 0.97 for GRNN). The proposed technique provides a stratification strategy for predictions made by CAD tools and can be applied to any type of decision algorithms.

Measuring the similarity between discretely sampled intensity values of different images as a function of geometric transformations is necessary for performing automatic image registration. Arbitrary spatial transformations require a continuous model for the intensity values of the discrete images. Because of computation cost most researchers choose to use low order basis functions, such as the linear hat function or low order B-splines, to model the discrete images. Using the theory of random processes we show that low order interpolators cause undesirable local optima artifacts in similarity measures based on the L₂ norm, linear correlation coefficient, and mutual information. We show how these artifacts can be significantly reduced, and at times completely eliminated, by using sinc approximating kernels.

We present a set of new methods for efficient and precise registration of any X-Ray modality (fluoroscopy, portal imaging or regular X-Ray imaging) to a CT data set. These methods require neither feature extraction nor 2D or 3D segmentation. Our main contribution is to directly perform the computations on the gradient vector volume of the CT data, which has several advantages. It can increase the precision of the registration as it assesses mainly the alignment of intensity edges in both CT and X-Ray images. By using only significant areas of the gradient vector volume, the amount of information needed in each registration step can be reduced up to a factor of 10. This both speeds up the registration process and allows for using the CT data with full precision, e.g. 512³ voxels. We introduce a Volume Gradient Rendering (VGR) as well as a Volume Gradient Correlation (VGC) method, where the latter one can be used directly for computing the image similarity without DBR generation.

The evaluation of tumor growth as regression under therapy is an important clinical issue. Rigid registration of sequentially acquired 3D-images has proven its value for this purpose. Existing approaches to rigid image registration use the whole volume for the estimation of the rigid transform. Non-rigid soft tissue deformation, however, will imply a bias to the registration result, because local deformations cannot be modeled by rigid transforms. Anatomical substructures, like bones or teeth, are not affected by these deformations, but follow a rigid transform. This important observation is incorporated in the proposed registration algorithm. The selection of anatomical substructure is done by manual interaction of medical experts adjusting the transfer function of the volume rendering software. The parameters of the transfer function are used to identify the voxels that are considered for registration. A rigid transform is estimated by a quaternion gradient descent algorithm based on the intensity values of the specified tissue classes. Commonly used voxel intensity measures are adjusted to the modified registration algorithm. The contribution describes the mathematical framework of the proposed registration method and its implementation in a commercial software package. The experimental evaluation includes the discussion of different similarity measures, the comparison of the proposed method to established rigid registration techniques and the evaluation of the efficiency of the new method. We conclude with the discussion of potential medical applications of the proposed registration algorithm.

The issue of this paper is about real-time or interactive 2D-2D resp. 3D-3D matching. Based on Viola's sample-based stochastic Mutual Information (MI) gradient matching we developed a technique that allows to optimally set all necessary parameters in a short preprocessing step using typical images. In this paper we concentrate on finding an optimal parameter set for Rprop, the underlying stochastic optimizer. The relevant parameters are the start and the minimum learning rate given a pair of aligned images. Rprop is modelled by a Markov chain whose transition kernel is estimated by a mean gradient. We introduce a truncated recursion to simulate Rprop and obtain an expectation for the number of iterations for each parameter combination. This way near optimal parameters are found within 20-50 seconds, depending on the data. Using automatically set parameters for Rprop and the sample size, matching requires 0.3-1.3 s for 2D-2D and 0.6-2.1 s for 3D-3D on our test data using an Athlon 800 MHz processor. Altogether we get a real-time registration algorithm that optimizes its control parameters for the given data within less than a minute.

Much literature on image registration has worked with purely geometric image deformation models. For such models, interpolation/resampling operations are often the computationally intensive steps when iteratively minimizing the deformation cost function. This article discusses some techniques for efficiently implementing and accelerating these operations. To simplify presentation, we discuss our ideas in the context of 2D imaging. However, the concepts readily generalize to 3D. Our central technique is a table-lookup scheme that makes somewhat liberal use of RAM, but should not strain the resources of modern processors if certain design parameters are appropriately selected. The technique works by pre-interpolating and tabulating the grid values of the reference image onto a finer grid along one of the axes of the image. The lookup table can be rapidly constructed using FFTs. Our results show that this technique reduces iterative computation by an order of magnitude. When a minimization algorithm employing coordinate block alternation is used, one can obtain still faster computation by storing certain intermediate quantities as state variables. We refer to this technique as state variable hold-over. When combined with table-lookup, state variable hold-over reduces CPU time by about a factor two, as compared to table-lookup alone.

Nonrigid registration of medical images by maximisation of their mutual information, in combination with a deformation field parameterised by cubic B-splines, has been shown to be robust and accurate in many applications. However, the high computation time is a big disadvantage. This work focusses on the optimisation procedure. Many implementations follow a gradient-descent like approach. The time needed for computing the derivative of the mutual information with respect to the B-spline parameters is the bottleneck in this process. We investigate the influence of several gradient approximation techniques on the number of iterations needed and the computation time per iteration. Three methods are studied: a simple finite difference strategy, the so-called simultaneous perturbation method, and a more analytic computation of the gradient based on a continuous, and differentiable representation of the joint histogram. In addition, the effect of decreasing the number of image samples, used for computing the gradient in each iteration, is investigated. Two types of experiments are performed. Firstly, the registration of an image to itself, after application of a known, randomly generated deformation, is considered. Secondly, experiments are performed with 3D ultrasound brain scans, and 3D CT follow-up scans of the chest. The experiments show that the method using an analytic gradient computation outperforms the other two. Furthermore, the computation time per iteration can be extremely decreased, without affecting the rate of convergence and final accuracy, by using very few samples of the image (randomly chosen every iteration) to compute the derivative. With this approach, large data sets (256³) can be registered within 5 minutes on a moderate PC.

In this paper, we examine the problem of non-rigid, image-to-image registration for CT images of the abdomen. This problem has been previously addressed in many different contexts (e.g., visualization using different imaging modalities, modelling of organ deformation after surgical resection). The particular application in which we are interested is modelling of respiratory motion of abdominal organs, so that we may achieve a more accurate representation of the dose distribution in both targeted structures and non-targeted areas during radiosurgical treatment. Our goal is to register two CT images, each acquired at different positions in the respiratory cycle. We use a transformation model based on B-splines, and take advantage of B-splines' "locality" or "compact support" property to ensure computational efficiency and robust convergence to a satisfactory result. We demonstrate that, although a purely intensity-based registration metric performs well in matching the deformation of the lungs during the respiratory cycle, the movement of other organs (e.g., liver and kidney) is poorly represented due to the poor contrast within and between these organs in the CT images. We introduce a registration metric that is a weighted combination of intensity difference and distance between corresponding points that are manually identified in the two images being registered; we show how the influence of these points can be elegantly added to the metric so that it remains differentiable with respect to the spline control points. Visual inspection reveals that resulting registrations appear to be superior to the intensity-only ones in terms of representation of visceral organ deformation and movement.

In present-day medical practice it is often necessary to nonrigidly align image data, either intra- or inter-patient. Current registration algorithms usually do not take different tissue types into account. A problem that might occur with these algorithms is that rigid tissue, like bone, also deforms elastically. We propose a method to correct a deformation field, that is calculated with a nonrigid registration algorithm. The correction is based on a second feature image, which represents the tissue stiffness. The amount of smoothing of the deformation field is related to this stiffness coefficient. By filtering the deformation field on rigid tissue, the deformation field will represent a local rigid transformation. Other parts of the image containing nonrigid tissue are smoothed less, which leaves the original elastic deformation (almost) untouched. It is shown on a synthetic example and on inspiration-expiration CT data of the thorax, that a filtering of the deformation field based on tissue type indeed keeps rigid tissue rigid, thus improving the registration results.

Breath holding (BH) allows to immobilize organs during radiotherapy treatment of lung cancer. Deformable registration methods applied on 3D Computerized Tomography (CT) scans acquired in BH can be used to evaluate the breath holding reproducibility. Resulting 3D vector fields could then be used to adapt internal margins for each patient. In this work we compare three non-rigid registration schemes with Gaussian, linear-elastic and Nagel-Enckelmann based regularizations. As we do not dispose of gold standard, we analyze vector fields by several operators (transitivity, symmetry, volume dilatation, Jacobian). Experiments were based on clinical data sets of two patients: one with normal lung behavior and second with lung discrepancies which lead to bad BH reproducibility. Results show that none of operators allows to clearly highlight the superiority of a method, except for convergence rapidity and Jacobian.

We introduce an image structure analysis technique suitable in cases where anisotropy plays an important role. The so-called Locally Adapted Scaling Vector Method (LSVM) comprises two steps. First, a procedure to estimate the local main orientation at every point of the image is applied. These orientations are then incorporated in a structure characterization procedure. We apply this methodology to High Resolution Magnetic Resonance Images (HRMRI) of human proximal femoral specimens IN VITRO. We extract a 3D local texture measure to establish correlations with the biomechanical properties of bone specimens quantified via the bone maximum compressive strength. The purpose is to compare our results with the prediction of bone strength using similar isotropic texture measures, bone mineral density, and standard 2D morphometric parameters. Our findings suggest that anisotropic texture measures are superior in cases where directional properties are relevant.

We extend the recently introduced scaling vector method (SVM) to improve the textural characterization of oriented trabecular bone structures in the context of osteoporosis. Using the concept of scaling vectors one obtains non-linear structural information from data sets, which can account for global anisotropies. In this work we present a method which allows us to determine the local directionalities in images by using scaling vectors. Thus it becomes possible to better account for local anisotropies and to implement this knowledge in the calculation of the scaling properties of the image. By applying this adaptive technique, a refined quantification of the image structure is possible: we test and evaluate our new method using realistic two-dimensional simulations of bone structures, which model the effect of osteoblasts and osteoclasts on the local change of relative bone density. The partial differential equations involved in the model are solved numerically using cellular automata (CA). Different realizations with slightly varying control parameters are considered. Our results show that even small changes in the trabecular structures, which are induced by variation of a control parameters of the system, become discernible by applying the locally adapted scaling vector method. The results are superior to those obtained by isotropic and/or bulk measures. These findings may be especially important for monitoring the treatment of patients, where the early recognition of (drug-induced) changes in the trabecular structure is crucial.

Osteoporosis is a metabolic bone disease leading to de-mineralization and increased risk of fracture. The two major factors that determine the biomechanical competence of bone are the degree of mineralization and the micro-architectural integrity. Today, modern imaging modalities (high resolution MRI, micro-CT) are capable of depicting structural details of trabecular bone tissue. From the image data, structural properties obtained by quantitative measures are analysed with respect to the presence of osteoporotic fractures of the spine (in-vivo) or correlated with biomechanical strength as derived from destructive testing (in-vitro). Fairly well established are linear structural measures in 2D that are originally adopted from standard histo-morphometry. Recently, non-linear techniques in 2D and 3D based on the scaling index method (SIM), the standard Hough transform (SHT), and the Minkowski Functionals (MF) have been introduced, which show excellent performance in predicting bone strength and fracture risk. However, little is known about the performance of the various parameters with respect to monitoring structural changes due to progression of osteoporosis or as a result of medical treatment. In this contribution, we generate models of trabecular bone with pre-defined structural properties which are exposed to simulated osteoclastic activity. We apply linear and non-linear texture measures to the models and analyse their performance with respect to detecting architectural changes. This study demonstrates, that the texture measures are capable of monitoring structural changes of complex model data. The diagnostic potential varies for the different parameters and is found to depend on the topological composition of the model and initial “bone density”. In our models, non-linear texture measures tend to react more sensitively to small structural changes than linear measures. Best performance is observed for the 3rd and 4th Minkowski Functionals and for the scaling index method.

The goal of this study was to compare the performance of decision tree (DT) classifiers with artificial neural network (ANN) and linear discriminant analysis (LDA) classifiers under different conditions for the class distributions, feature space dimensionality, and training sample size using a Monte Carlo simulation study. We also investigated a bagging technique for improving the accuracy of the DT. The resubstitution (training) and test accuracies of the classifiers were compared by using the area A_z under the ROC curve as the performance measure. Three types of feature space distributions were studied: the Gaussian feature space, a mixture of Gaussians, and a mixture of uniform distributions. The feature space dimensionality was varied between 2 and 12. In a given experiment, 1000 cases were randomly sampled from each distribution, N_t trainers per class was used for classifier design, and the remaining cases were used to test the classifier. The effect of the training sample size was investigated by varying N_t between 30 and 500. Performance measures from 100 experiments were averaged. Our results indicated that, in the Gaussian feature space, the LDA outperformed the other two classifiers, especially when the number of trainers was low. For the mixture of uniform distributions, the A_z value of the DT was in general higher than that of the ANN and the LDA. For the mixture of Gaussians, the performances of the DT and ANN classifiers were comparable. Our results indicate that a DT can be a viable alternative to ANN and LDA classifiers in certain feature spaces.

Micro-array protein crystal images are now routinely acquired automatically by CCD cameras. High-throughput automatic classification of protein crystals requires to alleviation of the time-consuming task of manual visual inspection. We propose a classification framework combined with a multi-scale image processing method for recognizing protein crystals and precipitates versus clear drops. The main two points of the processing method are the multi-scale Laplacian pyramid filters and histogram analysis techniques to find an effective feature vector. The processing steps include: 1. Tray well cropping using Radon Transform; 2. Droplet cropping using an ellipsoid Hough Transform; 3. Multi-scale image separation with Laplacian pyramidal filters; 4. Feature vector extraction from the histogram of the multi-scale boundary images. The feature vector combines geometric and texture features of each image and provides input to a feed forward binomial neural network classifier. Using human (expert crystallographers) classified images as ground truth, the current experimental results gave 86% true positive and 94% true negative rates (average true percentage is 90%) using an image database which contained over 2,000 images. To enable NESG collaborators to carry our crystal classification, a web-based Matlab server was also developed. Users at other locations on the internet can input micro-array crystal image folders and parameters for training and testing processes through a friendly web interface. Recognition results are shown on the client side website and may be downloaded by a remote user as an Excel spreadsheet file.

Wireless Capsule Endoscopy (WCE) is a new colour imaging technology that enables close examination of the interior of the entire small intestine. Typically, the WCE operates for ~8 hours and captures ~40,000 useful images. The images are viewed as a video sequence, which generally takes a doctor over an hour to analyse. In order to activate certain key features of the software provided with the capsule, it is necessary to locate and annotate the boundaries between certain gastrointestinal (GI) tract regions (stomach, intestine and colon) in the footage. In this paper we propose a method of automatically discriminating stomach, intestine and colon tissue in order to significantly reduce the video assessment time. We use hue saturation chromaticity histograms which are compressed using a hybrid transform, incorporating the Discrete Cosine Transform (DCT) and Principal Component Analysis (PCA). The performance of two classifiers is compared: k-nearest neighbour (kNN) and Support Vector Classifier (SVC). After training the classifier, we applied a narrowing step algorithm to converge to the points in the video where the capsule firstly passes through the pylorus (the valve between the stomach and the intestine) and later the ileocaecal valve (IV, the valve between the intestine and colon). We present experimental results that demonstrate the effectiveness of this method.

Advances in video technology are being incorporated into today’s healthcare practice. For example, colonoscopy is an important screening tool for colorectal cancer. Colonoscopy allows for the inspection of the entire colon and provides the ability to perform a number of therapeutic operations during a single procedure. During a colonoscopic procedure, a tiny video camera at the tip of the endoscope generates a video signal of the internal mucosa of the colon. The video data are displayed on a monitor for real-time analysis by the endoscopist. Other endoscopic procedures include upper gastrointestinal endoscopy, enteroscopy, bronchoscopy, cystoscopy, and laparoscopy. However, a significant number of out-of-focus frames are included in this type of videos since current endoscopes are equipped with a single, wide-angle lens that cannot be focused. The out-of-focus frames do not hold any useful information. To reduce the burdens of the further processes such as computer-aided image processing or human expert’s examinations, these frames need to be removed. We call an out-of-focus frame as non-informative frame and an in-focus frame as informative frame. We propose a new technique to classify the video frames into two classes, informative and non-informative frames using a combination of Discrete Fourier Transform (DFT), Texture Analysis, and K-Means Clustering. The proposed technique can evaluate the frames without any reference image, and does not need any predefined threshold value. Our experimental studies indicate that it achieves over 96% of four different performance metrics (i.e. precision, sensitivity, specificity, and accuracy).

This paper describes a novel framework to build 3D Point Distribution Model (PDM) from a set of segmented volumetric images. This method is based on a deformable model algorithm. Each training sample deforms to approximate all other training shapes. The training sample with best approximation results is then chosen as the template. Finally, the poor approximation results from this template are improved by the "bridge over" scheme, which deforms the template to approximate intermediate training shapes and then deforms the approximations to outliers. The method is applied to construct a 3D PDM of 20 human brain ventricles. The results show that the algorithm leads to more accurate representation than traditional framework. Also, the performance of the PDM of soft tissue is comparable with the PDM of bone structures by a previous method. The traditional framework of deformable model based approach selects the template arbitrarily and deforms the template to approximate training shapes directly. The limitation of the traditional framework is that the representation accuracy of the PDM entirely depends on the direct approximation. Moreover, the arbitrary template selection deteriorates the accuracy of the approximation. Our framework that features template selection and indirect approximation solves the shortcomings and improves the PDM representation accuracy. Furthermore, the "bridge over" framework could be used with any deformable model algorithm. In this sense, the method is a generic framework open to future investigation.

Deformable models are powerful approaches to medical image analysis, particularly segmentation. However, the outcome of applying a parametric or geometric deformable model is often significantly dependent on its initialization. This is an obstacle to robust automatic segmentation. Based on theoretical analyses of the watershed transform, we propose a novel approach to reducing this sensitivity to initialization by deriving a vector field from topographic and Euclidean distance transforms. This vector field is aimed to extend the influence of the gradients at the boundary of the segmentation target over the entire image in a consistent fashion, while ignoring any irrelevant gradients in the original image. Initiated by one or more segmentation seeds, the vector field is first computed using an efficient numerical method, and subsequently participates in the model's evolution process. Integration of the vector field has so far been performed with a two-dimensional (2D) parametric deformable model and with a three-dimensional (3D) geodesic active contour level set model. We believe that our approach will enable a higher degree of automation for deformable-model-based segmentation, particularly in situations where the seeds can be placed automatically based on, for example, a priori knowledge regarding the anatomy and the intensity differentiation between the target and the background. Experiments on segmenting organs and tumors from CT and MR images using the integrated models have shown that this is a promising approach.

Previously, "Deformable organisms" were introduced as a novel paradigm for medical image analysis that uses artificial life modelling concepts. Deformable organisms were designed to complement the classical bottom-up deformable models methodologies (geometrical and physical layers), with top-down intelligent deformation control mechanisms (behavioral and cognitive layers). However, a true physical layer was absent and in order to complete medical image segmentation tasks, deformable organisms relied on pure geometry-based shape deformations guided by sensory data, prior structural knowledge, and expert-generated schedules of behaviors. In this paper we introduce the use of physics-based shape deformations within the deformable organisms framework yielding additional robustness by allowing intuitive real-time user guidance and interaction when necessary. We present the results of applying our physics-based deformable organisms, with an underlying dynamic spring-mass mesh model, to segmenting and labelling the corpus callosum in 2D midsagittal magnetic resonance images.

Rapid and unsupervised quantitative analysis is of utmost importance to ensure clinical acceptance of many examinations using cardiac magnetic resonance imaging (MRI). We present a framework that aims at fulfilling these goals for the application of left ventricular ejection fraction estimation in four-dimensional MRI. The theoretical foundation of our work is the generative two-dimensional Active Appearance Models by Cootes et al., here extended to bi-temporal, three-dimensional models. Further issues treated include correction of respiratory induced slice displacements, systole detection, and a texture model pruning strategy. Cross-validation carried out on clinical-quality scans of twelve volunteers indicates that ejection fraction and cardiac blood pool volumes can be estimated automatically and rapidly with accuracy on par with typical inter-observer variability.

Segmentation and labelling of the left atrium from pre-operative images could be a valuable source of information for the planning of electrophysiology procedures to cure atrial fibrillation. A method is presented that uses multi-slice computed tomography (MSCT) images for this purpose that were initially acquired for coronary assessment. The method combines the power of active shape models (robustness by use of prior anatomical knowledge) with the advantages of solely data driven segmentation methods (accuracy). A triangular shape model was built for the human left atrium and its pulmonary vein trunks. It was automatically adapted to the MSCT images, labelling these structures and segmenting them coarsely. In addition, a segmentation of the blood pool by a Hounsfield threshold was applied to the images. The enclosed volumes were triangulated to get a fine surface representation yet still including many distracting objects (the artery tree, coronaries, adjacent chambers, and bones). A correspondence between surface triangles of the coarse, but anatomically labelled model surface and those of the fine iso-surface was established by a similarity criterion on position and orientation. This allows for the refinement of the model-based segmentation showing more anatomical details by selection of corresponding parts of the iso-surface. Vice versa, the correspondence could be used to assign anatomical labels to each iso-surface patch.

This paper presents a method for automated deformation recovery of the left and right ventricular wall from a time sequence of anatomical images of the heart. The deformation is recovered within the heart wall, i.e. it is not limited only to the epicardium and endocardium. Most of the suggested methods either ignore or approximately model incompressibility of the heart wall. This physical property of the cardiac muscle is mathematically guaranteed to be satisfied by the proposed method. A scheme for decomposition of a complex incompressible geometric transformation into simpler components and its application to cardiac deformation recovery is presented. A general case as well as an application specific solution is discussed. Furthermore, the manipulation of the constructed incompressible transformations, including the computation of the inverse transformation, is computationally inexpensive. The presented method is mathematically guaranteed to generate incompressible transformations which are experimentally shown to be a very good approximation of actual cardiac deformations. The transformation representation has a relatively small number of parameters which leads to a fast deformation recovery. The approach was tested on six sequences of two-dimensional short-axis cardiac MR images. The cardiac deformation was recovered with an average error of 1.1 pixel. The method is directly extendable to three dimensions and to the entire heart.

The original FDK algorithm has been extensively employed in medical and industrial imaging applications. With an increased cone angle, cone beam (CB) artifacts in images reconstructed by the original FDK algorithm deteriorate, since the circular trajectory does not satisfy the so-called data sufficiency condition (DSC). A few “circular plus” trajectories have been proposed in the past to reduce CB artifacts by meeting the DSC. However, the circular trajectory has distinct advantages over other scanning trajectories in practical CT imaging, such as cardiac, vascular and perfusion applications. In addition to looking into the DSC, another insight into the CB artifacts of the original FDK algorithm is the inconsistency between conjugate rays that are 180° apart in view angle. The inconsistence between conjugate rays is pixel dependent, i.e., it varies dramatically over pixels within the image plane to be reconstructed. However, the original FDK algorithm treats all conjugate rays equally, resulting in CB artifacts that can be avoided if appropriate view weighting strategy is exercised. In this paper, a modified FDK algorithm is proposed, along with an experimental evaluation and verification, in which the helical body phantom and a humanoid head phantom scanned by a volumetric CT (64 x 0.625 mm) are utilized. Without extra trajectories supplemental to the circular trajectory, the modified FDK algorithm applies reconstruction-plane-dependent view weighting on projection data before 3D backprojection, which reduces the inconsistency between conjugate rays by suppressing the contribution of one of the conjugate rays with a larger cone angle. Both computer-simulated and real phantom studies show that, up to a moderate cone angle, the CB artifacts can be substantially suppressed by the modified FDK algorithm, while advantages of the original FDK algorithm, such as the filtered backprojection algorithm structure, 1D ramp filtering, and data manipulation efficiency, can be maintained.

C-Arm CT systems suffer from artifacts due to truncated projections caused by a finite detector size. One method used to mitigate the truncation artifacts is projection extrapolation without a priori knowledge. This work focuses on estimating the 0^th and 1^st moments of an image, which can be used to extrapolate a set of truncated projections. If some projections are not truncated, then accurate estimation of the moments can be achieved using only those projections. The more difficult case arises when all projections are truncated by some amount. For this case we make simplifying assumptions and fit the truncated projections with elliptical profiles. From this fit, we estimate the 0^th and 1^st moments of the original image. These estimated moments are then used to perform an extrapolation of the truncated projections, where each projection meets a constraint based on the 0^th and 1^st moments (moment extrapolation). This work focuses on how accurate moment estimates must be for moment extrapolation to be effective. The algorithm was tested on simulated and real data for the head, thorax, and abdomen, and those results were compared to symmetric mirroring by Ohnesorge et al., another extrapolation technique that requires no a priori knowledge. Overall, moment estimation and mass extrapolation alleviates a large amount of image artifact, and can improve on other extrapolation techniques. For the real CT head and abdominal data, the average reconstruction error for mass extrapolation was 48% less than the reconstruction error for symmetric mirroring.

X-ray 3D rotational angiography based on C-arm systems has become a versatile and established tomographic imaging modality for high contrast objects in interventional environment. Improvements in data acquisition, e.g. by use of flat panel detectors, will enable C-arm systems to resolve even low-contrast details. However, further progress will be limited by the incompleteness of data acquisition on the conventional short-scan circular source trajectories. Cone artifacts, which result from that incompleteness, significantly degrade image quality by severe smearing and shading. To assure data completeness a combination of a partial circle with one or several line segments is investigated. A new and efficient reconstruction algorithm is deduced from a general inversion formula based on 3D Radon theory. The method is theoretically exact, possesses shift-invariant filtered backprojection (FBP) structure, and solves the long object problem. The algorithm is flexible in dealing with various circle and line configurations. The reconstruction method requires nothing more than the theoretically minimum length of scan trajectory. It consists of a conventional short-scan circle and a line segment approximately twice as long as the height of the region-of-interest. Geometrical deviations from the ideal source trajectory are considered in the implementation in order to handle data of real C-arm systems. Reconstruction results show excellent image quality free of cone artifacts. The proposed scan trajectory and reconstruction algorithm assure excellent image quality and allow low-contrast tomographic imaging with C-arm based cone-beam systems. The method can be implemented without any hardware modifications on systems commercially available today.

Recently, there has been interest in estimating kinetic model parameters for each voxel in a PET image. To do this, the activity images are first reconstructed from PET sinogram frames at each measurement time, and then the kinetic parameters are estimated by fitting a model to the reconstructed time-activity response of each voxel. However, this indirect approach to kinetic parameter estimation tends to reduce signal-to-noise ratio (SNR) because of the requirement that the sinogram data be divided into individual time frames. In 1985, Carson and Lange proposed, but did not implement, a method based on the EM algorithm for direct parametric reconstruction. More recently, researchers have developed semi-direct methods which use spline-based reconstruction, or direct methods for estimation of kinetic parameters from image regions. However, direct voxel-wise parametric reconstruction has remained a challenge due to the unsolved complexities of inversion and required spatial regularization. In this work, we demonstrate an efficient method for direct voxel-wise reconstruction of kinetic parameters (as a parametric image) from all frames of the PET data. The direct parametric image reconstruction is formulated in a Bayesian framework, and uses the parametric iterative coordinate descent (PICD) algorithm to solve the resulting optimization problem. This PICD algorithm is computationally efficient and allows the physiologically important kinetic parameters to be spatially regularized. Our experimental simulations demonstrate that direct parametric reconstruction can substantially reduce estimation error of kinetic parameters as compared to indirect methods.

In many applications of circular cone-beam CT, it is not uncommon that the size of the field of view (FOV) is smaller than that of the imaging object, thus leading to transverse truncation in projection data. Exact reconstruction in any region is not possible from such truncated data using conventional algorithms. Recently, an exact algorithm for image reconstruction on PI-line segments in helical cone-beam CT has been proposed. This algorithm, which we refer to as the backprojection-filtration (BPF) algorithm, can naturally address the problem of exact region of interest (ROI) reconstruction from such truncated data. In this work, we modified this algorithm to reconstructing images in circular cone-beam scan. The unique property of this modified algorithm is that it can reconstruct exact ROIs in midplane and approximate ROIs in other planes from transversely truncated data. We have performed computer-simulation studies to validate the theoretical assertions. Preliminary results demonstrate that the proposed algorithm provides a solution to the truncation problems caused by limited FOV size.

We present a method to incorporate nonlinear shape prior constraints into segmenting different anatomical structures in medical images. Kernel space density estimation (KSDE) is used to derive the nonlinear shape statistics and enable building a single model for a class of objects with nonlinearly varying shapes. The object contour is coerced by image-based energy into the correct shape sub-distribution (e.g., left or right lung), without the need for model selection. In contrast to an earlier algorithm that uses a local gradient-descent search (susceptible to local minima), we propose an algorithm that iterates between dynamic programming (DP) and shape regularization. DP is capable of finding an optimal contour in the search space that maximizes a cost function related to the difference between the interior and exterior of the object. To enforce the nonlinear shape prior, we propose two shape regularization methods, global and local regularization. Global regularization is applied after each DP search to move the entire shape vector in the shape space in a gradient descent fashion to the position of probable shapes learned from training. The regularized shape is used as the starting shape for the next iteration. Local regularization is accomplished through modifying the search space of the DP. The modified search space only allows a certain amount of deformation of the local shape from the starting shape. Both regularization methods ensure the consistency between the resulted shape with the training shapes, while still preserving DP’s ability to search over a large range and avoid local minima. Our algorithm was applied to two different segmentation tasks for radiographic images: lung field and clavicle segmentation. Both applications have shown that our method is effective and versatile in segmenting various anatomical structures under prior shape constraints; and it is robust to noise and local minima caused by clutter (e.g., blood vessels) and other similar structures (e.g., ribs). We believe that the proposed algorithm represents a major step in the paradigm shift to object segmentation under nonlinear shape constraints.

We have developed an active shape model (ASM)-based segmentation scheme that uses the original Cootes et al. formulation for the underlying mechanics of the ASM but improves the model by fixating selected nodes at specific structural boundaries called transitional landmarks. Transitional landmarks identify the change from one boundary type (such as lung-field/heart) to another (lung-field/diaphragm). This results in a multi-segmented lung-field boundary where each segment correlates to a specific boundary type (lung-field/heart, lung-field/aorta, lung-field/rib-cage, etc.). The node-specified ASM is built using a fixed set of equally spaced feature nodes for each boundary segment. This allows the nodes to learn local appearance models for a specific boundary type, rather than generalizing over multiple boundary types, which results in a marked improvement in boundary accuracy. In contrast, existing lung-field segmentation algorithms based only on ASM simply space the nodes equally along the entire boundary without specification. We have performed extensive experiments using multiple datasets (public and private) and compared the performance of the proposed scheme with other contour-based methods. Overall, the improved accuracy is 3-5 &percent; over the standard ASM and, more importantly, it corresponds to increased alignment with salient anatomical structures. Furthermore, the automatically generated lung-field masks lead to the same fROC for lung-nodule detection as hand-drawn lung-field masks. The accurate landmarks can be easily used for detecting other structures in the lung field. Based on the related landmarks (mediastinum-heart transition, heart-diaphragm transition), we have extended the work to heart segmentation.

Despite the advantages that 3D medical image analysis methods offer and the fast introduction of CT and MRI, to date most hospitals use radiographs to perform preoperative planning of hip surgeries and automatic analysis of hip radiographs is still of interest. In this paper, we present a novel method for segmentation of bone structures in anterior-posterior (AP) radiographs based on Active Appearance Models. The pelvis shape is decomposed in circular regions which reflect convex local arrangement of shape points. A priori global knowledge of the geometric structure of this region representation is captured by a statistical deformable template integrating a set of admissible deformations. The texture of each region is modeled separately, and we build a local Active Appearance Model for each region. A leave-one-out test was used to evaluate the performance of the proposed method and to compare it with conventional Active Appearance Model. The results demonstrate that the method is precise and very robust to large-scale noise present in radiographs, and that it can be useful in the context of preoperative planning of hip surgery.

We propose a new deformable model Deformable Associate Net (DAN). It is represented by a set of nodes which are associated by deformation constrains such as topology association, inter-part association, intra-part association, and geometry to atlas association. Each node in the model is given a priority, and hence DAN is a hierarchical model in which each layer is decided by nodes with same priority. Directional edges and dynamic generated local atlases are used in energy function to incorporate knowledge about tissue and image acquisition. A fast digital topology based method is designed to check whether topology of the model is changed under deformation. The deformation procedure hierarchically combines global and local deformations. Layers with high priority deform first. Once a higher layer is deformed to its target position in an image, the nodes in this layer are fixed, and then used as reference to help lower layers deform to their initial positions. At a particular layer, the model is first deformed by using global affine transformation to fit the image roughly, and then is warped by using a local deformation to fit the image better. The proposed method has been used to segment chest CT images for thoracic surgical planning, and it is also promising for other medical applications, such as model based image registration, and model-based 3D modeling.

This paper describes a method for automatic contour detection in reformatted short-axis (SA) cardiac computed tomography (CT) using a virtual exploring robot. The robot is a tricycle with a steering front wheel. Its motion obeys a set of kinematic equations and is subject to the non-holonomic constraints. The robot is designed to navigate in the binary representation of a cardiac image, consisting of the allowed navigational and obstacle spaces. It is initially positioned inside the allowed navigational space. Avoiding obstacles, the robot autonomously cruises through the navigational space and collects information about the location of the left ventricular (LV) boundaries. Consequently, the obtained information is used to reconstruct the endocardial and epicardial contours. Validation of the method was performed on in-vivo multislice multiphase short-axis cardiac CT images of ten subjects. Results showed good correlation between the quantitative parameters, computed from manual and automatic segmentation: for end-diastolic volume (EDV) r=0.99, for end-systolic volume (ESV) r=0.98, ejection fraction (EF) r=0.83, and LV mass (LVM) r=0.95.

The presence of speckle (a spatial stochastic process in an ultrasound image) makes ultrasound segmentation difficult. Speckle introduces local minima in the MAP energy function of an active contour, and when evolving under gradient descent, the contour gets trapped in a spurious local minimum. In this paper, we propose an alternate technique for evolving a MAP active contour. The technique has two parts: a deterministic evolution strategy called tunneling descent which escapes from spurious local minima, and a stopping rule for terminating the evolution. The combination gives an algorithm that is robust and gives good segmentations. The algorithm also benefits from having only a few free parameters which do not require tweaking. We present the conceptual framework of the algorithm in this paper, and study the impact of different stopping rules on the performance of the algorithm. The algorithm is used to segment the endocardium in cardiac ultrasound images. We present segmentation results in this paper and an experimental evaluation of different stopping rules on the performance of the algorithm. Although the algorithm is presented as an ultrasound segmentation technique, in fact, it can be used to segment any first-order texture boundary.

The efficacy of image processing and analysis on skull stripped MR images vis-a-vis the original images is well established. Additionally, compliance with the Health Insurance Portability and Accountability Act (HIPAA) requires neuroimage repositories to anonymise the images before sharing them. This makes the non-trivial skull stripping process all the more significant. While a number of optimal approaches exist to strip the skull from T1-weighted MR images to the best of our knowledge, there is no simple, robust, fast, parameter free and fully automatic technique to perform the same on T2-weighted images. This paper presents a strategy to fill this gap. It employs a fast parameterization of the T2 image intensity onto a standardized T1 intensity scale. The parametric "T1-like" image obtained via the transformation, which takes only a few seconds to compute, is subsequently processed by any of the many T1-based brain extraction techniques to derive the brain mask. Masking the original T2 image with this brain mask strips the skull. By standardizing the intensity of the parametric image, preset algorithm-specific parameters (if any) could be used across multiple datasets. The proposed scheme has been used in a number of phantom and clinical T2 brain datasets to successfully strip the skull.

In vivo studies of the relationships that exist among vascular geometry, plaque morphology, and hemodynamics have recently been made possible through the development of a system that accurately reconstructs coronary arteries imaged by x-ray angiography and intravascular ultrasound (IVUS) in three dimensions. Currently, the bottleneck of the system is the segmentation of the IVUS images. It is well known that IVUS images contain numerous artifacts from various sources. Previous attempts to create automated IVUS segmentation systems have suffered from either a cost function that does not include enough information, or from a non-optimal segmentation algorithm. The approach presented in this paper seeks to strengthen both of those weaknesses -- first by building a robust, knowledge-based cost function, and then by using a fully optimal, three-dimensional segmentation algorithm. The cost function contains three categories of information: a compendium of learned border patterns, information theoretic and statistical properties related to the imaging physics, and local image features. By combining these criteria in an optimal way, weaknesses associated with cost functions that only try to optimize a single criterion are minimized. This cost function is then used as the input to a fully optimal, three-dimensional, graph search-based segmentation algorithm. The resulting system has been validated against a set of manually traced IVUS image sets. Results did not show any bias, with a mean unsigned luminal border positioning error of 0.180 ± 0.027 mm and an adventitial border positioning error of 0.200 ± 0.069 mm.

Coupling the geodesic active contours model with statistical information based on regions introduces robustness in the segmentation of images with weak or inhomogeneous gradients. In the estimation of the probability density function for each region take part the definition of the features that describe the image inside the different regions and the method of density estimation itself. A Gaussian Mixture Model is frequently proposed for density estimation. This approach is based on the assumption that the intensity distribution of the image is the most discriminant feature in a region. However, the use of second order features provides a better discrimination of the different regions, as these features represent more accurately the local properties of the image manifold. Due to the high dimensionality of the problem, the use of non parametric density estimation methods becomes necessary. In this article, we present a novel method of introducing the second order information of an image for non parametric estimation of the probability density functions of the different tissues that are present in medical images. The novelty of the method stems on the use of the response of the image under an orthogonal harmonic operator set projected onto a prototype space for feature generation. The technique described here is applied to the segmentation of brain aneurysms in Computed Tomography Angiography (CTA) and 3D Rotational Angiography (3DRA) showing a qualitative improvement from the Gaussian Mixture Model approach.

Tumor vasculature has a high degree of irregularity as compared to normal vasculature. The quantification of the morphometric complexity in tumor images can be useful in diagnosis. Also, it is desirable in several other medical applications to have an automated complexity analysis to aid in diagnosis and prognosis under treatment. e.g. in diabetic retinopathy and in arteriosclerosis. In addition, prior efforts at segmentation of the tumor vasculature using matched filtering, template matching and splines have been hampered by the irregularity of these vessels. We try to solve both problems by introducing a novel technique for vessel detection, followed by a tracing-independent complexity analysis based on a combination of ideas. First, the vessel cross-sectional profile is modeled using a continuous and everywhere differentiable family of super-Gaussian curves. This family generates rectangular profiles that can accurately localize the vessel boundaries in microvasculature images. Second, a robust non-linear regression algorithm based on M-estimators is used to estimate the parameters that optimally characterize the vessel’s shape. A framework for the quantitative analysis of the complexity of the vasculature based on the vessel detection is presented. A set of measures that quantify the complexity are proposed viz. Squared Error, Entropy-based and Minimum Description Length-based Shape Complexities. They are completely automatic and can deal with complexities of the entire vessel unlike existing tortuousity measures which deal only with vessel centerlines. The results are validated using carefully constructed phantom and real image data with ground truth information from an expert observer.

We introduce a novel node-structural map method for digital image-based analysis of angiogenesis, and apply it to quantitate the effect of pro-angiogenic (e.g. VEGF) and anti-angiogenic (e.g. anti-VEGF) treatments on blood vessel patterns in the quail chorioallantoic membrane (CAM) in vivo model. After vessel segmentation and skeletonization, a node-structural map is derived to represent the structural property of each pixel in the skeletonized image, such as end nodes (root and leaf), branch and furcation/junction nodes. By combining the node-structural map with vascular thickness map, our proposed method provides more detailed morphometric and structural measurement of vascular angiogenesis, such as average diameter, average branch length, branch thickness distribution, density of branch and furcation nodes, etc. Furthermore, the concept of vascular generations will be proposed. Accordingly, more detailed measurements related to the generations of the vascular tree can be easily evaluated. From the quantitative analysis results, it correctly shows the fact that VEGF/anti-VEGF can modify the blood vessel pattern and increase/decrease the vessel density in CAM significantly, compared to the phosphate-buffered saline treated controls.

Segmentation of medical images is fundamental for many high-level applications. Unsupervised techniques such as region growing or merging allow automated processing of large data amounts. The regions are usually described by a mean feature vector, and the merging decisions are based on the Euclidean distance. This kind of similarity model is strictly local, since the feature vector of each region is calculated without evaluating the region's surrounding. Therefore, region merging often fails to extract visually comprehensible and anatomically relevant regions. In our approach, the local model is extended. Regional similarity is calculated for a pair of adjacent regions, e.g. considering the contrast along their common border. Global similarity components are obtained by analyzing the entire image partitioning before and after a hypothetical merge. Hierarchical similarities are derived from the iteration history. Local, regional, global, and hierarchical components are combined task-specifically guiding the iterative region merging process. Starting with an initial watershed segmentation, the process terminates when the entire image is represented as a single region. A complete segmentation takes only a few seconds. Our approach is evaluated contextually on plain radiographs that display human hands acquired for bone age determination. Region merging based on a local model fails to detect most bones, while a correct localization and delineation is obtained with the combined model. A gold standard is computed from ten manual segmentations of each radiograph to evaluate the quality of delineation. The relative error of labeled pixels is 15.7%, which is slightly more than the mean error of the ten manual references to the gold standard (12%). The flexible and powerful similarity model can be adopted to many other segmentation tasks in medical imaging.

The human lungs are divided into five distinct anatomic compartments called lobes. The physical boundaries between the lobes are called the lobar fissures. Detection of lobar fissure positions in pulmonary X-ray CT images is of increasing interest for the diagnosis of lung disease. We have developed an automatic method for segmentation of all five lung lobes simultaneously using a 3D watershed transform on the distance transform of a previously generated vessel mask, linearly combined with the original data. Due to the anatomically separate airway sub-trees for individual lobes, we can accurately and automatically place seed points for the watershed segmentation based on the airway tree anatomical description, due to the fact that lower generation airway and vascular tree segments are located near each other. This, along with seed point placement using information on the spatial location of the lobes, can give a close approximation to the actual lobar fissures. The accuracy of the lobar borders is assessed by comparing the automatic segmentation to manually traced lobar boundaries. Averaged over all volumes, the RMS distance errors for the left oblique fissure, right oblique fissure and right horizontal fissure are 3.720 mm, 0.713 mm and 1.109 mm respectively.

This paper describes methods for automatic localization of the mid-sagittal plane (MSP) and mid-sagittal surface (MSS). The data used is a subset of the Leukoaraiosis And DISability (LADIS) study consisting of three-dimensional magnetic resonance brain data from 62 elderly subjects (age 66 to 84 years). Traditionally, the mid-sagittal plane is localized by global measures. However, this approach fails when the partitioning plane between the brain hemispheres does not coincide with the symmetry plane of the head. We instead propose to use a sparse set of profiles in the plane normal direction and maximize the local symmetry around these using a general-purpose optimizer. The plane is parameterized by azimuth and elevation angles along with the distance to the origin in the normal direction. This approach leads to solutions confirmed as the optimal MSP in 98 percent of the subjects. Despite the name, the mid-sagittal plane is not always planar, but a curved surface resulting in poor partitioning of the brain hemispheres. To account for this, this paper also investigates an optimization strategy which fits a thin-plate spline surface to the brain data using a robust least median of squares estimator. Albeit computationally more expensive, mid-sagittal surface fitting demonstrated convincingly better partitioning of curved brains into cerebral hemispheres.

We propose a new fuzzy segmentation framework that incorporates the idea of super-resolution image reconstruction. The new framework is designed to segment data sets comprised of orthogonally acquired magnetic resonance (MR) images by taking into account their different system point spread functions. Formulating the reconstruction within the segmentation framework improves its robustness and stability, and makes it possible to incorporate multispectral scans that possess different contrast properties into the super-resolution reconstruction process. Our method has been tested on both simulated and real 3D MR brain data.

A level-set-based segmentation framework for Computer Aided Dental X-rays Analysis (CADXA) is proposed. In this framework, we first employ level set methods to segment the dental X-ray image into three regions: Normal Region (NR), Potential Abnormal Region (PAR), Abnormal and Background Region (ABR). The segmentation results are then used to build uncertainty maps based on a proposed uncertainty measurement method and an analysis scheme is applied. The level set segmentation method consists of two stages: a training stage and a segmentation stage. During the training stage, manually chosen representative images are segmented using hierarchical level set region detection. The segmentation results are used to train a support vector machine (SVM) classifier. During the segmentation stage, a dental X-ray image is first classified by the trained SVM. The classifier provides an initial contour which is close to the correct boundary for the coupled level set method which is then used to further segment the image. Different dental X-ray images are used to test the framework. Experimental results show that the proposed framework achieves faster level set segmentation and provides more detailed information and indications of possible problems to the dentist. To our best knowledge, this is one of the first results on CADXA using level set methods.

This paper presents a novel level set method for contour detection in multiple-object scenarios applied to the segmentation of teeth images. Teeth segmentation from 2D images of dental plaster cast models is a difficult problem because it is necessary to independently segment several objects that have very badly defined borders between them. Current methods for contour detection which only employ image information cannot successfully segment such structures. Being therefore necessary to use prior knowledge about the problem domain, current approaches in the literature are limited to the extraction of shape information of individual objects, whereas the key factor in such a problem are the relative positions of the different objects composing the anatomical structure. Therefore, we propose a novel method for introducing such information into a level set framework. This results in a new energy term which can be explained as a regional term that takes into account the relative positions of the different objects, and consequently creates an attraction or repulsion force that favors a determined configuration. The proposed method is compared with balloon and GVF snakes, as well as with the Geodesic Active Regions model, showing accurate results.

In this contribution, we propose a novel approach to the segmentation of tomographic image data considering topological properties of binarized image components expressed in terms of the Minkowski Functionals in 3D. Electron tomography is a non-invasive method for three-dimensional (3D) reconstruction of cellular sub-structures from a series of projection images (i.e. from a tilt series) recorded with a transmission electron microscope. Data obtained by electron tomography provide a rich source of quantitative information concerning the structural composition and organization of cellular components. It allows to obtain 3D information on structural cellular arrangements at a significantly higher resolution than any other of the currently available imaging modalities. A major challenge, in this context, is the segmentation of the image data with respect to the identification macro-molecular structures such as the actin-cytoskeleton or cell organelles. We introduce a morphological filtering algorithm based on the Minkowski Functionals in 3D for segmentation of macromolecular structures in intact eukaryotic cells depicted by cryo-electron tomography. In mathematical topology, multi-dimensional convex objects can be characterized with respect to shape, structure, and the connectivity of their components using a set of morphological descriptors known as the Minkowski functionals. In a 3D-Euclidian space, these correspond to volume, surface area, mean integral curvature, and the Euler-Poincare characteristic. The morphological filtering procedure is applied to a 3D image data of an intact, ice-embedded Dictyostelium cell obtained by low dose transmission electron microscopy using a tilt series of -50° to +41.5° with an increment of 1.5°. Our method allows to separate cellular components with predefined textural properties, e.g. filamentary or globular structures, from the image data, which may then be studied and interpreted further.

We have developed a method that uses a large amount of a priori information to generate super resolution radiographs. We measured and modeled analytically the point spread function of a low-dose gas microstrip x-ray detector at several beam energies. We measured the relationship between the local image intensity and the noise variance in the radiographs. The soft-tissue signal in the images was modeled using a minimum-curvature filtering technique. These results were then combined into an image deconvolution procedure using wavelet filtering to reduce restoration noise while keeping the enhanced small-scale features. The method was applied to a resolution grid image to measure its effects on the detector’s modulation transfer function. The restored images of a radiological human-torso phantom revealed small-scale details on the bones that were not seen before, and this, with improved SNR and image contrast. Dual-energy imaging was integrated to the restoration process in order to generate separate high-resolution images of the bones, the soft tissues, and the mean atomic number. This information could be used to detect bone micro-fractures in athletes and to assess bone demineralization in seniors due to osteoporosis. Super resolution radiographs are easier to segment due to their enhanced contrasts and uniform backgrounds; the boundaries of the features of interest can be delimited with a sub-pixel accuracy. This is highly relevant to the morphometric analysis of complex bone structures like individual vertebrae. The restoration method can be automated for a clinical environment use.

Many medical images suffer from the partial volume effect where a boundary between two structures of interest falls in the middle of a voxel giving a signal value that is a mixture of the two. We propose a method to restore the ideal boundary by splitting a voxel into sub-voxels and reapportioning the signal into the sub-voxels. We designed this method to correct MRI 2D slice images where partial volume can be a considerable limitation. Each voxel is divided into four (or more) sub-voxels by nearest neighbor interpolation. The gray level of each sub-voxel is considered as “materials” able to move between sub-voxels but not between voxels. A partial differential equation is written to allow the material to flow towards the highest gradient direction, creating a “reverse” diffusion process. Flow is subject to constraints that tend to create step edges. Material is conserved in the process thereby conserving MR signal. The method proceeds until the flow decreases to a low value. To test the method, synthetic images were down-sampled to simulate the partial volume artifact and restored. Corrected images were remarkably closer both visually and quantitatively to the original images than those obtained from common interpolation methods: on simulated data mean square errors were 0.35, 1.09, and 1.24 for the proposed method, bicubic, and bilinear interpolation respectively. The method was relatively insensitive to noise. On MRI physical phantom and brain images, restored images processed with the new method were visually much closer to high-resolution counter-parts than those obtained with common interpolation methods.

The data-driven Pixon noise-reduction method is applied to nuclear studies. By using the local information content, it preserves all statistically justifiable image features without generating artifacts. Statistical measures provide the user a feedback to judge if the processing parameters are optimal. The present work focuses on planar nuclear images with known Poisson noise characteristics. Its ultimate goals are to: (a) increase sensitivity for detection of lesions of small size and/or of small activity-to-background ratio, (b) reduce data acquisition time, and (c) reduce patient dose. Data are acquired using Data Spectrum’s cylinder phantom in two configurations: (a) with hot and cold rod inserts at varying total counts and (b) with hot sphere inserts at varying activity-to-background ratios. We show that the method adapts automatically to both hot and cold lesions, concentration ratios, and different noise levels and structure dimensions. In clinical applications, slight adjustment of the parameters may be needed to adapt to the specific clinical protocols and physician preference. Visually, the processed images are comparable to raw images with ~16 times as many counts, and quantitatively the reduced noise equals that obtained with ~50 times as many counts. We also show that the Pixon method allows for identification of spheres at low concentration ratios, where raw planar imaging fails and matched filtering underperforms. Conclusion: The Pixon method significantly improves the image quality of data at either reduced count levels, or low target-to-background ratios. An analysis of clinical studies is now warranted to assess the clinical impact of the method.

If the phase error at each pixel in a complex-valued MRI image is known the noise in the image can be reduced resulting in improved detection of medically significant details. However, given a complex-valued MRI image, estimating the phase error at each pixel is a difficult problem. Several approaches have previously been suggested including non-linear least squares fitting and smoothing filters. We propose a new scheme based on iteratively applying a series of non-linear filters, each used to modify the estimate into greater agreement with one piece of knowledge about the problem, until the output converges to a stable estimate. We compare our results with other phase estimation and MRI denoising schemes using synthetic data.

Image registration refers to the process of finding the spatial correspondence between two or more images. This is usually done by applying a spatial transformation, computed automatic or manually, to a given image using a continuous image model computed either with interpolation or approximation methods. We show that noise induced signal variance in interpolated images differs significantly from the signal variance of the original images in native space. We describe a simple approach to compute the signal variance in registered images based on the signal variance and covariance of the original images, the spatial transformations computed by the registration procedure, and the interpolation or approximation kernel chosen. Our approach is applied to diffusion tensor (DT) MRI data. We show that incorrect noise variance estimates in registered diffusion weighted images can affect the estimated DT parameters, their estimated uncertainty, as well as indices of goodness of fit such as chi-square maps. In addition to DT-MRI, we believe that this methodology would be useful any time parameter extraction methods are applied to registered or interpolated data.

When a patient is examined at different times using different protocols, how can we know whether the observed differences in the area or volume estimate are due to the patient, the protocol, or both? Specifically, we ask what is the smallest difference in lung volume that can be computed reliably when two sets of CT data are acquired by varying the number and thickness of the slices, but while holding constant the in-plane resolution. The accuracy and precision of the total lung volume estimates are calculated based on the principles of stereology using uniform design sampling. Comparisons of the lung volume estimate based on fewer slices using stereological principles are employed. A formal test made of the hypothesis that the use of fewer slices can yield satisfactory precision of the lung estimate. It is known that estimation of lung volume based on CT images is sensitive to the acquisition parameters used during scanning: dose, scan time, number of cross-sectional slices, and slice collimation. Those parameters are very different depending on the lung examination required: routine studies or high-resolution detailed studies. Thus, if different protocols are to be used confidently for volume estimation, it is important to understand the factors that influence volume estimate accuracy and to provide the associated confidence intervals for the measurements.

This paper introduces a method for constructing confidence intervals for possible clinical outcomes around the FROC curve of a mammography CAD system. Given the architecture of a CAD classifying machine, there is one and only one system threshold that will yield a desired sensitivity on a certain population. The limited training sample size leads to a sampling error and an uncertainty in determining the optimal system threshold. This leads to an uncertainty in the operating point in the direction along the FROC curve which can be captured by a Bayesian approach where the distribution of possible thresholds is estimated. This uncertainty contributes to a large and spread-out confidence interval which is important to consider when one is intending to make comparisons between CAD algorithms trained on different data sets. The method is validated using a Monte Carlo method designed to capture the effect of correctly determining the system threshold.

This paper presents a novel, completely unsupervised fMRI brain mapping approach that addresses the three problems of hemodynamic response function (HRF) shape variability, neural event timing, and fMRI response linearity. To make it robust, the method takes into account spatial and temporal information directly into the core of the activation detection process. In practice, activation detection is formulated in terms of temporal alignment between the sequence of hemodynamic response onsets (HROs) detected in the fMRI signal at υ and in the spatial neighbourhood of υ, and the sequence of "off-on" transitions observed in the input blocked stimulation paradigm (when considering epoch-related fMRI data), or the sequence of stimuli of the event-based paradigm (when considering event-related fMRI data). This multiple event sequence alignment problem, which comes under multisensor data fusion, is solved within the probabilistic framework of hidden Markov multiple event sequence models (HMMESMs), a special class of hidden Markov models. Results obtained on real and synthetic data compete with those obtained with the popular statistical parametric mapping (SPM) approach, but without necessitating any prior definition of the expected activation patterns, the HMMESM mapping approach being completely unsupervised.

Accurate quantitation of PET tracer uptake levels in small tumors remains a challenge. This work uses an improved reconstruction algorithm to reduce the quantitative errors due to limited system resolution and due to necessary image noise reduction. We propose a method for finding and using the detection system response in the projection matrix of a statistical reconstruction algorithm. In addition we use aligned anatomical information, available in PET/CT scanners, to govern the penalty term applied during each image update. These improvements are combined with FORE rebinning in a clinically feasible algorithm for reconstructing fully 3D PET data. Simulated results show improved tumor bias and variance characteristics with the new algorithm.

The accurate characterization of pulmonary airways on CT is potentially very useful for diagnosis and evaluation of lung diseases. The task is challenging due to their small size and variable orientation. We propose a probabilistic modeling technique and a set of measurement tools to quantitate airway morphology. We extract the airway tree structure from high resolution CT scans with a seeded region growing algorithm. Individual airway branches are identified by reducing the airway tree to a set of central axes. Properties such as lumen diameter and branch angle are measured from these central axes. The structure of the Bayesian model is inferred from a set of equations representing the parent-daughter relationships between branches, such as equations of air flow ratio and flow conservation. The CT measurements are used to instantiate the conditional probability tables of the Bayesian model. To evaluate the model, it was used to predict the airway diameter for the 2nd, 3rd, 4th, 5th, and 6th generations of the airway tree. We show that the model can reasonably predict the diameter of a particular airway branch, given information of its parent.

Object detection in ultrasound fetal images is a challenging task for the relatively low resolution and low signal-to-noise ratio. A direct inverse randomized Hough transform (DIRHT) is developed for filtering and detecting incomplete curves in images with strong noise. The DIRHT combines the advantages of both the inverse and the randomized Hough transforms. In the reverse image, curves are highlighted while a large number of unrelated pixels are removed, demonstrating a “curve-pass filtering” effect. Curves are detected by iteratively applying the DIRHT to the filtered image. The DIRHT was applied to head detection and measurement of the biparietal diameter (BPD) and head circumference (HC). No user input or geometric properties of the head were required for the detection. The detection and measurement took 2 seconds for each image on a PC. The inter-run variations and the differences between the automatic measurements and sonographers’ manual measurements were small compared with published inter-observer variations. The results demonstrated that the automatic measurements were consistent and accurate. This method provides a valuable tool for fetal examinations.

Accurate and reliable identification and quantification of vertebral fractures constitute a challenge both in clinical trials and in diagnosis of osteoporosis. Various efforts have been made to develop reliable, objective, and reproducible methods for assessing vertebral fractures, but at present there is no consensus concerning a universally accepted diagnostic definition of vertebral fractures. In this project we want to investigate whether or not it is possible to accurately reconstruct the shape of a normal vertebra, using a neighbouring vertebra as prior information. The reconstructed shape can then be used to develop a novel vertebra fracture measure, by comparing the segmented vertebra shape with its reconstructed normal shape. The vertebrae in lateral x-rays of the lumbar spine were manually annotated by a medical expert. With this dataset we built a shape model, with equidistant point distribution between the four corner points. Based on the shape model, a multiple linear regression model of a normal vertebra shape was developed for each dataset using leave-one-out cross-validation. The reconstructed shape was calculated for each dataset using these regression models. The average prediction error for the annotated shape was on average 3%.

In medical imaging, low signal-to-noise ratio (SNR) and/or contrast-to-noise ratio (CNR) often cause many image processing algorithms to perform poorly. Postacquisition image filtering is an important off-line image processing approach widely employed to enhance the SNR and CNR. A major drawback of many filtering techniques is image degradation by diffusing/blurring edges and/or fine structures. In this paper, we introduce a scale-based filtering method that employs scale-dependent diffusion conductance to perform filtering. This approach utilizes novel object scale information via a concept called generalized ball scale, which imposes no shape, size, or anisotropic constraints unlike previously published ball scale-based filtering strategies. The object scale allows us to better control the filtering process by constraining smoothing in regions with fine details and in the vicinity of boundaries while permitting effective smoothing in the interior of homogeneous regions. Quantitative evaluations based on the Brainweb data sets show superior performance of generalized ball scale-based diffusive filtering over two existing methods, namely, ball scale-based and nonlinear complex diffusion process. Qualitative experiments based on both phantom and patient magnetic resonance images demonstrate that the generalized ball scale-based approach leads to better preserved fine details and edges.

The rupture mechanism of intracranial aneurysms is still not fully understood. Although the size of the aneurysm is the shape index most commonly used to predict rupture, some controversy still exists about its adequateness as an aneurysm rupture predictor. In this work, an automatic method to geometrically characterize the shape of cerebral saccular aneurysms using 3D moment invariants is proposed. Geometric moments are efficiently computed via application of the Divergence Theorem over the aneurysm surface using a non-structured mesh. 3D models of the aneurysm and its connected parent vessels have been reconstructed from segmentations of both 3DRA and CTA images. Two alternative approaches have been used for segmentation, the first one based on isosurface deformable models, and the second one based on the level set method. Several experiments were also conducted to both assess the influence of pre-processing steps in the stability of the aneurysm shape descriptors, and to know the robustness of the proposed method. Moment invariants have proved to be a robust technique while providing a reliable way to discriminate between ruptured and unruptured aneurysms (Sensitivity=0.83, Specificity=0.74) on a data set containing 55 aneurysms. Further investigation over larger databases is necessary to establish their adequateness as reliable predictors of rupture risk.

The dependence on macro-optical imaging of the human body in the assessment of possible disease is rapidly increasing concurrent with, and as a direct result of, advancements made in medical imaging technologies. Assessing the pulmonary airways through bronchoscopy is performed extensively in clinical practice however remains highly subjective due to limited visualization techniques and the lack of quantitative analyses. The representation of 3D structures in 2D visualization modes, although providing an insight to the structural content of the scene, may in fact skew the perception of the structural form. We have developed two methods for visualizing the optically derived airway mucosal features whilst preserving the structural scene integrity. Shape from shading (SFS) techniques can be used to extract 3D structural information from 2D optical images. The SFS technique presented addresses many limitations previously encountered in conventional techniques resulting in high-resolution 3D color images. The second method presented to combine both color and structural information relies on combined CT and bronchoscopy imaging modalities. External imaging techniques such as CT provide a means of determining the gross structural anatomy of the pulmonary airways, however lack the important optically derived mucosal color. Virtual bronchoscopy is used to provide a direct link between the CT derived structural anatomy and the macro-optically derived mucosal color. Through utilization of a virtual and true bronchoscopy matching technique we are able to directly extract combined structurally sound 3D color segments of the pulmonary airways. Various pulmonary airway diseases are assessed and the resulting combined color and texture results are presented demonstrating the effectiveness of the presented techniques.

This paper presents a method for detecting suspicious nodules based on successive low-dose helical CT images. The method uses both initial and follow-up images to improve nodule detection performance. The basic idea of the detection is to register nodule images measured at different time and to assess the changes in size, shape, and density of the nodule. Since there are several variations of nodule changes, such as stable, shrinking, expansion in size, disappearance, appearance, and separation, a coarse-to-fine registration technique was adopted to deal with large nodule deformation. Especially, the fine registration is performed by excluding nodule regions and using nodule surroundings to avoid effects of nodule deformations in alignment task. In a preliminary experiment, the method was applied to ten cases with successive scans. From visual inspection, the corresponding results between initial and follow-up images were acceptable in clinical use. More researches using a large data set will be required. Still, we believe that the method has the potential of detecting suspicious nodules for use in a computer-aide diagnosis system.

This paper presents a method for automated detection of liver cancer regions based on transition of density at each point obtained from multi-phase X-ray CT images. For describing transition of density, two kinds of feature vectors named Density Transition (DT) and Density Change Transition (DCT) are introduced. DCT is used for extraction of cancer candidates and DT is used for suppression of false candidates. In the experiments using 14 real abdominal CT images with cancer, it was shown that the detection rate was 100% and the number of false-positives was 0.71 regions per case.

We are developing a computer-aided detection (CAD) system for lung nodules in thoracic CT volumes. During false positive (FP) reduction, the image structures around the identified nodule candidates play an important role in differentiating nodules from vessels. In our previous work, we exploited shape and first-order derivative information of the images by extracting ellipsoid and gradient field features. The purpose of this study was to explore the object shape information using second-order derivatives and the Hessian matrix to further improve the performance of our detection system. Eight features related to the eigenvalues of the Hessian matrix were extracted from a volume of interest containing the object, and were combined with ellipsoid and gradient field features to discriminate nodules from FPs. A data set of 82 CT scans from 56 patients was used to evaluate the usefulness of the FP reduction technique. The classification accuracy was assessed using the area A_z under the receiving operating characteristic curve and the number of FPs per section at 80% sensitivity. In the combined feature space, we obtained a test A_z of 0.97 ± 0.01, and 0.27 FPs/section at 80% sensitivity. Our results indicate that combining the Hessian, ellipsoid and gradient field features can significantly improve the performance of our FP reduction stage.

The purpose of this paper is to evaluate the post-operative Magnetic Resonance Perfusion (MRP) scans of patients undergoing carotid endarterectomy (CEA), using a novel image-analysis algorithm, to determine if post-operative neurocognitive decline is associated with cerebral blood flow changes. CEA procedure reduces the risk of stroke in appropriately selected patients with significant carotid artery stenosis. However, 25% of patients experience subtle cognitive deficits after CEA compared to a control group. It was hypothesized that abnormalities in cerebral blood flow (CBF) are responsible for these cognitive deficits. A novel algorithm for analyzing MR-perfusion (MRP) scans to identify and quantify the amount of CBF asymmetry in each hemisphere was developed and to quantify the degree of relative difference between three corresponding vascular regions in the ipsilateral and contralateral hemispheres, the Relative Difference Map (RDM). Patients undergoing CEA and spine surgery (controls) were examined preoperatively, and one day postoperatively with a battery of neuropsychometric (NPM) tests, and labeled “injured” patients with significant cognitive deficits, and “normal” if they demonstrated no decline in neurocognitive function. There are apparently significant RDM differences with MRP scans between the two hemispheres in patients with cognitive deficits which can be used to guide expert reviews of the imagery. The proposed methodology aids in the analysis of MRP parameters in patients with cognitive impairment.

We are developing a computer-aided detection (CAD) system to detect microcalcification clusters automatically on full field digital mammograms (FFDMs). The CAD system includes five stages: preprocessing, image enhancement and/or box-rim filtering, segmentation of microcalcification candidates, false positive (FP) reduction, and clustering. In this study, we investigated the performance of a nonlinear multiscale Laplacian pyramid enhancement method in comparison with a box-rim filter at the image enhancement stage and the use of a new error metric to improve the efficiency and robustness of the training of a convolution neural network (CNN) at the FP reduction stage of our CAD system. A data set of 96 cases with 200 images was collected at the University of Michigan. This data set contained 215 microcalcification clusters, of which 64 clusters were proven by biopsy to be malignant and 151 were proven to be benign. The data set was separated into two independent data sets. One data set was used to train and validate the CNN in our CAD system. The other data set was used to evaluate the detection performance. For this data set, Laplacian pyramid multiscale enhancement did not improve the performance of the microcalcification detection system in comparison with our box-rim filter previously optimized for digitized screen-film mammograms. With the new error metric, the training of CNN could be accelerated and the classification performance in validation was improved from an A_z value of 0.94 to 0.97 on average. The CNN in combination with rule-based classifiers could reduce FPs with a small tradeoff in sensitivity. By using the free-response receiver operating characteristic (FROC) methodology, it was found that our CAD system can achieve a cluster-based sensitivity of 70%, 80%, and 88% at 0.23, 0.39, and 0.71 FP marks/image, respectively. For case-based performance evaluation, a sensitivity of 80%, 90%, and 98% can be achieved at 0.17, 0.27, and 0.51 FP marks/image, respectively.

Osteoporotic fractures are associated with substantial morbidity, increased medical cost and high mortality risk. Several equipments of bone assessment have been developed to identify individuals, especially postmenopausal women, with high risk of osteoporotic fracture; however, a large segment of women with low skeletal bone mineral density (BMD), namely women with high risk of osteoporotic fractures, cannot be identified sufficiently because osteoporosis is asymptomatic. Recent studies have been demonstrating that mandibular inferior cortical width manually measured on panoramic radiographs may be useful for the identification of women with low BMD. Automatic measurement of cortical width may enable us to identify a large number of asymptomatic women with low BMD. The purpose of this study was to develop a computer-aided system for measuring the mandibular cortical width on panoramic radiographs. Initially, oral radiologists determined the region of interest based on the position of mental foramen. Some enhancing image techniques were applied so as to measure the cortical width at the best point. Panoramic radiographs of 100 women who had BMD assessments of the lumbar spine and femoral neck were used to confirm the efficacy of our new system. Cortical width measured with our system was compared with skeletal BMD. There were significant correlation between cortical width measured with our system and skeletal BMD. These correlations were similar with those between cortical width manually measured by the dentist and skeletal BMD. Our results suggest that our new system may be useful for mass screening of osteoporosis.

We present an automatic diagnosis method of liver cancer by using sequential images with contrast material of dynamic CT. Our method identifies and classifies liver tumors by extracting temporal change of CT values [Hounsfield Unit(HU)] of tumors from four kinds of CT images (i.e. plain CT, early phase, portal phase, late phase of dynamic CT images) in addition to morphological features of tumors. Automatic diagnosis of liver tumors is very difficult, because contrast of liver tumors is very small compared with liver background, shapes of tumors are diverse, and extraction of temporal change of CT values is very difficult due to morphological and contrast complexity of temporal change of tumor segments. Our method extracts temporal change of CT values of objects by mapping segments of same objects in different CT phase based on overlap ratio and position adjustment. We also implemented a graphical user interface for searching such images from an image database that include tumors similar to an image given as a search condition with respect to features of morphorogical and temporal change of contrast.

Microcalcification (MC) clusters in mammograms can be important early signs of breast cancer in women. Accurate detection of MC clusters is an important but challenging problem. In this paper, we propose the use of a recently developed machine learning technique -- relevance vector machine (RVM) -- for automatic detection of MCs in digitized mammograms. RVM is based on Bayesian estimation theory, and as a feature it can yield a decision function that depends on only a very small number of so-called relevance vectors. We formulate MC detection as a supervised-learning problem, and use RVM to classify if an MC object is present or not at each location in a mammogram image. MC clusters are then identified by grouping the detected MC objects. The proposed method is tested using a database of 141 clinical mammograms, and compared with a support vector machine (SVM) classifier which we developed previously. The detection performance is evaluated using the free-response receiver operating characteristic (FROC) curves. It is demonstrated that the RVM classifier matches closely with the SVM classifier in detection performance, and does so with a much sparser kernel representation than the SVM classifier. Consequently, the RVM classifier greatly reduces the computational complexity, making it more suitable for real-time processing of MC clusters in mammograms.

This paper presents a new algorithm for Gleason grading of pathological images of prostate. Structural features of the glands are extracted and used in a tree-structured (TS) algorithm to classify the images into five Gleason grades of 1 to 5. In this algorithm the image is first segmented to locate the glandular regions using texture features and a K-means clustering algorithm. The glands are then labeled from the glandular regions. In each stage of the proposed TS algorithm, shape and intensity-based features of the glands are extracted and used in a linear classifier to classify the image into two groups. Despite some proposed methods in the literature which use only texture features, this technique uses the features like roundness and shape distribution, which are related to the structure of the glands in each grade and are independent of the magnification. The proposed method is therefore robust to illumination and magnification variations. To evaluate the performance of the proposed method, we use two datasets. Data set 1 contains 91 images with similar magnifications and illuminations. Data set 2 contains 199 images with different magnifications and illuminations. Using leave-one-out technique, we achieve 95% and 85% accuracy for dataset 1 and 2, respectively.

Automatic identification of the pectoral muscle on MLO view is an essential step for computerized analysis of mammograms. It can reduce the bias of mammographic density estimation, will enable region-specific processing in lesion detection programs, and also may be used as a reference in image registration algorithms. We are developing a computerized method for the identification of pectoral muscle on mammograms. The upper portion of the pectoral edges was first detected to estimate the direction of the pectoral muscle boundary. A gradient-based directional (GD) filter was used to enhance the linear texture structures, and then a gradient-based texture analysis was designed to extract a texture orientation image that represented the dominant texture orientation at each pixel. The texture orientation image was enhanced by a second GD filter. An edge flow propagation method was developed to extract edges around the pectoral boundary using geometric features and anatomic constraints. The pectoral boundary was finally generated by a second-order curve fitting. 118 MLO view mammograms were used in this study. The pectoral muscle boundary identified on each image by an experienced radiologist was used as the gold standard. The accuracy of pectoral boundary detection was evaluated by two performance metrics. One is the overlap percentage between the computer-identified area and the gold standard, and the other is the root-mean-square (RMS) distance between the computer and manually identified pectoral boundary. For 118 MLO view mammograms, 99.2% (117/118) of the pectoral muscles could be identified. The average of the overlap percentage is 94.8% with a standard deviation of 20.9%, and the average of the RMS distance is 4.3 mm with a standard deviation of 5.9 mm. These results indicate that the pectoral muscle on mammograms can be detected accurately by our automated method.

The purpose of this study was to investigate factors that impact the generalization of breast cancer computer-aided diagnosis (CAD) systems that utilize the Breast Imaging Reporting and Data System (BI-RADS). Data sets from four institutions were analyzed: Duke University Medical Center, University of Pennsylvania Medical Center, Massachusetts General Hospital, and Wake Forest University. The latter two data sets are subsets of the Digital Database for Screening Mammography. Each data set consisted of descriptions of mammographic lesions according to the BI-RADS lexicon, patient age, and pathology status (benign/malignant). Models were developed to predict pathology status from the BI-RADS descriptors and the patient age. Comparisons between the models built on data from the different institutions were made in terms of empirical (non-parametric) receiver operating characteristic (ROC) curves. Results suggest that BI-RADS-based CAD systems focused on specific classes of lesions may be more generally applicable than models that cover several lesion types. However, better generalization was seen in terms of the area under the ROC curve than in the partial area index (>90% sensitivity). Previous studies have illustrated the challenges in translating a BI-RADS-based CAD system from one institution to another. This study provides new insights into possible approaches to improve the generalization of BI-RADS-based CAD systems.

In mesothelioma, response is usually assessed by computed tomography (CT). In current clinical practice the Response Evaluation Criteria in Solid Tumors (RECIST) or WHO, i.e., the uni-dimensional or the bi-dimensional measurements, is applied to the assessment of therapy response. However, the shape of the mesothelioma volume is very irregular and its longest dimension is almost never in the axial plane. Furthermore, the sections and the sites where radiologists measure the tumor are rather subjective, resulting in poor reproducibility of tumor size measurements. We are developing an objective three-dimensional (3D) computer algorithm to automatically identify and quantify tumor volumes that are associated with malignant pleural mesothelioma to assess therapy response. The algorithm first extracts the lung pleural surface from the volumetric CT images by interpolating the chest ribs over a number of adjacent slices and then forming a volume that includes the thorax. This volume allows a separation of mesothelioma from the chest wall. Subsequently, the structures inside the extracted pleural lung surface, including the mediastinal area, lung parenchyma, and pleural mesothelioma, can be identified using a multiple thresholding technique and morphological operations. Preliminary results have shown the potential of utilizing this algorithm to automatically detect and quantify tumor volumes on CT scans and thus to assess therapy response for malignant pleural mesothelioma.

In research and development of computer-aided differential diagnosis, there is now a widespread interest in the use of nodule doubling time for measuring the volumetric changes of pulmonary nodule. To assess nodule status requires not only the measurement of volume changes but also one of nodule density variations. This paper proposes a computerized approach to measure nodule density variation inside small pulmonary nodule using CT images. The approach consists of five steps: (1) nodule segmentation, (2) computation of CT density histogram, (3) nodule classification based on CT density histogram and size, (4) computation of doubling time based on CT density histogram, and (5) classification between benign and malignant. Our approach was applied to follow-up scans of lung nodules. The preliminary experimental result demonstrated that our approach has a highly potential usefulness to assess the nodule evolution using high-resolution CT images.

There is growing interest in computer aided diagnosis applications including automatic detection of lung nodules from multislice computed tomography (CT). However the increase in the number and size of CT datasets introduces high costs for data storage and transmission, and becomes an obstacle to routine clinical exam as well as hindering widespread utilization of computerized applications. We investigated the effects of 3D lossy region-based JPEG2000 standard compression on the results of an automatic lung nodule detection system. As the algorithm detects the lungs within the datasets, we used this lung segmentation to define a region of interest (ROI) where the compression should be of higher fidelity. We tested 4 methods of 3D compression: 1) default compression of the whole image, 2) default compression of segmented lungs with masking out all non-lung regions, 3) ROI-based compression as specified in the JPEG2000 standard and 4) compression where voxels in the ROI are weighted to be given emphasis in the encoding. We tested 7 compression ratios per method: 1, 4, 6, 8, 10, 20, and 30 to 1. We then evaluated our experimental CAD algorithm on 10 patients with 67 documented nodules initially identified on the decompressed data. Sensitivities and false positive rates were compared for the various compression methods and ratios. We found that region-based compression generally performs better than default compression. The sensitivity with default compression decreased from 85% at no compression to 61% at 30:1 compression, a decrease of 25%, whereas the masked compression method saw a decreased in sensitivity on only 13.5% at maximum compression. At compression levels up to 10:1, all 3 region-based compression methods had decreases in sensitivity of 7.5% or less. Detection of small nodules (< 4mm in diameter) was more affected by compression than detection of large nodules; sensitivity to calcified nodules was less affected by compression than to non-calcified nodules.

Achieving consistent and correct database cases is crucial to the correct evaluation of any computer-assisted diagnostic (CAD) paradigm. This paper describes the application of artificial intelligence (AI), knowledge engineering (KE) and knowledge representation (KR) to a data set of ≈2500 cases from six separate hospitals, with the objective of removing/reducing inconsistent outlier data. Several support vector machine (SVM) kernels were used to measure diagnostic performance of the original and a “cleaned” data set. Specifically, KE and ER principles were applied to the two data sets which were re-examined with respect to the environment and agents. One data set was found to contain 25 non-characterizable sets. The other data set contained 180 non-characterizable sets. CAD system performance was measured with both the original and “cleaned” data sets using two SVM kernels as well as a multivariate probabilistic neural network (PNN). Results demonstrated: (i) a 10% average improvement in overall Az and (ii) approximately a 50% average improvement in partial Az.

We present a method for detecting circumscribed masses in digital mammograms. Morphological hierarchical watersheds are used in the segmentation process. Oversegmentation is prevented by employing a reconstructive open/close alternating sequential filter to simplify the image. The advantage of this method of simplification is that the object shapes and edges are preserved. The regional maxima of the simplified input image are then extracted and used as internal markers for the hierarchical watershed transform, which is applied to the gradient image of the simplified input image. An image-based classification technique is applied to reduce the number of false positives. The method is applied to 18 mammograms from the MIAS database, containing 20 circumscribed masses in background tissue of varying density. We obtain a high true detection rate using combined with a low number of false positives per image.

The cardio-thoracic ratio (CTR) is commonly measured manually for the evaluation of cardiomegaly. To determine the CTR automatically, we have developed a computerized scheme based on gray-level histogram analysis and an edge detection technique with feature analysis. The database used in this study consisted of 392 chest radiographs, which included 304 normals and 88 abnormals with cardiomegaly. The pixel size and the quantization level of the image were 0.175 mm and 1024, respectively. We performed a nonlinear density correction to maintain consistency in the density and contrast of the image. Initial heart edge detection was performed by selection of a certain range of pixel values in the histogram of a rectangular area at the center of a low-resolution image. Feature analysis with use of an edge gradient and with the orientation obtained by a Sobel operator was applied for accurate identification of the heart edges, which tend to have large edge gradients in a certain range of orientations. In addition, to determine the CTR, we detected the ribcage edges automatically by using image profile analysis. In 94.9% of all of the cases, the heart edges were detected accurately by use of this scheme. The area under the ROC curve (A_z value) in distinguishing between normals and abnormals with cardiomegaly based on the CTR was 0.912. Because the CTR is measured automatically and quickly (in less than 1 sec.), radiologists could save reading time. The computerized scheme will be useful for the assessment of cardiomegaly on chest radiographs.

Computed Tomography (CT) has become the new reference standard for quantification of emphysema. The most popular measure for emphysema derived from CT is the Pixel Index (PI), which expresses the fraction of the lung volume with abnormally low intensity values. As PI is calculated from a single, fixed threshold on intensity, this measure is strongly influenced by noise. This effect shows up clearly when comparing the PI score for a high-dose scan to the PI score for a low-dose (i.e. noisy) scan of the same subject. This paper presents a class of noise filters that make use of a local noise estimate to specify the filtering strength: Local Noise Variance Weighted Averaging (LNVWA). The performance of the filter is assessed by comparing high-dose and low-dose PI scores for 11 subjects. LNVWA improves the reproducibility of high-dose PI scores: For an emphysema threshold of -910 HU, the root-mean-square difference in PI score drops from 10% of the lung volume to 3.3% of the lung volume if LNVWA is used.

Several studies have demonstrated the fractal properties of screening mammograms. The purpose of this study was to investigate fractal texture analysis for the automated detection of architectural distortion (AD) in screening mammograms. The study was based on the Digital Database for Screening Mammography (DDSM). Initially, a database of 708 mammographic regions with confirmed pathology was created. They were all 512x512 pixel regions of interest (ROIs). The ROI size was determined empirically. Fifty-two regions were extracted around biopsy-proven architectural distortion. The remaining 656 ROIs depicted normal breast parenchyma. Fractal analysis was performed on each ROI at multiple resolutions (512x512, 256x256, 128x128, and 64x64). The fractal dimension of each ROI was calculated using the circular average power spectrum technique. Overall, the average fractal dimension (FD) estimate of the normal ROIs was statistically significantly higher than the average FD of the ROIs with AD. This result was consistent across all resolutions. However, best detection performance was achieved when the fractal dimension was estimated on ROIs subsampled with a factor of 2 (ROC area index A_z=0.89±0.02). Specifically, there was perfect performance in fatty breasts (A_z=1.0), A_z=0.95±0.02 in fibroglandular breasts, A_z=0.84±0.05 in heterogeneous breasts, and A_z=0.66±0.10 in dense breasts. Overall, the present study demonstrates that the presence of AD disrupts the normal parenchymal structure, thus resulting in a lower fractal dimension. Consequently, fractal texture analysis could play an important role in the development of computer-assisted detection tools tailored towards architectural distortion.

The purpose of this study was to use objective quantitative MR imaging methods to develop a computer-aided diagnosis tool to differentiate white matter (WM) hyperintensities as either leukoencephalopathy (LE) or normal maturational processes in children treated for acute lymphoblastic leukemia with intravenous high dose methotrexate. A combined imaging set consisting of T1, T2, PD, and FLAIR MR images and WM, gray matter, and cerebrospinal fluid a priori maps from a spatially normalized atlas were analyzed with a neural network segmentation based on a Kohonen Self-Organizing Map. Segmented regions were manually classified to identify the most hyperintense WM region and the normal appearing genu region. Signal intensity differences normalized to the genu within each examination were generated for two time points in 203 children. An unsupervised hierarchical clustering algorithm with the agglomeration method of McQuitty was used to divide data from the first examination into normal appearing or LE groups. A C-support vector machine (C-SVM) was then trained on the first examination data and used to classify the data from the second examination. The overall accuracy of the computer-aided detection tool was 83.5% (299/358) with sensitivity to normal WM of 86.9% (199/229) and specificity to LE of 77.5% (100/129) when compared to the readings of two expert observers. These results suggest that subtle therapy-induced leukoencephalopathy can be objectively and reproducibly detected in children treated for cancer using this computer-aided detection approach based on relative differences in quantitative signal intensity measures normalized within each examination.

The purpose of this work was to evaluate an information-theoretic computer-aided detection (CAD) scheme for improving the specificity of mass detection in screening mammograms. The study was based on images from the Lumisys set of the Digital Database for Screening Mammography (DDSM). Initially, the craniocaudal views of 49 DDSM mammograms were analyzed using an automated detection algorithm developed to prescreen mammograms. The prescreening algorithm followed a morphological concentric layer analysis and resulted in 319 false positive detections at 92% sensitivity. These 319 suspicious yet normal regions were extracted for further analysis with our information-theoretic CAD scheme. Our scheme follows a knowledge-based decision strategy. The strategy relies on information theoretic principles for similarity assessment between a query case and a knowledge databank of cases with known ground truth. Receiver Operating Characteristic (ROC) analysis was performed to determine how well the CAD scheme can discriminate the false positive regions from 681 true masses. The overall ROC area index of the information-theoretic CAD system was 0.75±0.02. At 97%, 95%, and 90% sensitivity, the system eliminated safely 20%, 30%, and 42% of the previously identified false positives respectively. Thus, information-theoretic CAD analysis can yield a significant reduction in false-positive detections while maintaining reasonable sensitivity.

Fractures of bone are a common affliction. In most developed countries the number of fractures associated with age-related bone loss is increasing rapidly. Each year many fractures are missed during x-ray diagnosis, resulting in ineffective patient management and expensive litigation. From both an orthopaedic and radiologic point of view, the fully automatic detection and classification of fractures in long-bones is an important but difficult problem. In this paper, a fully automated method of detecting fractures in the diaphysis of a long-bone is described. X-rays are very difficult to process automatically, so to extract the required information a non-linear anisotropic diffusion method, the Affine Morphological Scale Space, was implemented to smooth the image without losing information about the location of boundaries within the image. Next, an iterative peak detection algorithm is used to accurately locate the bone centreline and articular surfaces. A method based on orthogonal projections calculated from a modified Hough transform is used to automatically locate the long-bone diaphysis. At this point, our algorithm accurately localises the area of the fracture, and would allow further image registration if necessary. Finally, a gradient-based algorithm is used to detect fractures present in the region of interest. The magnitude and direction of the gradient are combined to produce a measure of the likelyhood of the presence of a fracture. A library of long-bone fracture images was created. Experimental tests performed on a series of x-ray images show that the method is capable of accurately segmenting the diaphysis from the epiphyses, and is also able to detect many mid-shaft fractures of long-bones.

We have developed a computer-aided diagnostic (CAD) scheme for detection of unruptured intracranial aneurysms in magnetic resonance angiography (MRA) based on findings of short branches in vessel skeletons, and a three-dimensional (3D) selective enhancement filter for dots (aneurysms). Fifty-three cases with 61 unruptured aneurysms and 62 non-aneurysm cases were tested in this study. The isotropic 3D MRA images with 400 x 400 x 128 voxels (a voxel size of 0.5 mm) were processed by use of the dot enhancement filter. The initial candidates were identified not only on the dot-enhanced images by use of a multiple gray-level thresholding technique, but also on the vessel skeletons by finding short branches on parent skeletons, which can indicate a high likelihood of small aneurysms. All candidates were classified into four categories of candidates according to effective diameter and local structure of the vessel skeleton. In each category, a number of false positives were removed by use of two rule-based schemes and by linear discriminant analysis on localized image features related to gray level and morphology. Our CAD scheme achieved a sensitivity of 97% with 5.0 false positives per patient by use of a leave-one-out-by-patient test method. This CAD system may be useful in assisting radiologists in the detection of small intracranial aneurysms as well as medium-size aneurysms in MRA.

There is a lot of interest in developing computer-aided detection (CAD) techniques for mammography that use multiple view information. During the development of such techniques we have noticed that they are hampered by the phenomena that mass lesions are sometimes detected by multiple regions. This has encouraged us to develop a technique to regroup initial CAD detections to facilitate the final classification of suspicious regions. The regrouping technique searches for detections that belong to the same structure. Therefore, it takes into account the distance between the detections and the image structure along a path between the detections. When correspondence is found, the two detections are replaced by a new detection in between the initial detections. Our regrouping technique correctly regrouped the detections in 48 percent of the masses initially detected by multiple regions. Of the false positive detections two percent were combined, and the percentage of true positive - false positive combinations was one. Incorporation of the algorithm into our CAD scheme resulted in a slight increase in detection performance. In addition, in our multiple view scheme it also resulted in a decrease in the number of incorrectly linked regions in corresponding mammographic views.

Our group developed the computer aided diagnosis (CAD) system for lung cancer in 1996, and has been used in clinical field since 1997. From this CAD system (conventional system), we discovered problem and we attempted to solve the problem by using our proposed algorithm. The proposed algorithm succeeded in the improvement of the following three problems of the conventional system. (1) Weak extraction algorithm of region of interest (ROI) with noise, (2) Poor knowledge of chest structure, and (3) diagnostic processing for nodule of limited size. In this paper, the algorithm that solves problem (2) and (3) is described. We evaluated the proposed algorithm, which was applied to the following four databases. (A) Lung cancer database, (B) detailed examination database, (C) a large-scale screening database by 10mm-thickness images reconstructed from single-slice CT scan, and (D) a large-scale screening database by 10mm-thickness images reconstructed from multi-slice CT scan. The proposed method obtained the following successful results: Lung cancer database 95.7% TP and detailed examination 94.8% TP. For the large-scale screening database, we evaluated each examination process from physicians’ reading to cancer decision. The extraction rate of proposed algorithm improved as the examinations proceed. Two false positive results were obtained. False positive 1 (6.8-9.2 shadows/case) needed for a detailed examination and the object of false positive 2 (2.6-4.0 shadows/case) was an abnormal shadow.

Automated segmentation and classification of diagnostic markers in medical imagery are challenging tasks. Numerous algorithms for segmentation and classification based on statistical approaches of varying complexity are found in the literature. However, the design of an efficient and automated algorithm for precise classification of desired diagnostic markers is extremely image-specific. The National Library of Medicine (NLM), in collaboration with the National Cancer Institute (NCI), is creating an archive of 60,000 digitized color images of the uterine cervix. NLM is developing tools for the analysis and dissemination of these images over the Web for the study of visual features correlated with precancerous neoplasia and cancer. To enable indexing of images of the cervix, it is essential to develop algorithms for the segmentation of regions of interest, such as acetowhitened regions, and automatic identification and classification of regions exhibiting mosaicism and punctation. Success of such algorithms depends, primarily, on the selection of relevant features representing the region of interest. We present color and geometric features based statistical classification and segmentation algorithms yielding excellent identification of the regions of interest. The distinct classification of the mosaic regions from the non-mosaic ones has been obtained by clustering multiple geometric and color features of the segmented sections using various morphological and statistical approaches. Such automated classification methodologies will facilitate content-based image retrieval from the digital archive of uterine cervix and have the potential of developing an image based screening tool for cervical cancer.

Uterine cervical cancer is the second most common cancer among women worldwide. Colposcopy is a diagnostic method used to detect cancer precursors and cancer of the uterine cervix, whereby a physician (colposcopist) visually inspects the metaplastic epithelium on the cervix for certain distinctly abnormal morphologic features. A contrast agent, a 3-5% acetic acid solution, is used, causing abnormal and metaplastic epithelia to turn white. The colposcopist considers diagnostic features such as the acetowhite, blood vessel structure, and lesion margin to derive a clinical diagnosis. STI Medical Systems is developing a Computer-Aided-Diagnosis (CAD) system for colposcopy -- ColpoCAD, a complex image analysis system that at its core assesses the same visual features as used by colposcopists. The acetowhite feature has been identified as one of the most important individual predictors of lesion severity. Here, we present the details and preliminary results of a multi-level acetowhite region detection algorithm for RGB color images of the cervix, including the detection of the anatomic features: cervix, os and columnar region, which are used for the acetowhite region detection. The RGB images are assumed to be glare free, either obtained by cross-polarized image acquisition or glare removal pre-processing. The basic approach of the algorithm is to extract a feature image from the RGB image that provides a good acetowhite to cervix background ratio, to segment the feature image using novel pixel grouping and multi-stage region-growing algorithms that provide region segmentations with different levels of detail, to extract the acetowhite regions from the region segmentations using a novel region selection algorithm, and then finally to extract the multi-levels from the acetowhite regions using multiple thresholds. The performance of the algorithm is demonstrated using human subject data.

Screening programs produce large amount of mammographic data, and good compression schemes would be beneficial for both storage and transmission purposes. In medical data it is crucial that diagnostic important information is preserved. In this work we have implemented two different region-of-interest (ROI) coding methods together with a Set Partitioning in Hierarchical Trees (SPIHT) scheme to be used for compression of mammograms. Region-of-interest coding allows a region of the image to be compressed with higher fidelity than the rest of the image. This is useful in medical data to be able to compress a region containing a possibly cancer area with very high fidelity, but still manage an overall good compression ratio. Both the ROI methods, the basic SPIHT method as well as JPEG compression standard, the latter two without possibility of ROI coding, are evaluated by studying the results from a Computer Aided Detection (CAD) system for microcalcifications tested on the original and the compressed mammograms. In addition a visual inspection is performed as well as Peak Signal-to-Noise-Ratio (PSNR) calculations. Mammograms from the MIAS database is used. We show that mammograms can be compressed to less than 0.5 (0.3) bpp without any visual degradation and without significantly influence on the performance of the CAD system.

Many medical imaging techniques available today generate 4D data sets. One such technique is functional magnetic resonance imaging (fMRI) which aims to determine regions of the brain that are activated due to various cognitive and/or motor functions or sensory stimuli. These data sets often require substantial resources for storage and transmission and hence call for efficient compression algorithms. fMRI data can be seen as a time-series of 3D images of the brain. Many different strategies can be employed for compressing such data. One possibility is to treat each 2D slice independently. Alternatively, it is also possible to compress each 3D image independently. Such methods do not fully exploit the redundancy present in 4D data. In this work, methods using 4D wavelet transforms are proposed. They are compared to different 2D and 3D methods. The proposed schemes are based on JPEG2000, which is included in the DICOM standard as a transfer syntax. Methodologies to test the effects of lossy compression on the end result of fMRI analysis are introduced and used to compare different compression algorithms.

One of the problems in fluoroscopy based 3D seed reconstruction for prostate brachytherapy is the clustering of seeds in the fluoroscopic images. A template matching based method is proposed in this study to reconstruct the orientations and locations of individual seed images in the cluster. An idealized projection image of implanted seeds was used as a template to reconstruct the cluster, and different optimization strategies were implemented to find the best orientation and location for individual seeds. The four search methods compared were: 1) Down hill simplex method; 2) Powell method; 3) Multi-resolution based method with exhaustive initial search; and 4) Multi-resolution based method without exhaustive initial search. These methods were applied to 10 test images. Five of the ten images had only 2-seed clusters and five had 3-seed clusters. The results demonstrate that the first two methods didn’t perform well, and that the results were dependent on the initial guesses used to start the optimization process. The third method successfully found the best configurations for all of the 10 images while the fourth method succeeded in 9 of the 10 cases. We conclude therefore that multi-resolution approaches are appropriate for the seed image reconstruction problem. Since the possible configurations of the template are pre-computed, an additional advantage is that less execution time was needed for the multi-resolution methods. When applied to the seed image reconstruction process, this method will potentially significantly improve the accuracy of the 3D reconstruction of implanted seeds from fluoroscopic images used in prostate brachytherapy.

Magnetic resonance (MR) images acquired with a high temporal resolution or high spatial resolution are usually with a penalty of low signal to noise ratio (SNR). It is necessary to remove the noise artifacts with important image features such as edges preserved. In this paper, we propose to use the improved wavelet-based multiscale anisotropic diffusion algorithm for MR imaging. Experimental results demonstrate that this denoising algorithm can significantly improve the SNR for MR images while preserving edges with good visual quality. The denoising results indicate that in MR imaging applications, we can almost doubly improve the temporal resolution or improve the spatial resolution while achieving high SNR and acceptable image quality.

The influence of quantum mottle appears as degradation of graininess with reduction of the amount of incidence X-rays. The results of a Wiener spectrum study showed that graininess increased as the dose was reduced, and noise affected all frequencies. However, in clinical images, these effects are seen only in the high-frequency domain above 0.3 cycle/mm. Moreover, the effects of a grid are restricted to a parallel component or a perpendicular component based on its structure. And the influence appears in the decomposition wavelet image of H or V. From these result, the decomposed wavelet coefficients at complete binary tree are divided into seven frequency-coefficient bands. About the noise processing method, we tried to reduce noise by applying modification of Wavelet Transform Modules Maxima method proposed by Mallet, et al. Then we tried to the adaptive nonlinear threshold based on wiener spectrum study and power spectrum study. Based on the above considerations, evaluation was performed using clinical radiographs obtained at a standard dose and reduced dose with the noise reduction processing applied. The results showed that noise caused by quantum mottle and the grid can be reduced by this method without the need for threshold processing based on clinical experience.

A new method of noise reduction based on shrinkage in the wavelet domain has been created for the application in projection radiography. The method is based on comparing two similar or quasi-identical images of the same object. Using an appropriate measure of similarity, these images are compared with each other in order to produce the weighting matrices. The weighting factors for the wavelet coefficients are chosen to be proportional to the elements of the weighting matrices. One image of the pair is then reconstructed from the weighted wavelet coefficients. The effect of this kind of de-noising is a suppression of those structures in the image which don’t correlate with the structures in the other image of the pair. Normally the suppressed structures are quantum or scatter noise, while the correlated structures which are not affected at all, are the real anatomical structures.

Human movement analysis is generally performed through the utilization of marker-based systems, which allow reconstructing, with high levels of accuracy, the trajectories of markers allocated on specific points of the human body. Marker based systems, however, show some drawbacks that can be overcome by the use of video systems applying markerless techniques. In this paper, a specifically designed computer vision technique for the detection and tracking of relevant body points is presented. It is based on the Gauss-Laguerre Decomposition, and a Principal Component Analysis Technique (PCA) is used to circumscribe the region of interest. Results obtained on both synthetic and experimental tests provide significant reduction of the computational costs, with no significant reduction of the tracking accuracy.

For 2-dimensional B-mode ultrasound images, we propose an image enhancement algorithm based on a multi-resolution approach. In the proposed algorithm, we perform the directional filtering and noise reducing procedures from the coarse to fine resolution images that are obtained from the wavelet-transformed data. For directional filtering, the structural feature at each pixel is examined through the eigen-analysis. Then, if the pixel belongs to the edge region, we perform two-step directional filtering, namely, directional smoothing along the tangential direction of the edge to improve its continuity, and directional sharpening along the normal direction to enhance the contrast. Meanwhile, speckle noise is alleviated by reducing the wavelet coefficients corresponding to the homogeneous region. The reducing rate of the wavelet coefficients is determined by considering the frequency characteristics of speckle. Thereby, the algorithm reduces speckle noise efficiently without affecting the edge sharpness and enhances edges regardless their size. Note that the proposed speckle reduction scheme is based on the structural information rather than the statistics of the magnitude of wavelet coefficients as in the existing methods. The proposed algorithm is compared to the algorithm based on nonlinear anisotropic diffusion filtering and the one based on the wavelet shrinkage scheme. The experimental results show that the proposed algorithm considerably improves the subjective image quality without providing any noticeable artifact.

Due to the increasing demand for high-quality ceramic crowns and bridges, the CAD/CAM-based production of dental restorations has been a subject of intensive research during the last fifteen years. A prerequisite for the efficient processing of the 3D measurement of prepared teeth with a minimal amount of user interaction is the automatic determination of the preparation line, which defines the sealing margin between the restoration and the prepared tooth. Current dental CAD/CAM systems mostly require the interactive definition of the preparation line by the user, at least by means of giving a number of start points. Previous approaches to the automatic extraction of the preparation line rely on single contour detection algorithms. In contrast, we use a combination of different contour detection algorithms to find several independent potential preparation lines from a height profile of the measured data. The different algorithms (gradient-based, contour-based, and region-based) show their strengths and weaknesses in different clinical situations. A classifier consisting of three stages (range check, decision tree, support vector machine), which is trained by human experts with real-world data, finally decides which is the correct preparation line. In a test with 101 clinical preparations, a success rate of 92.0% has been achieved. Thus the combination of different contour detection algorithms yields a reliable method for the automatic extraction of the preparation line, which enables the setup of a turn-key dental CAD/CAM process chain with a minimal amount of interactive screen work.

The aim of this study is to provide feature images of digital mammograms in which regions corresponding to masses are enhanced. Subsequently, the feature images can be segmented and classified into two classes; masses and normal tissue. Our proposed feature extraction method is based on a local energy measure as texture feature. The local energy measure is extracted using a filter optimized with respect to the relative distance between the average feature values. In order to increase the sensitivity of the texture feature extraction scheme each mammogram is preprocessed using wavelet transformation, adaptive histogram equalization, and a morphology based enhancement technique. Initial experiments indicate that our scheme is able to provide useful feature images of digital mammograms. In order to quantify the system performance the feature images of 38 mammograms from the MIAS database -- 19 containing circumscribed masses, and 19 containing spiculated masses -- were segmented using simple gray level thresholding. For the circumscribed masses a true positive (TP) rate of 89% with a corresponding 2.3 false detections (false positives, FPs) per image was achieved. For the spiculated masses the performance was somewhat lower, yielding an overall TP rate of 84% with a corresponding 2.6 FPs per image.

We present a quantitative method for the comparison of vascular topology and geometry measured from retinal fundus photographs. The measure compares the difference between distributions taken from a graph representation of the vasculature, which is derived by image segmentation. The measure uses the Kullback-Leibler distance between statistical measures on the reference and test segmentations which can be geometrical, like the distribution of vessel widths, or topological, like local connectivity or a combination of the two. The user is free to build any meaningful description and here we illustrate two local topology measures graphically. Using this assessment method, we also show that our model based segmentation method has better geometrical accuracy than a technique based on matched filtering. We have tested out the measures on a set of 20 images from the STARE project data.

Understanding gene expression in the mouse brain should provide a better understanding of the underlying topology of the mammalian brain, thereby opening previously unexplored avenues in neuroscience and brain informatics. An important step in this direction is to develop robust algorithms to quantify gene expression in in-situ hybridization (ISH) data. ISH methodology involves the use of labeled nucleic acid probes that bind to specific mRNA transcripts in tissue sections. The bound probe is detected using colorimetric methods and the resulting stained tissue sections are imaged at high resolution. The goals of the present study are to first identify a staining method that produces maximum signal to noise ratio (SNR), and second to develop a method for gene expression detection for a wide range of ISH images spanning different intensity and expression patterns. A simple k-means based clustering method is used to separate foreground labeling from background and non-expressing tissues in a variety of images stained with different staining/counter staining techniques. We found NBT/BCIP with no counterstain produces the best signal to background separation. The foreground cluster detected using the k-means algorithm was further modeled using a normal distribution. A novel one sided Mahalanobis distance based metric with majority partial ordered voting method was then developed to generate a fuzzy segmentation of the gene expression in each ISH image. This algorithm is fully automatic and facilitates high throughput analysis of large amount of image data. Using this methodology a cluster of 10 PCs was able to process approximately 10% of the mouse genome (17 TBytes of JPEG2000 lossless compressed images) over a period of 2 weeks. The results may be visualized at the Allen Institute for Brain Science web site www.brain-map.org.

Automated medical image processing and analysis offer a powerful tool for medical diagnosis. In this work, a decision-tree based white blood cell (WBC) classification scheme for peripheral blood images is developed. Based on the sufficient analysis on the characteristics of white blood cells, 10 efficient features are extracted, including size, shape, intensity and color, and a classification scheme based on decision-tree is designed to classify 6 different types of normal white blood cells. Especially, an efficient approach to separate two types of neutrophil is presented. The presented scheme is tested on 59 WBCs coming from 3 sets of blood images, which are obtained under different dying and imaging conditions. Results show classification accuracy above 96%.

To be clinically viable, computer-aided diagnosis (CAD) systems must be as automated and user-friendly as possible. CAD systems for breast ultrasound are still preliminary and are not adapted for use in a standard clinical environment. For example, computer detection and classification schemes need the pixel size of each image to operate correctly, and while the DICOM standard allows pixel size to be encoded in the image file, some equipment manufacturers neglect to utilize the encoding. As a result, the pixel size is calculated from user input. In order to increase clinical efficiency and reduce the likelihood of error due to incorrect image specifications, automating this input process is a highly desirable asset. We developed and applied a character recognition algorithm to the annotation region of each ultrasound image in our database. A set of numerical masks, which corresponded to the characters used in the annotation information, enabled the filtering of each image. Numerical masks yielding the maximum output from the comparison operation between image data and mask were output to obtain the annotation information. Each image was then automatically cropped to remove the annotation banner and leave only the image data. The cropped image matrix dimensions and character recognition output were used to determine the corresponding pixel size. The algorithm was tested on 1110 images with various pixel sizes. In every case, the value output by the algorithm corresponded exactly to the true value. Our recognition algorithm now allows for the clinical translation of our fully-automated breast ultrasound CAD system.

Signal intensity in magnetic resonance images (MRIs) is affected by random noise. Assessing noise-induced signal variance is important for controlling image quality. Knowledge of signal variance is required for correctly computing the chi-square value, a measure of goodness of fit, when fitting signal data to estimate quantitative parameters such as T1 and T2 relaxation times or diffusion tensor elements. Signal variance can be estimated from measurements of the noise variance in an object- and ghost-free region of the image background. However, identifying a large homogeneous region automatically is problematic. In this paper, a novel, fully automated approach for estimating the noise-induced signal variance in magnitude-reconstructed MRIs is proposed. This approach is based on the histogram analysis of the image signal intensity, explicitly by extracting the peak of the underlining Rayleigh distribution that would characterize the distribution of the background noise. The peak is extracted using a nonparametric univariate density estimation like the Parzen window density estimation; the corresponding peak position is shown here to be the expected signal variance in the object. The proposed method does not depend on prior foreground segmentation, and only one image with a small amount of background is required when the signal-to-noise ratio (SNR) is greater than three. This method is applicable to magnitude-reconstructed MRIs, though diffusion tensor (DT)-MRI is used here to demonstrate the approach.

Image intensity standardization is a post-acquisition processing operation designed for correcting acquisition-to-acquisition signal intensity variations (non-standardness) inherent in Magnetic Resonance (MR) images. While existing standardization methods based on histogram landmarks have been shown to produce a significant gain in the similarity of resulting image intensities, their weakness is that, in some instances the same histogram-based landmark may represent one tissue, while in other cases it may represent different tissues. This is often true for diseased or abnormal patient studies in which significant changes in the image intensity characteristics may occur. In an attempt to overcome this problem, in this paper, we present two new intensity standardization methods based on the concept of generalized scale. In reference 1 we introduced the concept of generalized scale (g-scale) to overcome the shape, topological, and anisotropic constraints imposed by other local morphometric scale models. Roughly speaking, the g-scale of a voxel in a scene was defined as the largest set of voxels connected to the voxel that satisfy some homogeneity criterion. We subsequently formulated a variant of the generalized scale notion, referred to as generalized ball scale (g_B-scale), which, in addition to having the advantages of g-scale, also has superior noise resistance properties. These scale concepts are utilized in this paper to accurately determine principal tissue regions within MR images, and landmarks derived from these regions are used to perform intensity standardization. The new methods were qualitatively and quantitatively evaluated on a total of 67 clinical 3D MR images corresponding to four different protocols and to normal, Multiple Sclerosis (MS), and brain tumor patient studies. The generalized scale-based methods were found to be better than the existing methods, with a significant improvement observed for severely diseased and abnormal patient studies.

We present a novel semi-automated registration technique for 3D volume reconstruction from fluorescent laser scanning confocal microscope (LSCM) imagery. The developed registration procedure consists of (1) highlighting segmented regions as salient feature candidates, (2) defining two region correspondences by a user, (3) computing a pair of region centroids, as control points for registration, and (4) transforming images according to estimated transformation parameters determined by solving a set of linear equations with input control points. The presented semi-automated method is designed based on our observations that (a) an accurate point selection is much harder for a human than an accurate region (segment) selection, (b) a centroid selection of any region is less accurate by a human than by a computer, and (c) registration based on structural shape of a region rather than on intensity-defined point is more robust to noise and to morphological deformation of features across stacks. We applied the method to image mosaicking and image alignment registration steps and evaluated its performance with 20 human subjects on LSCM images with stained blood vessels. Our experimental evaluation showed significant benefits of automation for 3D volume reconstruction in terms of achieved accuracy, consistency of results and performance time. In addition, the results indicate that the differences between registration accuracy obtained by experts and by novices disappear with an advanced automation while the absolute registration accuracy increases.

The purpose of this study was to extend previous work to detect migration of total wrist arthroplasty non-invasively, and with greater accuracy. Two human cadaverous arms, each with a cemented total wrist implant, were used in this study. In one of the arms, 1 mm tantalum balls were implanted, six in the carpal bones and five in the radius. Five CT scans of each arm were acquired, changing the position of the arm each time to mimic different positions patients might take on repeated examinations. Registration of CT volume data sets was performed using an extensively validated, 3D semi-automatic volume fusion tool in which co-homologous point pairs (landmarks) are chosen on each volume to be registered. Three sets of ten cases each were obtained by placing landmarks on 1) bone only (using only arm one), 2) tantalum implants only, and 3) bone and tantalum implants (both using only arm two). The accuracy of the match was assessed visually in 2D and 3D, and numerically by calculating the distance difference between the actual position of the transformed landmarks and their ideal position (i.e., the reference landmark positions). All cases were matched visually within one width of cortical bone and numerically within one half CT voxel (0.32 mm, p = 0.05). This method matched only the bone/arm and not the prosthetic component per se, thus making it possible to detect prosthetic movement and wear. This method was clinically used for one patient with pain. Loosening of the carpal prosthetic component was accurately detected and this was confirmed at surgery.

SPECT (Single Photon Emission Computed Tomography) imagery has become widely available and is particularly useful for regional cerebral blood flow (rCBF) studies. Distribution of rCBF is still essentially studied by visual observation, searching for abnormalities, and comparing with other studies. In order to facilitate the localization of these abnormalities, we propose a simple, automatic and direct method to register a SPECT rCBF study with a commonly used atlas in the neurological community, the Talairach Atlas. The Talairach atlas still gives today the most extensive information of regions of interests, coupled with a coordinate system. The proposed method will therefore allow a physician to precisely navigate in a SPECT image by interpreting the abnormalities coordinates. The registration of these two volumes is carried out in two steps, a rough alignment followed by an elastic registration. The rough alignment step consists in computing the mass centroid of each volume and in scaling the volumes accordingly if necessary. A simple threshold method (30% of the maximum intensity of the SPECT image) is used to determine the volume of the brain being studied. In order to facilitate the fine registration, the Talairach atlas was previously segmented in three classes: cerebrospinal fluid (CSF), white and gray matters. Then, an automatic intensity transformation as well as a low-pass filtering is performed to closely resemble the spatial resolution and intensities of the SPECT volume. This intensity transformation is a simple method which combines the use of a joint 2D histogram of the segmented atlas and the individual volume as well as a clustering algorithm. The fine registration is then computed with an optical flow methodology. The effectiveness of this scheme was tested on a database of virtual patients, simulated from a database of 45 healthy and diseased brains. The rate of pixels misclassification in each class within a one pixel neighborhood (CSF 0.5%; white matter 1.37%, gray matter 2.80%) indicates that this proposed method will be useful for the nuclear physician in helping localize abnormalities.

A conventional SPECT image of lung is obtained by accumulating the detected count of gamma rays over long acquisition time that contains many respiratory cycles. The lung motion due to respiration during the acquisition makes reconstructed image blurred and may lead to a misdiagnosis. If a respiratory-gated SPECT is used, reconstructed images at various phase of respiration are obtained and the blur in an image can be avoided. However, the respiratory-gated SPECT requires long time to accumulate sufficient number of counts at each phase. If the acquisition time is not long enough, the detected count becomes inadequately small and hence the reconstructed image becomes noisy. We propose a method for correcting the motion between different phase images obtained with the respiratory-gated SPECT. In this method, an objective function consisting of both the degree of similarity between a reference and a deformed image and the smoothness of deformation is defined and optimized. The expansion ratio defined as a ratio of the change of the local volume due to the deformation is introduced to preserve the total activity during the motion correction process. By summing each phase images corrected by this method, a less noisy and less blurred SPECT image can be obtained. Furthermore, this method allows us to analyze the local movement of lung. This method was applied to the computer phantom, the real phantom and some clinical data and the motion correction and visualization of local movements between inspiration and expiration phase images were successfully achieved.

In this paper we propose a new method of elastic registration of anatomical structures that bears an inner skeleton, such as the knee, hand or spine. Such a method has to deal with great degrees of variability, specially for the case of inter-subject registration; but even for the intra-subject case the degree of variability of images will be large since the structures we bear in mind are articulated. Rigid registration methods are clearly inappropriate for this problem, and well-known elastic methods do not usually incorporate the restriction of maintaining long skeletal structures straight. A new method is therefore needed to deal with such a situation; we call this new method "articulated registration". The inner bone skeleton is modeled with a wire model, where wires are drawn by connecting landmarks located in the main joints of the skeletal structure to be registered (long bones). The main feature of our registration method is that within the bone axis (specifically, where the wires are) an exact registration is guaranteed, while for the remaining image points an elastic registration is carried out based on a distance transform (with respect to the model wires); this causes the registration on long bones to be affine to all practical purposes, while the registration of soft tissue -- far from the bones -- is elastic. As a proof-of-concept of this method we describe the registration of hands on radiographs.

Registration and segmentation are two most important problems in the field of medical image analysis. Traditionally, they were treated as separate problems. In this paper, we introduce a unified variational framework for simultaneously carrying out image segmentation and registration. Segmentation information is integrated into the process of registration in leading to a more stable and noise-tolerant shape evolution, while a diffusion model is used to infer the volumetric deformation across the image. One of the major advantages of our model is its robustness against image noise. We present several 2D examples on synthetic and real data.

Optimization is an important component in linear and nonlinear medical image registration. While common non-derivative approaches such as Powell's method are accurate and efficient, they cannot easily be adapted for parallel hardware. In this paper, new optimization strategies are proposed for parallel, shared-memory (SM) architectures. The Dividing Rectangles (DIRECT) global method is combined with the local Generalized Pattern Search (GPS) and Multidirectional Search (MDS) and to improve efficiency on multiprocessor systems. These methods require no derivatives, and can be used with all similarity metrics. In a multiresolution framework, DIRECT is performed with relaxed convergence criteria, followed by local refinement with MDS or GPS. In 3D-3D MRI rigid registration of simulated MS lesion volumes to normal brains with varying noise levels, DIRECT/MDS had the highest success rate, followed by DIRECT/GPS. DIRECT/GPS was the most efficient (5-10 seconds with 8 CPUs, and 10-20 seconds with 4 CPUs). DIRECT followed by MDS or GPS greatly increased efficiency while maintaining accuracy. Powell's method generally required more than 30 seconds (1 CPU) with a low success rate (0.3 or lower). This work indicates that parallel optimization on shared memory systems can markedly improve registration speed and accuracy, particularly for large initial misorientations.

Mutual information is currently one of the most widely used image similarity measures for multimodality image registration. An important step in the calculation of the mutual information of two images is the estimation of their joint histogram. Most algorithms use lateral joint histogram sizes that are smaller than the actual number of intensity levels present in the images being registered. Using a reduced joint histogram size is especially useful when registering small portions of the images to obtain local deformations in nonrigid registration algorithms, and when implementing hardware solutions for acceleration of mutual information calculation. The most commonly used method for reducing the size of the joint histogram is to perform a linear rescaling of intensity values. The main problem with this method is that image regions with similar intensity values but corresponding to distinct tissue types tend to merge, thus compromising the accuracy of registration. We present new algorithms for reducing the number of gray levels present in 3D medical images, and compare their performance with previously reported ones. The tested algorithms are classified in three categories: histogram shape preserving algorithms, entropy maximization algorithms and quantization error minimization algorithms. Results show that in CT and MRI registration the best accuracy is achieved using entropy maximization algorithms, while in PET and MRI registration the best accuracy is achieved using histogram shape preservation algorithms.

The implementation of Magnetic Resonance Spectroscopic Imaging (MRSI) for diagnostic imaging benefits from close integration of the lower-spatial resolution MRSI information with information from high-resolution structural MRI. Since patients can commonly move between acquisitions, it is necessary to account for possible mis-registration between the datasets arising from differences in patient positioning. In this paper we evaluate the use of 4 common multi-modality registration criteria to recover alignment between high resolution structural MRI and 3D MRSI data of the brain with sub-voxel accuracy. We explore the use of alternative MRSI water reference images to provide different types of structural information for the alignment process. The alignment accuracy was evaluated using both synthetically created MRSI and MRI data and a set of carefully collected subject image data with known ground truth spatial transformation between image volumes. The final accuracy and precision of estimates were assessed using multiple random starts of the registration algorithm. Sub voxel accuracy was found by all four similarity criteria with normalized mutual information providing the lowest target registration error for the 7 subject images. This effort supports the ongoing development of a database of brain metabolite distributions in normal subjects, which will be used in the evaluation of metabolic changes in neurological diseases.

Registration of medical images, i.e. the integration of two or more images into a common geometrical system of reference so that corresponding image structures correctly align, is an active field of current research. Registration algorithms in general are composed of three main building blocks: a geometrical transformation is applied in order to transform the images into the geometrical system of reference, a similarity measure puts the comparison of the images into quantifiable terms, and an optimization algorithm searches for that transformation that leads to optimal similarity between the images. Whereas in the literature fixed configurations of registration algorithms are investigated, here we present a modular toolbox containing several similarity measures, transformation classes and optimization strategies. Derivative-free optimization is applicable for any similarity measure, but is not fast enough in clinical practice. Hence we consider much faster derivative-based Gauss-Newton and Levenberg-Marquardt optimization algorithms that can be used in conjunction with frequently needed similarity measures for which derivatives can be easily obtained. The implemented similarity measures, geometrical transformations and optimization methods can be freely combined in order to configure a registration algorithm matching the requirements of a particular clinical application. Test examples show that particular algorithm configurations out of this toolbox allow e.g. for an improved lesion identification and localization in PET-CT or MR registration applications.

Introduction: Indocyanin green fundus angiographic images (ICGA) of the eyes is useful method in detecting and characterizing the choroidal neovascularization (CNV), which is the major cause of the blindness over 65 years of age. To investigate the quantitative analysis of the blood flow on ICGA, systematic approach for automatic registration of using mutual information and a quantitative analysis was developed. Methods: Intermittent sequential images of indocyanin green angiography were acquired by Heidelberg retinal angiography that uses the laser scanning system for the image acquisition. Misalignment of each image generated by the minute eye movement of the patients was corrected by the mutual information method because the distribution of the contrast media on image is changing throughout the time sequences. Several region of interest (ROI) were selected by a physician and the intensities of the selected region were plotted according to the time sequences. Results: The registration of ICGA time sequential images is required not only translate transform but also rotational transform. Signal intensities showed variation based on gamma-variate function depending on ROIs and capillary vessels show more variance of signal intensity than major vessels. CNV showed intermediate variance of signal intensity and prolonged transit time. Conclusion: The resulting registered images can be used not only for quantitative analysis, but also for perfusion analysis. Various investigative approached on CNV using this method will be helpful in the characterization of the lesion and follow-up.

Joint histogram is the only quantity required to calculate the mutual information (MI) between two images. For MI based image registration, joint histograms are often estimated through linear interpolation or partial volume interpolation (PVI). It has been pointed out that both methods may result in a phenomenon known as interpolation induced artifacts. In this paper, we implemented a wide range of interpolation/approximation kernels for joint histogram estimation. Some kernels are nonnegative. In this case, these kernels are applied in two ways as the linear kernel is applied in linear interpolation and PVI. In addition, we implemented two other joint histogram estimation methods devised to overcome the interpolation artifact problem. They are nearest neighbor interpolation with jittered sampling with/without histogram blurring and data resampling. We used the clinical data obtained from Vanderbilt University for all of the experiments. The objective of this study is to perform a comprehensive comparison and evaluation of different joint histogram estimation methods for MI based image registration in terms of artifacts reduction and registration accuracy.

In image registration of medical data a common and challenging problem is handling intensity-inhomogeneities. These inhomogeneities appear for instance in images of serially sectioned brains caused by the histological staining process or in medical imaging with contrast agents. Beneath this, natural outliers (for instance cells or vessels) produced by the underlying material itself may be mistaken as noise. Both image registration applications have in common that the well known sum of squared differences (SSD) measure would detect false differences. To deal with these kinds of problems, we supplement the common SSD-measure with image derivatives of higher order. Additionally we introduce a non-quadratic penalizer function to the distance measure leading to robust energy. The concepts are well known in optical flow. Overall, we present a variational model which combines all of these properties. This formulation leads to a fast and efficient algorithm. We demonstrate its applicability at the problems described above.

Prostate cancer has been ranked as the second leading cause of cancer death in men. The existence of cancer in prostate is usually examined by a biopsy procedure under the transrectal ultrasound (TRUS) guidance. Development of a prostate biopsy robotics can alleviate urologists' labor and guarantee accuracy. However, it is usually impossible to identify cancer region in the noisy ultrasound images, thus leading to a random biopsy protocol for prostate. It is being recognized that Magnetic Resonance Imaging (MRI) and Magnetic Resonance Spectroscopy Imaging (MRSI) techniques are potential to diagnose cancer distribution in prostate. So navigating the biopsy needle towards those cancer-suspected sites could improve the cancer detection rate and reduce the possibility of false negative diagnosis results. As the prostate usually deforms under the different rectal filling of probes and change of patient postures, a deformable registration scheme is implemented for the integration of the pre-operative MRI/MRSI information with the intra-operative TRUS images. A framework including a global rigid alignment and a sequent non-rigid transformation was described in this paper to match the cross-modal prostate surfaces and thereafter their volumes. For validation, an elastic prostate phantom that simulated the human condition was built up, with fiducial markers implanted inside the phantom prostate as the "ground truth". It shows that our method can achieve at least 30% improvement in accuracy compared with an affine transformation. Preliminary study was also conducted on patient data but with visual assessments presented only due to the current lack of "ground truth".

In this paper, we study image description based on compound mutual information (CMI) and its application to multimodal image registration. CMI is an aggregate information measure derived from the multiple marginal densities of image distributions. It extends histogram-based mutual information to that based on various marginal densities, which encode spatial intensity distribution and other image characteristics. Therefore, CMI can overcome the difficulties (such as the lack of spatial information) inherent in the color histogram, enrich the vocabulary of image description, and help improve registration accuracy and robustness. CMI is not sensitive to illumination and absolute appearance, and it is particularly suited for multimodal applications.

The use of affine image registration based on normalized mutual information (NMI) has recently been proposed by Frangi et al. as an automatic method for assessing brachial artery flow mediated dilation (FMD) for the characterization of endothelial function. Even though this method solves many problems of previous approaches, there are still some situations that can lead to misregistration between frames, such as the presence of adjacent vessels due to probe movement, muscle fibres or poor image quality. Despite its widespread use as a registration metric and its promising results, MI is not the panacea and can occasionally fail. Previous work has attempted to include spatial information into the image similarity metric. Among these methods the direct estimation of α-MI through Minimum Euclidean Graphs allows to include spatial information and it seems suitable to tackle the registration problem in vascular images, where well oriented structures corresponding to vessel walls and muscle fibres are present. The purpose of this work is twofold. Firstly, we aim to evaluate the effect of including spatial information in the performance of the method suggested by Frangi et al. by using α-MI of spatial features as similarity metric. Secondly, the application of image registration to long image sequences in which both rigid motion and deformation are present will be used as a benchmark to prove the value of α-MI as a similarity metric, and will also allow us to make a comparative study with respect to NMI.

Spatial normalization is a very important step in the processing of magnetic resonance imaging (MRI) data. So the quality of brain templates is crucial for the accuracy of MRI analysis. In this paper, using the classical protocol and the optimized protocol plus nonlinear deformation, we constructed the T1 whole brain templates and apriori brain tissue data from 69 Chinese pediatric MRI data (age 7-16 years). Then we proposed a new assessment method to evaluate our templates. 10 pediatric subjects were chosen to do the assessment as the following steps. First, the cerebellum region, the region of interest (ROI), was located on both the pediatric volume and the template volume by an experienced neuroanatomist. Second, the pediatric whole brain was mapped to the template with affine and nonlinear deformation. Third, the parameter, derived from the second step, was used to only normalize the ROI of the child to the ROI of the template. Last, the overlapping ratio, which described the overlapping rate between the ROI of the template and the normalized ROI of the child, was calculated. The mean of overlapping ratio normalized to the classical template was 0.9687, and the mean normalized to the optimized template was 0.9713. The results show that the two Chinese pediatric brain templates are comparable and their accuracy is adequate to our studies.

Standard Mutual Information function contains local maxima, which make against to convergence of registration transformation parameters for automated multimodality image registration problems. We proposed Feature Potential Mutual Information (FPMI) to increases the smoothness of the registration measure function and use Particle Swarm Optimization to search the optimal registration transformation parameter in this paper. At first, Edges of images are detected. Next, edge feature potential is defined by expanding edges to the neighborhood region using potential function. Each edge point influences the whole potential field, just like the particle of physics in the gravitation field space. FPMI is computed on the edge feature potential of two images. It substitutes the edge feature potential values for gray values in images. It can avoid great change of joint probability distribution and has less local maxima. The registration accuracy of FPMI is analyzed under different edge detection cases. It is shown that the registration accuracy of FPMI is more accurate and more robust than that of MI. Maximization of FPMI is done by PSO. PSO combines local search methods with global search methods, attempting to balance exploration and exploitation. Its complex behavior follows from a few simple rules and has less computational complexity. Multimodal medical images are used to compare the response of FPMI and MI to translation and rotation. Experiments show that FPMI is smoother and has less local fluctuations than that of MI. Registration results show that PSO do it better than Powell’s method to search the optimal registration parameters.

Medical image fusion is increasingly enhancing diagnostic accuracy by synergizing information from multiple images, obtained by the same modality at different times or from complementary modalities such as structural information from CT and functional from PET. An active, crucial research topic in fusion is validation of the registration (point-to-point correspondence) used. Phantoms and other simulated studies are useful in the absence of, or as a preliminary to, definitive clinical tests. Software phantoms in specific have the added advantage of robustness, repeatability and reproducibility. Our virtual-lung-phantom-based scheme can test the accuracy of any registration algorithm and is flexible enough for added levels of complexity (addition of blur/anti-alias, rotate/warp, and modality-associated noise) to help evaluate the robustness of an image registration/fusion methodology. Such a framework extends easily to different anatomies. The feature of adding software-based fiducials both within and outside simulated anatomies prove more beneficial when compared to experiments using data from external fiducials on a patient. It would help the diagnosing clinician make a prudent choice of registration algorithm.

Pixel intensities of MR images reconstructed by Fourier Transform and Projection methods have been proved to be spatially asymptotically independent (S.A.I.) and to have exponential correlation coefficient (E.C.C.). Based on S.A.I. and E.C.C., the MR image has been proved to be embedded in an MRF with respect to a proper neighborhood system. Further, the MR image is proved to be modeled by a Finite Normal Mixture (FNM) with an MRF as its prior. A unified Expectation-Maximization (EM) algorithm is presented for performing image segmentation. S.A.I., E.C.C. and Markovianity provide means for selecting the order of neighborhood systems and the values of clique potentials. The use of the 3rd-order neighborhood system and the correlation coefficient-based assignments of clique potentials strike an optimal trade-off between good accuracy and sufficient simplicity in image segmentation.

Landmarking MR images is crucial in registering brain structures from different images. It consists in locating the voxel in the image that corresponds to a well-defined point in the anatomy, called the landmark. Example of landmarks are the apex of the head (HoH) of Hippocampus, the tail and the tip of the splenium of the corpus collosum (SCC). Hand landmarking is tedious and time-consuming. It requires an adequate training. Experimental studies show that the results are dependent on the landmarker and drifting with time. We propose a generic algorithm performing automated detection of landmarks. The first part consists in learning from a training set of landmarked images the parameters of a probabilistic model, using the EM algorithm. The second part inputs the estimated parameters and a new image, and outputs a voxel as a predicted location for the landmark. The algorithm is demonstrated on the HoH and the SCC. In contrast with competing approaches, the algorithm is generic: it can be used to detect any landmark, given a hand-labeled training set of images.

In this paper we seek to improve upon the standard method of assessing the degree of calcification in the lumbar aorta, which is commonly used on lateral 2-D x-rays. The necessity for improvement arises from the fact that the existing method can not measure subtle progressions in the plaque development; neither is it possible to express the density of individual plaques. Both of these qualities would be desireable to assess, since they are the key for making progression studies as well as for testing the effect of drugs in longitudinal studies. Our approach is based on inpainting, a technique used in image restoration as well as postprocessing of film. In this study we discuss the potential implications of total variation inpainting for characterizing aortic calcification.

This paper presents a study of the texture information of high-resolution FLAIR images of the brain with the aim of determining the abnormality and consequently the candidacy of the hippocampus for temporal lobe epilepsy (TLE) surgery. Intensity and volume features of the hippocampus from FLAIR images of the brain have been previously shown to be useful in detecting the abnormal hippocampus in TLE. However, the small size of the hippocampus may limit the texture information. High-resolution FLAIR images show more details of the abnormal intensity variations of the hippocampi and therefore are more suitable for texture analysis. We study and compare the low and high-resolution FLAIR images of six epileptic patients. The hippocampi are segmented manually by an expert from T1-weighted MR images. Then the segmented regions are mapped on the corresponding FLAIR images for texture analysis. The 2-D wavelet transforms of the hippocampi are employed for feature extraction. We compare the ability of the texture features from regular and high-resolution FLAIR images to distinguish normal and abnormal hippocampi. Intracranial EEG results as well as surgery outcome are used as gold standard. The results show that the intensity variations of the hippocampus are related to the abnormalities in the TLE.

The goal of this work is to find a new parametric model of three-dimensional (3D) bone micro-architecture which is both representative of some bone characteristics and can be used as a prior model in tomographic reconstruction or image segmentation. For this purpose, Markov Random Fields (MRFs) are adapted. Bone micro-architecture is typically organized as a pseudo-periodic arrangement of parallel plates with a given thickness (≈ 100 μm), and a given spacing (≈ 700 μm). As standard MRFs could not represent the complexity of bone, we investigated appropriated energy functions, parameters and a design of the neighborhood system including local and distant interactions. Then, parameters of the new model were related to the standard morphometric parameters of the bone microarchitecture using an usual bone quantification tool.