This PDF file contains the front matter associated with SPIE Proceedings Volume 12088 including the Title Page, Copyright information, and Table of Contents.

Neuromuscular function is typically evaluated by recording the surface electromyogram (sEMG) in different motor tasks. Linear stochastic signal analysis is largely employed to extract information from sEMG and motor unit activities, even though there are well known nonlinearities in the neuromuscular system. Measures based on information theory have been used in the neuroscience. These measures are model-independent that can capture both linear and nonlinear features of time series. Here, we used Shannon’s entropy to quantify the neural information present in (1) the single motor unit spike trains, (2) compound spike trains (a population signal), and (3) sEMG envelope (a gross measure of muscle electrical activity). Participants performed isometric force control tasks in two contraction intensities (2.5% and 5% of maximum voluntary contraction). We also calculated the correlation (linear regression model) between the entropy and either the coefficient of variation of muscle force or mean motor unit firing rate. Results showed that entropy increases with force level and it is higher for the spike trains. In addition, spike train entropy has moderate correlation with the mean motor unit firing rate. The results are quite promising, and they show that measures from information theory can be applied to signals from motor unit activities to differentiate contraction intensities and, with further analysis, to understand the nonlinear properties of the neuromuscular system.

In this paper, we present a new non-contact strategy to estimate the Peripheral Oxygen Saturation (SPO₂) based on the Eulerian motion video magnification technique and a signal processing technique, The magnification procedure was carried out using two approaches : the Hermite decomposition and the Gaussian decomposition. The SpO₂ is estimated from the signals extracted after magnification process using the red and the blue Chanel of the frame. We have tested the method on five healthy subjects using videos obtained from the google-meet video conference platform. To compare the performance of the methods, we compute the mean average error and metrics issues from the Bland and Altman analysis to investigate the agreement of the methods with respect to a contact pulse oximeter device as reference. The proposed solution shows an agreement with respect to the reference of most of 98%

For the gait phases characterization and for the gait classification (slow, normal, fast, upstairs, downstairs), it is of fundamental importance to determine the Heel-Strike and Toe-Off gait events, which correspond to the moment when the ankle hits the ground, and to the instant when the toes leave the ground, respectively. These two events allow calculating various quantitative parameters of the gait, such as number of steps and cadence, and qualitative parameters, such as the symmetry indices. In this work, we present a method for detecting the above-mentioned events, using simple LSTM neural network architectures and windows filtering through a rule set. Mean accuracies obtained for Heel-Strike and Toe-Off 97 events, respectively, were 98,50% and 97,47%.

Motor disorders in Parkinson’s Disease (PD) show high inter-individual variability which challenges the current observational-based strategies in the clinical setting to determine the actual disease evolution and monitoring the therapy response. In spite the recent development of the motion capture technology, it is still hardly transferable to the routine exam and the non-linear disease patterns are poorly explored. Because gait pattern could be approached as deterministic chaotic system, this work aimed to non-linearly represent lower limb kinematic standing out the differences among PD stages. For doing so, a widely used deep learning framework was implemented for obtaining the body landmarks and their temporal series and thereafter, construing the phase space based on the first order derivatives. Largest Lyapunov exponent, correlation dimension and approximate entropy were then computed resulting in statistically significant differences (Wilcoxon rank test, p < 0.05), particularly between healthy controls and stages 3, the most advanced stage, and comparing stage 1 face to stage 3. These finding providing insights how the complex patterns may be related with the disease progression in PD and easily implemented using affordable video devices.

Parkinson’s disease is a neurodegenerative disease that affects more than 6.1 million people worldwide. In the clinical routine, the main tool to diagnose and monitor disease progression is based on motor impairments, such as postural instability, bradykinesia, tremor, among others. Besides, new biomarkers based on motion patterns have emerged to describe disease findings. Nonetheless, this motor characterization has low sensitivity, especially at early stages, and is largely expert-dependent, because protocols are mainly based on visual observations. However, most of these analyses require complex and some invasive systems that additionally only bring global information of complete recordings. This work introduces a multimodal approach that integrates gait and eye motion videos to quantify and predict patient stage on-the-fly. This method starts by computing dense apparent velocity maps that represent the local displacement of the person seen from the gait in a sagittal plane and as micro-movements during the fixation experiment. Then, each frame is described as a covariance descriptor of deep feature activation maps computed over the motion field at each video time. Then, the covariance video manifold is mapped to a recurrent LSTM network to learn higher non-local dependencies and quantify a motion descriptor. Also, an end-to-end scheme allows to lately fuse both modalities (gait and fixational eye) to obtain a more sensitive Parkinson disease descriptor. In a study with 25 subjects, the proposed approach reaches an average F1-score of 0.83 with an average recall of 0.78. In a temporal prediction analysis, the approach reports major correlations with the disease considering swing phase.

The constant increase in the volume of data generated by various medical modalities has generated discussions regarding the space needed for storage. Although the storage and network bandwidth costs are decreasing, medical data production grows faster, thus forcing an increase in spending. With the application of image compression and decompression techniques, such expectations and challenges overcoming can preserve all clinically relevant information. This research proposes and evaluates a method that combines an adaptive normalization for each DICOM slice and volume compression using a video CODEC. Similarity metrics results show that the best result achieved for these tests was the method combines the normalization function and the H264 using as parameters FPS 60, bitrate 120, with an image in PNG format, where SSIM and CC have the maximum value, PSNR has a high value and CR higher than competitors in the literature.

In 2020, COVID-19 emerged, and several health units faced problems with the demand for chest Computed Tomography (CT), one of the ways of detecting lung lesions. One of the initial processes for treating CT is lung segmentation. Several computational methods have been proposed since they would accelerate the patient’s diagnosis process. This paper proposes a method for lung segmentation on CT in patients affected by COVID-19 using deep learning and adaptive convex hull. In this work, two databases were used: one with 38 and the other with 72 patients, without and with COVID-19, respectively. The promising results may assist in future work to detect pneumonia.

The Coronavirus Disease 2019 (COVID-19) pandemic that affects the world since 2020 generated a great amount of research interest in how to provide aid to medical staff on triage, diagnosis, and prognosis. This work proposes an automated segmentation model over Computed Tomography (CT) scans, segmenting the lung and COVID-19 related lung findings at the same time. Manual segmentation is a time-consuming and complex task, especially when applied to high-resolution CT scans, resulting in a lack of gold standards annotation. Thanks to data provided by the RadVid19 Brazilian initiative, providing over a hundred annotated High Resolution CT (HRCT), we analyze the performance of three convolutional neural networks for the segmentation of lung and COVID findings: a 3D UNet architecture; a modified EfficientDet (2D) architecture; and 3D and 2D variations of the MobileNetV3 architecture. Our method achieved first place in the RadVid19 challenge, among 13 other competitors’ submissions. Additionally, we evaluate the model with the best result on the challenge in four public CT datasets, comparing our results against other related works, and studying the effects of using different annotations in training and testing. Our best method achieved on testing upwards of 0.98 Lung and 0.73 Findings 3D Dice and reached state-of-the-art performance on public data.

The COVID-19 corona virus has claimed 4.1 million lives, as of July 24, 2021. A variety of machine learning models have been applied to related data to predict important factors such as the severity of the disease, infection rate and discover important prognostic factors. Often the usefulness of the findings from the use of these techniques is reduced due to lack of method interpretability. Some recent progress made on the interpretability of machine learning models has the potential to unravel more insights while using conventional machine learning models.1–3 In this work, we analyze COVID-19 blood work data with some of the popular machine learning models; then we employ state-of-the-art post-hoc local interpretability techniques(e.g.- SHAP, LIME), and global interpretability techniques(e.g. - symbolic metamodeling) to the trained black-box models to draw interpretable conclusions. In the gamut of machine learning algorithms, regressions remain one of the simplest and most explainable models with clear mathematical formulation. We explore one of the most recent techniques called symbolic metamodeling to find the mathematical expression of the machine learning models for COVID-19. We identify Acute Kidney Injury (AKI), initial Albumin level (ALB I), Aspartate aminotransferase (AST I), Total Bilirubin initial (TBILI) and D-Dimer initial (DIMER) as major prognostic factors of the disease severity. Our contributions are - (i) uncover the underlying mathematical expression for the black-box models on COVID-19 severity prediction task (ii) we are the first to apply symbolic metamodeling to this task, and (iii) discover important features and feature interactions. Code repository: https://github.com/ananyajana/interpretable covid19.

Deep learning (DL) is now widely used to perform tasks involving the analysis of biomedical imaging. However, the small amounts available of annotated examples of these types of images make it difficult to use DL-based systems, since large amounts of data are required for adequate generalization and performance. For this reason, in recent years, Generative Adversarial Networks (GANs) have been used to obtain synthetic images that artificially increase the amount available. Despite this, the usual training instability in GANs, in addition to their empirical design, does not always allow for high-quality results. Through the neuroevolution of GANs it has been possible to reduce these problems, but many of these works use benchmark datasets with thousands of images, a scenario that does not reflect the real conditions of cases in which it is necessary to increase the data due to the limited amount available. In this work, cDCGAN-PSO is presented, an algorithm for the neuroevolution of GANs that adapts the concepts of the DCGAN-PSO to a conditional-DCGAN that allows the synthesis of three classes of chest X-ray images and that is trained with only 600 images of each class. The synthetic images obtained from evolved GANs show good similarity with real chest X-ray images.

Panoramic X-rays are an essential tool to assist dentistry experts in their diagnostic procedures. Dentists can analyze the anatomical and pathological structures while planing orthodontic, periodontal, and surgical treatments. Even though detecting, numbering, and segmenting teeth are essential tasks to leverage automatic analysis on panoramic X-rays, it is lacking in the literature a study and a data set that considers at the same time deciduous and permanent teeth in a wide variety of panoramic X-rays. To fill this gap, this work introduces a novel, challenging, and high-variable public data set labeled from scratch. This data set incorporates new elements such as instance overlapping and deciduous teeth, supporting our study on tooth numbering and segmentation. Our efforts aim to improve the segmentation on the boundaries because they are the main hurdle of the instance segmentation methods. For that, we investigate and compare (quantitatively and qualitatively) two Mask R-CNN-based solutions: the standard one, with a fully convolutional network, and another one that employs the PointRend module on the top. Our findings attest to the feasibility of extending segmentation and numbering to deciduous teeth through end-to-end deep learning architectures, as well as, the higher performance of the Mask R-CNN with PointRend either on instance segmentation (mAP of +2 percentage points) or the numbering (mAP of +1.2 percentage points) on the test data set. We hope that our findings and our new data set support the development of new tools to assist professionals in faster diagnosis, making upon panoramic X-rays.

Given a public database of medical images, testing and comparing different neural network algorithms using that database without a region of interest is a challenging task. This work aims to use the CBIS-DDSM (Curated Breast Imaging Subset of the DDSM - Database for Screening Mammography) and InBreast database, to test a lightweight model of convolutional neural networks (CNN) to segment the masses of tumors. The proposed model is a reduced version of the “U-Net” architecture to be lighter and faster to segment images in computers of routine use of health professionals. The proposed method, available in Google Colaboratory format for easy replication and modifications, can achieve competitive results comparing to the default “U-Net”. This work is a reproducible tool and does not achieve state of art results that uses other methods, but can be customized to enhance accuracy. Results showed that the model can predict tumors masks of both Medial Lateral Oblique (MLO) and Craniocaudal (CC) cases. We define a premise using region of interest to define what is a true positive and with that premise the model achieved a mean dice coefficient of 0.60 and a mean accuracy of 0.40. With test CPU hardware the model can predict 32 images per second, with dedicated GPU the model can predict 237 images per second.

Mammography is the main exam used to perform early detection of breast cancer. Using digital mammographic images processing, it’s possible to improve visualization of structures and facilitate diagnosis, using noise-reduction filters, contrast enhancement and segmentation. Segmentation divides the image in regions of interest (lesions) and background. Watershed segmentation performs a topographic analysis of the image and object’s boundaries extraction. Watershed demands using markers, that limit the regions where processing will be applied, avoiding issues such as over-segmentation. The markers can be defined using morphological operators, which use structuring elements to identify interest pixels. Therefore, the aim of this work is to evaluate watershed segmentation of breast lesions, by changing the pre-processing and structuring elements. There were applied disk and diamond-shaped structuring elements, with 25, 50 and 100 pixels as size, in addition to noise-reduction Wiener filter and contrast enhancement with Wavelet transform and contrast limited adaptive histogram equalization (CLAHE). The results were evaluated by using accuracy, ROC curve and visual analysis of the images, allowing to more easily compare the efficiency of each processing. For the disk format obtained AUC equal to 0.891, while the diamond format has 0.909, it was possible to identify the relationship between size and pre-processing with the number of detected regions. Moreover, small structuring elements are more effective in watershed segmentation, requiring research with different denoising methods and marker definition algorithms. Finally, the results allowed us to establish how breast density and shape, size and type of lesion interfere in watershed segmentation results.

Breast cancer is the second leading cause of cancer death in women worldwide. Ultrasound is one of the most used tools for image-based assessment of this disease, it helps to discriminate between benign or malignant masses, however, it depends largely on the radiologists experience. In recent years there has been a special interest to develop automatic segmentation systems (radiologist independent) for ultrasound images. The main challenges to get a clear ultrasound segmentation are: speckle noise, low contrast, blurred edges, shadows, etc. In this work, we use the Small Tumor Aware Network (STAN) architecture to automatically segment the images and we validate the network using eleven metrics on five combined breast ultrasound datasets. Dice’s coefficient indicates that predicted segmentations are 84% similar to the ground truth while high recall, specificity and accuracy results are being obtained. We use this information to calculate the tortuosity to differentiate between malignant and benign lesions, the Wilcoxon test results are p < 0.05 with a z = 2.73.

Quantitative MRI (qMRI) has been shown to be crucial for assessing organ dysfunction in the body. Usually, in qMRI approaches, a few metrics are extracted to distinguish normal and abnormal tissues. In this study, we coupled four MRI protocols (mDIXON T1, T1 and T2 mapping and DTI) to obtain 34 complementary metrics including 20 shape metrics, 2 texture metrics and 12 water diffusivity metrics for thigh muscle analysis. These metrics were calculated on both thighs to detect a pathological difference between a pair of right and left muscles. The method is based on a dimension reduction method and a projection of shape and diffusivity metrics into a three-dimensional linear latent space, along with two texture metrics. 5 healthy individuals (10 thighs, each thigh 7 muscles, i.e., 4 flexors and 3 extensors) were scanned to provide the reference scores. The developed pipeline was used to analyse the thighs of 4 patients in order to suggest a specific muscle therapy before total knee arthroplasty (TKA) and for each of the 7 muscles studied. Preliminary results from the analysis of thigh muscle texture, shape and diffusivity showed that this qMRI protocol can help to suggest a targeted, patient-specific exercise plan to improve muscle recovery after TKA surgery. More healthy and pathological subjects are needed to confirm these encouraging results.

The respiratory motion and the resulting artifacts are considered to be a big problem in abdominal Magnetic Resonance Imaging (MRI), often burdening hospitals with unnecessary cost of time and money. Many deep learning techniques have recently been developed for removing these respiratory motion artifacts. However, fine-details, such as the vessels in liver MRIs, are often over-smoothed by many models, while these details are most important for medical diagnosis. Therefore, more realistic-looking reconstructed images are needed. To achieve this goal, in this paper, we design a Generative Adversarial Networks (GAN)-based model, and use the perceptual loss as part of our training loss to further enhance the images’ perceptual quality. To ensure the proper convergence of the GAN model, we add a pre-training step before the adversarial training. In addition, due to lack of image pairs of the clean image and its corresponding motion distorted image, we develop a motion simulation method via the k-space to mimic the motion generated artifacts. Benefiting from all these methods, our proposed model achieves high perceptual quality with an MSE of 0.002, an SSIM of 0.942, a PSNR of 33.894 and a perceptual distance metric of 0.015. Finally, it generates motion artifact reduced images with clearer and better fine-details, thus providing to radiologists more realistic MRI images for improved diagnosis. The code is available at [Code for Motion Reduction].

Temporal lobe epilepsy (TLE) patients with normal-appearing MRI scans on neuroradiological evaluation are very common (30% of all TLE), but pre-surgical evaluation is challenging for this TLE cohort, and as a result surgery does not often achieve a seizure-free outcome. The purpose of this study is to analyze diffusion magnetic resonance imaging (dMRI) by constructing tractography models that were seeded from the hippocampus to superficial white matter underlying the neocortex and subcortical grey matter structures. We compared mean fractional anisotropy (FA) values in 96 regions of interest between patients with unilateral TLE (10 left lateralized TLE, LTLE and 16 right lateralized TLE, RTLE) and 18 healthy controls. We found that both LTLE and RTLE showed significantly decreased FA values in hippocampal pathways to putamen, pallidum, parahippocampal and entorhinal cortex. In particular, LTLE patients mainly displayed unilateral FA decreases in hippocampal connections ipsilateral to the epileptic focus, whereas RTLE showed bilateral FA decreases. No connectivity changes were found between the hippocampus and the neocortical regions. Our analysis provides novel evidence of alterations in connectivity between the hippocampus and its proximal grey matter structures in patients with MRI-negative TLE.

Preliminary evidence suggests individuals born very-premature have smaller hippocampi on MRI when compared to term-born controls. Moreover, these volumetric reductions have been associated with various cognitive deficits. The hippocampus undergoes an intense period of postnatal volumetric growth during the first year of life. However, this period of development has only been characterized in post-mortem studies. Although volume gain has been previously delineated, changes in hippocampal shape remain undescribed during this unique period. The objective of this study was to characterize and compare morphometric development between very-preterm born infant and healthy controls throughout the first year of life using multivariate tensor-based morphometry (mTBM). We segmented left and right hippocampi from 133 T1-weighted images acquired from 20 very-preterm infants and 67 term-born controls between atbirth or term-equivalent age and 12 months of age. MRI were performed on a 3 Tesla scanner at 3-month intervals (i.e., term-equivalence, 3, 6, 9, 12 months). We used mTBM to compare shape between groups at each time-point. We found that subregions of the hippocampus including the dentate gyrus, CA2, CA3 and subiculum were morphometrically different, especially at term-equivalence age. Morphometric differences were less prominent at 3 and 6 months but reappeared at 9 and 12 months, particularly in the left hippocampus. Although hippocampal shape differences between very-preterm and healthy term-born infants seem to decrease during the first 6 months of life, atypical shape development reappeared at 9-12 months which likely highlights altered periods of morphologic development. Future long-term studies will inform if these developmental differences continue to increase or disappear in subsequent years.

Diffusion-weighted magnetic resonance imaging (dMRI) is sensitive to microstructural changes in the brain that occur with normal aging and Alzheimer’s disease (AD). There is much interest in which dMRI measures are most strongly correlated with (1) AD diagnosis, (2) clinical measures of AD severity, such as the clinical dementia rating (CDR), and (3) biological processes that may be disrupted in AD, such as brain amyloid load measured using PET. Of these processes, some can be targeted using novel drugs. Since 2016, the Alzheimer’s Disease Neuroimaging Initiative (ADNI) has collected dMRI data from three scanner manufacturers across 58 sites using 7 different protocols that vary in angular resolution, scan duration, and distribution of diffusion-weighted gradients. Here, we assessed dMRI data from 730 of those individuals (447 healthy controls, 214 with mild cognitive impairment, 69 with dementia; age: 74.1±7.9 years; 381 female/349 male). To harmonize data from different protocols, we applied ComBat, ComBat-GAM, and CovBat to dMRI metrics from 28 brain regions of interest. We ranked all dMRI metrics in order of the strength of clinically relevant associations, and assessed how this depended on the harmonization methods employed. dMRI metrics were strongly associated with age and AD severity, but also with amyloid positivity. All harmonization methods gave comparable results when assessing associations with age, dementia and amyloid load, while enabling data integration across multiple scanners and protocols.

Alzheimer’s disease (AD) accounts for 60% of dementia cases worldwide; patients with the disease typically suffer from irreversible memory loss and progressive decline in multiple cognitive domains. With brain imaging techniques such as magnetic resonance imaging (MRI), microscopic brain changes are detectable even before abnormal memory loss is detected clinically. Patterns of brain atrophy can be measured using MRI, which gives us an opportunity to facilitate AD detection using image classification techniques. Even so, MRI scanning protocols and scanners differ across studies. The resulting differences in image contrast and signal to noise make it important to train and test classification models on multiple datasets, and to handle shifts in image characteristics across protocols (also known as domain transfer or domain adaptation). Here, we examined whether adversarial domain adaptation can boost the performance of a Convolutional Neural Network (CNN) model designed to classify AD. To test this, we used an Attention-Guided Generative Adversarial Network (GAN) to harmonize images from three publicly available brain MRI datasets - ADNI, AIBL and OASIS - adjusting for scanner-dependent effects. Our AG-GAN optimized a joint objective function that included attention loss, pixel loss, cycle-consistency loss and adversarial loss; the model was trained bidirectionally in an end-to-end fashion. For AD classification, we adapted the popular 2D AlexNet CNN to handle 3D images. Classification based on harmonized MR images significantly outperformed classification based on the three datasets in non-harmonized form, motivating further work on image harmonization using adversarial techniques.

The quantitative characterization of Alzheimer’s Disease (AD) in early stages allows timely detection and prediction of disease progression. These are important to disease intervention and monitoring before clinical diagnosis of dementia. We used cognitive, functional and behavioral data from 612 Alzheimer’s Disease Neuroimaging Initiative (ADNI) individuals. First, we standardized, based on normative data, and dichotomized the selected variables. Grouping abnormal variables, we learned possible disease features from a sample of 225 cognitively impaired patients. Then, we quantify the manifestation for each disease feature and evaluated this quantitative characterization in the automated prediction of future progression from Mild Cognitive Impairment (MCI) to AD dementia. Five groups of abnormal neuropsychological measures were established describing five possible disease features. The resulting quantitative characterization for AD at prodromal stages predicts disease progression within the next 36 months with an accuracy of 0.76.

Corpus callosum (CC) segmentation is an important first step of MRI-based analysis, however most available automated methods and tools perform its segmentation on the midsagittal slice only. Additionally, the few volumetric CC segmentation methods available work on T1-weighted images, what requires an additional step of registering the T1 segmentation mask over diffusion tensor images (DTI) when conducting any DTI-based analysis. This work presents a volumetric segmentation method of the corpus callosum using a modified U-Net on diffusion tensor data, such as Fractional Anisotropy (FA), Mean Difusivity (MD) and Mode of Anisotropy (MO). The model was trained on 70 DTI acquisitions and tested on a dataset composed of 14 acquisitions with manual volumetric segmentation. Results indicate that using multiple DTI maps as input channels is better than using a single one. The best model obtained a mean dice of 83,29% on the test dataset, surpassing the performance of available softwares.

Biomarker-assisted diagnosis and intervention in Alzheimer’s disease (AD) may be the key to prevention breakthroughs. One of the hallmarks of AD is the accumulation of tau plaques in the human brain. However, current methods to detect tau pathology are either invasive (lumbar puncture) or quite costly and not widely available (Tau PET). In our previous work, structural MRI-based hippocampal multivariate morphometry statistics (MMS) showed superior performance as an effective neurodegenerative biomarker for preclinical AD and Patch Analysis-based Surface Correntropy-induced Sparse coding and max-pooling (PASCS-MP) has excellent ability to generate low-dimensional representations with strong statistical power for brain amyloid prediction. In this work, we apply this framework together with ridge regression models to predict Tau deposition in Braak12 and Braak34 brain regions separately. We evaluate our framework on 925 subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Each subject has one pair consisting of a PET image and MRI scan which were collected at about the same times. Experimental results suggest that the representations from our MMS and PASCS-MP have stronger predictive power and their predicted Braak12 and Braak34 are closer to the real values compared to the measures derived from other approaches such as hippocampal surface area and volume, and shape morphometry features based on spherical harmonics (SPHARM).

Magnetic resonance imaging is a versatile diagnostic tool with numerous clinical applications. However, despite advances towards higher resolutions, it cannot resolve images on a cellular level. To nevertheless probe tissue microstructure, multidimensional correlation imaging emerges as a promising method. It takes advantage of the fact that each tissue compartment has a unique signal. Usually, these multi-compartmental characteristics are averaged over a macroscopic voxel. In contrast, correlation imaging aims to probe the true, heterogeneous nature of tissue. Based on image series acquired with varying inversion time T I and echo time T E, multiparametric spectra of T₁ and T₂ relaxation times in every voxel can be reconstructed, revealing sub-voxel tissue classes. However, even with impractically long acquisition times spent on dense sampling of the image (3D) and T IT E-space (2D), the inverse problem of retrieving these components from measured signal curves remains highly ill-conditioned and requires expensive regularized approaches. We formulate multiparametric correlation imaging as a classification problem and propose a flexible physics informed deep learning framework comprising a multilayer perceptron. This way, we efficiently reconstruct voxel-wise T₁-T₂-spectra with increased robustness to noise and undersampling in the T I-T E-space compared to state-of-the-art regression. Our results show feasibility of further acceleration of the acquisition by a factor of 4. After training on synthetic data that is not constraint by pre-defined tissue classes and independent of annotated data, we test our method on in-vivo brain data, revealing sub-voxel compartments in white and gray matter. This allows us to quantify tissue microstructure and will potentially lead to novel biomarkers.

The streamline methods in Diffusion tensor imaging (DTI) use principal eigenvectors for tracking white matter fibers. In comparison, the geodesics in a multivalued ray-tracing method are closer to the actual underlying white matter fibers. This method provides multiple solutions in the form of geodesics in a Riemannian space. These geodesics are robust in tracking high curvature structures in the presence of noise. In the DTI modality, the 2nd order tensor fails to model the heterogeneous regions, e.g., crossing/merging fibers. Therefore, the ray-tracing method is limited to guide the geodesics in these regions. This work, as a first step, utilizes 4th order tensor approximation for the white matter regions. Subsequently, a non-linear optimization is performed to decompose 4th order tensors into multiple 2nd order tensors keeping their symmetric positive definite property. We are using the initial fiber directions from the diagonal components of the 4th order tensor and use the 4th order tensor decomposition for fiber tracking. The experimental results on synthetic images show that geodesics can traverse in heterogeneous and high curvature structures.

Parkinson disease (PD) is a common neurodegenerative pathology, whose accurate diagnosis is still a challenge. PET imaging obtained with [18F]-fluorodeoxyglucose provides a metabolic pattern, highlighting the brain substructures related to PD, thus constituting a valuable diagnosis tool. Besides, it has been reported that incorporating MRI into the analysis enhances the performance of methods aiming to discriminate between healthy subjects and PD patients. In this research, a methodology is proposed that allows: to integrate structural and metabolic imaging information at specific substructures of interest; to spatially align both modalities; to normalize functional images and to extract the adequate biomarkers. Among structural parameters, compacity and tortuosity are proposed, while metabolic biomarkers are extracted from histogram analyses. The random forest algorithm is used for classification and feature selection tasks. The studied populations consisted of nine patients with PD diagnosis and 12 healthy controls. Structural biomarkers showed a small contribution to discriminate between groups, while metabolic biomarkers resulted in 85% (training) to 100% (final test) accuracies. The proposed methodology is promising to diagnose PD and can be extended to other movement disorders.

Variational segmentation algorithms require a prior imposed in the form of a regularisation term to enforce smoothness of the solution. Recently, it was shown in the Deep Image Prior work that the explicit regularisation in a model can be removed and replaced by the implicit regularisation captured by the architecture of a neural network. The Deep Image Prior approach is competitive, but is only tailored to one specific image and does not allow us to predict future images. We propose to incorporate the ideas from Deep Image Prior into a more traditional learning algorithm to allow us to use the implicit regularisation offered by the Deep Image Prior, but still be able to predict future images.

Diabetic retinopathy (DR) is the result of a complication of diabetes affecting the retina. It can cause blindness, if left undiagnosed and untreated. An ophthalmologist performs the diagnosis by screening each patient and analyzing the retinal lesions via ocular imaging. In practice, such analysis is time-consuming and cumbersome to perform. This paper presents a model for automatic DR classification on eye fundus images. The approach identifies the main ocular lesions related to DR and subsequently diagnoses the illness. The proposed method follows the same workflow as the clinicians, providing information that can be interpreted clinically to support the prediction. A subset of the kaggle EyePACS and the Messidor-2 datasets, labeled with ocular lesions, is made publicly available. The kaggle EyePACS subset is used as training set and the Messidor-2 as a test set for lesions and DR classification models. For DR diagnosis, our model has an area-under-the-curve, sensitivity, and specificity of 0.948, 0.886, and 0.875, respectively, which competes with state-of-the-art approaches.

Deep learning neural networks are a common tool in medical imaging and frequently used to solve a variety of complex problems. Magnetic resonance (MR) images are frequently employed to develop these networks because of their high spatial resolution and user selectable image contrast between tissues. More advanced deep learning models are being developed, which when combined with improvements in MR image acquisition techniques, will allow for image analysis techniques that are more efficient yet able to solve increasingly challenging problems. A current significant disadvantage of deep learning networks is that they are extremely sensitive to the distribution of the data used for training, therefore, network implementation can be challenging in clinical applications with heterogeneous images. The main problem is that, in a clinical environment, data distributions of target datasets can vary from subject-to-subject due to differences in scanner vendor, magnetic field strength, and the setting of specific MR acquisition parameters. These variations create inherent scan variability that diversifies the data distributions of different datasets. This effect can result in the model becoming inaccurate and producing undesirable outcomes. Thus, to improve model generalizability, we explored a supervised domain-adaptation approach. To test this method, we created a convolutional neural network model that performed a classification task and was composed of three components: (1) a feature extractor, (2) a pathology classifier, and (3) a domain classifier. In a single, unified training process, the pathology classifier was trained by minimizing the pathology loss function and the domain classifier was trained by maximizing the domain loss function. This procedure allows the model to penalize learning of features specific to the domain, and thus attempts to produce a domaininvariant feature vector. The performance of this domain-adapted model was compared to the same model but without domain classification (i.e., a baseline traditional model consisting of a feature extractor and a pathology classifier). We found that the domain-adapted model achieved a higher accuracy rate in predicting images from both source and target datasets.

The monitoring of people in health centers and geriatric homes is performed by rehabilitation professionals who manually evaluate the surveillance cameras for identifying one person’s position, and his physical condition. However, this task is tedious and demands the full attention of the rehabilitation staff because patients with neurological conditions need special care or in some cases the 24/7 monitoring. On the other hand, the use of artificial intelligence in the detection of objects and people through images or videos has presented a great performance. This article presents a methodology based on deep learning for the detection and monitoring of people in closed and open environments using video. The proposed method is non-invasive, low-cost, and evaluates the physical activity and inactivity of people in real-time. Preliminary results in public databases present outstanding results in the monitoring and estimation of caloric expenditure in people in indoor and outdoor spaces.

Parkinson’s disease (PD) and Alzheimer’s disease (AD) are progressive neurodegenerative disorders that affect millions of people worldwide. In this work, we propose a deep learning approach to classify these diseases based on 3D T1- weighted brain MRI. We analyzed several datasets including the Parkinson's Progression Markers Initiative (PPMI), an independent dataset from the University of Pennsylvania School of Medicine (UPenn), the Alzheimer’s Disease Neuroimaging Initiative (ADNI), and the Open Access Series of Imaging Studies (OASIS) dataset. PPMI and ADNI were partitioned to train (70%), validate (20%), and test (10%) a 3D convolutional neural network (CNN) for PD and AD classification. The UPenn and OASIS datasets were used as independent test sets to evaluate the model performance during inference. We also implemented a random forest classifier as a baseline model by extracting key radiomics features from the same T1-weighted MRI scans. The proposed 3D CNN model was trained from scratch for the classification tasks. For AD classification, the 3D CNN model achieved an ROC-AUC of 0.878 on the ADNI test set and an average ROC-AUC of 0.789 on the OASIS dataset. For PD classification, the proposed 3D CNN model achieved an ROC-AUC of 0.667 on the PPMI test set and an average ROC-AUC of 0.743 on the UPenn dataset. We also found that model performance was largely maintained when using only 25% of the training dataset. The 3D CNN outperformed the random forest classifier for both the PD and AD tasks. The 3D CNN also generalized better on unseen MRI data from different imaging centers. Our results show that the proposed 3D CNN model was less prone to overfitting for AD than for PD classification. This approach shows promise for screening of PD and AD patients using only T1-weighted brain MRI, which is relatively widely available. This model with additional validation could also be used to help differentiate between challenging cases of AD and PD when they present with similarly subtle motor and non-motor symptoms.

Deep learning algorithms rely on large amounts of annotations for learning and testing. In digital pathology, a ground truth is rarely available, and many tasks show large inter-expert disagreement. Using the Gleason2019 dataset, we analyse how the choices we make in getting the ground truth from multiple experts may affect the results and the conclusions we could make from challenges and benchmarks. We show that using undocumented consensus methods, as is often done, reduces our ability to properly analyse challenge results. We also show that taking into account each expert’s annotations enriches discussions on results and is more in line with the clinical reality and complexity of the application.

This work presents a novel quantification of the cancer extension using a latent space embedded metrics of a variational autoencoder which captures the invariant patterns of the disease and projects them into a smaller latent space where data relations are linear, making it possible to apply simple metrics to quantify complicated relations. Selected patches of non-small cell lung cancer are projected to such latent space and a logistic regression model assigns an Euclidean distance between the patches projected in space. A simple grouping strategy quantitatively stratifies the characteristic patterns of the most representative patches for both adenocarcinoma and squamous cell lung cancer classes but it also estimates the composition of a mixture of patterns. This approach is fully interpretable, integrable with a pathology work flow and an objective characterization of diseases with complex patterns.

Glomeruli are histological structures located at the beginning of the nephrons in the kidney, having primary importance in diagnosing many renal diseases. Classifying glomerular lesions is time-consuming and requires experienced pathologists. Hence automatic classification methods can support pathologists in the diagnosis and decision-making scenarios. Recently most of state-of-the-art medical imaging classification methods have been based on deep-learning approaches, which are prone to return overconfident scores, even for out-of-distribution (OOD) inputs. Determining whether inputs are OOD samples is of underlying importance so as to ensure the safety and robustness of critical machine learning applications. Bare this in mind, we propose a unified framework comprised of unbounded open-set recognition and multi-lesion glomerular classification (membranous nephropathy, glomerular hypercellularity, and glomerular sclerosis). Our proposed framework classifies the input into in- or OOD data: If the sample is an OOD image, the input is disregarded, indicating that the model “doesn’t know” the class; otherwise, if the sample is classified as in-distribution, an uncertainty method based on Monte-Carlo dropout is used for multi-lesion classification. We explored an energy-based approach that allows open-set recognition without fine-tuning the in-distribution weights to specific OOD data. Ultimately, our results suggest that uncertainty estimation methods (Monte-Carlo dropout, test-time data augmentation, and ensemble) combined with energy scores slightly improved our open-set recognition for in-out classification. Our results also showed that this improvement was achieved without decreasing the 4-lesion classification performance, with an F1-score of 0.923. Toward an unbounded open-set glomerular multi-lesion recognition, the proposed method also kept a competitive performance.

The sentinel lymph node is a predictor of breast cancer aggressiveness.¹ Patients with micro-metastasis (MM) are usually considered negative, yet their the hazard ratio has been reported to be 2.4 and 1.203.81 with a 95% confidence interval.^2–4 This work proposes an automatic detection of micro-metastasis by quantifying local cellular changes. The proposed strategy characterizes nuclei morphometry, color and texture to establish differences between MM and normal tissue. The color model is obtained from the plane [(r − b), g] while texture corresponds to the Haralick’s features from five different orders of the co-occurrence matrix.⁵ This description is complemented by the cellular area obtained from a conventional watershed segmentation. An AdaBoost model, trained with 300 patches of 350 × 350 pixels (56000 μm² ) randomly selected from 18 cases, was tested in a set of five different cases with approximately ten patches containing micro-metastasis. This approach obtained a best classification accuracy of 0.86, sensitivity of 0.89, specificity of 0, 83, and F-score of 0.86, while the baseline, a ResNet 50 model, obtained 0.74 of accuracy, 0.86 of sensitivity, 0, 63 of specificity, and F-score of 0.77 for exactly the same task.

This work reports to the development of a metric system which is able to detect the glottic squamous cell carcinoma tumor and establish its position. It was observed that the images of the glottis, seen through an endoscope, are shaped according to their geometric characteristics and by shades of color and brightness. Thus, by superimposing the metric of a diffractive grid with light spots on images containing the glottic region, the position and attributes of the tumor can be quantified. Thus, by implement computer-aided diagnostic techniques, it is possible to extract features of the pathology, as well as minimize operational costs, compared to empirical experiments. To this end, this study was carried out by analyzing narrow-band endoscope sample images from 10 patients with squamous cell carcinoma-like tumor confirmed on medical diagnosis. The light spot array was generated by a diffractive optical element, designed by an iterative Fourier-transform method. Images segmentation was based on methods of image filtering as well as on active contour region. By employing a 3 x 3 spot array it was possible to obtain ' 97 % tumor detection. The proposed metric system can be considered a powerful tool in order to assist in the prognosis of laryngeal pathologies by detecting attributes and position of laryngeal squamous cell carcinoma tumor.

This work presents the development of a microservices integration platform for e-health, called HealthGate. The architecture proposed for the platform based on microservices aims to help contextualized implementations in e-health, aiming to reach critical points related to the domain, such as the heterogeneity of services and specifications. We propose the use of the FHIR standard for the integration of microservices in the platform, contributing to the interoperability between systems. To validate the proposed architecture and the defined requirements, two case studies were carried out, deploying a microservice that detects episodes of atrial fibrillation in ECG segments and a microservice that calculates the Framingham Score.

Surface electromyography (sEMG) is the most common technology used in gesture recognition for hand prosthesis control. Machine learning models based on Convolutional Neural Networks (CNN) that classify the gestures from the sEMG signals can reach high accuracy. However, common changes in the condition of use of the prosthesis, such as electrode shifts, can drastically impact these metrics, causing the need to retrain the model. Considering the application of a model based on such characteristics which was originally trained using data from subjects (source domain) other than the data from the user of the prosthesis (target domain), a domain adaptation must be employed. To lower the time spent during the retraining, only a small amount of data from the target domain must be considered. A relevant factor to be taken into account is that, for the prosthesis to be economically viable, the computational effort required to solve this problem must be supported by a common and inexpensive hardware device. In the current work, the CapgMyo sEMG dataset was used to fine-tune a 2D CNN-based model in an edge device. Inter-subject gesture recognition is performed, where the source domain is composed of the data from 17 of the 18 subjects of the dataset, while the target domain consists of data from the remaining subject. During the fine-tuning, only the classifier layer was retrained, while all the other layers were frozen. The fine-tuning was performed in a common Raspberry Pi 3. Results show that the computational power of the device is enough for a good accuracy using a small amount of data collected by the subject.

Federated learning (FL) enables distributed computation of machine learning models over various disparate, remote data sources, without requiring to transfer any individual data to a centralized location. This results in an improved generalizability of models and efficient scaling of computation as more sources and larger datasets are added to the federation. Nevertheless, recent membership attacks show that private or sensitive personal data can sometimes be leaked or inferred when model parameters or summary statistics are shared with a central site, requiring improved security solutions. In this work, we propose a framework for secure FL using fullyhomomorphic encryption (FHE). Specifically, we use the CKKS construction, an approximate, floating point compatible scheme that benefits from ciphertext packing and rescaling. In our evaluation on large-scale brain MRI datasets, we use our proposed secure FL framework to train a deep learning model to predict a person’s age from distributed MRI scans, a common benchmarking task, and demonstrate that there is no degradation in the learning performance between the encrypted and non-encrypted federated models.

In this work, we propose a novel medical image encryption algorithm. It is based on the Jigsaw transform, Langton’s ant, cyclic permutation, and a novelty way to add deterministic noise. The robustness of our algorithm has been proven through several testing, such as statistical analysis (histograms and correlation distributions), visual testing, entropy testing and key space assessment, showing in each one a high-security level.

Medical imaging analysis is an effective technique and process for visualizing the human body’s interior to diagnose, monitor, and treat medical conditions. Artificial Intelligence (AI) brings new opportunities for improvement, with multiple applications in all levels of the radiology workflow. This paper presents a solution that leverages state-of-the-art models and architectures to assemble a modular pipeline for detection, segmentation, measurement, and scoring, that builds up to an optimized clinical report for medical imaging analysis and diagnosis. The proposed approach is designed to be flexible and tailor-made to an end facility’s needs and data, helping the radiologist’s effectiveness.

Parkinson's disease (PD) is a chronic neurodegenerative disorder that commonly affects older people. Pain is one of the most common non-motor symptoms in this population, and it is very important to be constantly monitored and evaluated. However, there is a scarcity of questionnaires that effectively address the impact of pain in this population, and many instruments superficially assess this symptom. Then, the aim of this study was to remotely apply the Portuguese version of an already validated pain assessment questionnaire (MPI) in individuals with PD, and, with factor analysis, adjust/validate the questionnaire for this population and assess the level of pain perception felt by the respondents. The questionnaire was answered by 309 individuals with PD, 251 of whom reported the presence of pain. Factor analysis and descriptive statistics were used to analyze the data. The factor analysis (FA) suggested the exclusion of 2 items, and the grouping of the questions into 2 factors: affectability and severity of pain, and self-control and energy. The value of Cronbach's alpha was higher than that considered adequate, indicating high internal consistency of the instrument. A model to calculate the final scores of the respondents was established and, according to this model, it was observed that the individuals that made up the sample presented a high perception of pain, since many scored high scores. In summary, the evaluated questionnaire in is a complete pain assessment tool, and easily applied remotely to individuals with PD.

Analysis of retinal fundus images have been proven to provide relevant information about the diagnoses of several pathologies. Among them, glaucoma stands out as an important pathology due to the need for early treatment. Moreover, the relationship between optic disc and optic cup regions provided by retinal fundus image analysis can aid in diagnosis. Automatically generating such a relation is, therefore, an important feature for ensuring quicker and more precise conclusions. This paper evaluates the use of Conditional GAN (Generative Adversarial Networks) for an optic disc and optic cup segmentation task. Conditional GANs are hybrid machine learning models that are able to generate data based on conditioned training. The results demonstrate that the addressed method generates valid segmentation images for optic disc and optic cup location, with approximately 95% and 85% accuracy, respectively

In this work, the use of different Machine Learning classifiers with several features sets is presented to categorize Dual X-Ray Absorptiometry images according to the texture of their bone structure in two classes: control and osteoporosis. Two different bone areas were used: Ward’s Triangle and Femoral Neck. The diagnosis of this disease employing methods that analyze the bone structure through images of different medical modalities is an area of great interest. Support Vector Machine and Random Forest showed better AUC values than the rest of the methods. The features that gave the best results were those related to first-order statistics, which is consistent with what was found in previous work. For the combination of features and classification methods, the best results were 0.955 for AUC and 0.916 for F-Score using the Femoral Neck area. The obtained results constitute a reasonable basis for continuing the contributions to the subject by applying other learning methods such as deep learning.

Infertility is becoming an issue for an increasing number of couples. The most common solution, in vitro fertilization, requires embryologists to carefully examine light microscopy images of human oocytes to determine their developmental potential. We propose an automatic system to improve the speed, repeatability, and accuracy of this process. We first localize individual oocytes and identify their principal components using CNN (U-Net) segmentation. Next, we calculate several descriptors based on geometry and texture. The final step is an SVM classifier. Both the segmentation and classification training is based on expert annotations. The presented approach leads to a classification accuracy of 70%.

Early-stage detection of Bayès Syndrome, for preventive purposes, is of great interest due to its associations with multiple medical conditions. A strategy for increasing the original dataset employed on this syndrome's discovery is presented in this work. Furthermore, two clustering techniques were applied: K-Means++ (two different implementations) and FAUM. In addition, FAUM was used with a fixed number of clusters. By applying data augmentation techniques, 2113 signals were obtained from the 49 initial quantity. For each ECG signal sample, derivative and integrative variations on II, III and AVF leads, were calculated. On these datasets both K-Means++ (with Matlab and FAUM implementations) and FAUM (with a fixed number of clusters) were applied to these datasets. Regarding the results, K-Means++ with the Matlab implementation reached a value of 0.82 for the F1-Score metric. Also, integrative variation performance was better than derivative in detecting the P-Wave morphology.

This article presents the application of the full multiresolution active shape model for the segmentation of the left ventricle in ultrasound images as well as a comparison between the classical active shape model and the full multiresolution active shape model. Our objective is to evaluate the performance of the full multiresolution framework in a complex image segmentation task such as ultrasound of the left ventricle. The accuracy of this method is evaluated through the DICE coefficient between the expert annotation and the final segmentation of the full multiresolution active shape model (FMR-ASM). The validation and training data were obtained from the CAMUS database, with 100 training and 60 validation images on which we obtained a mean DICE coefficient of 0.76 with the FMR-ASM.

This work proposes a novel analysis of the left ventricular chamber dynamics from ultrasound 2D videos, in four steps: first the left ventricular chamber is segmented and second a multi-orientation and multi-scale filtering is performed. A third step is the chamber partition in similar number of super-pixels or homogeneous regions. The final step extracts features from the velocity-acceleration phase plane constructed by tracking these regions along the cardiac cycle and estimating their velocity and acceleration. Finally, each case is characterized by dividing the phase plane into three disjoint areas along the radial direction and estimating the density of points per region. This approach was evaluated in actual videos of four subjects, two control and two patients. Results show density means for unhealthy and control as follows: 0.63 and 0.45 for the low motion region, 0.26 and 0.4 for the mid motion region, and 0.1 and 0.1 for the high motion region.

We have recently shown the feasibility of unsupervised classification to identify clusters of persistent atrial fibrillation (persAF) atrial electrograms (AEGs) using attributes calculated by electroanatomic mapping systems. In the present work, we sought to extend this analysis to a set of reduced dimension principal component scores (PCSs) estimated from AEG-derived markers. 956 bipolar AEGs were collected from 11 persAF patients. Principal component analysis (PCA) was implemented on 28 markers obtained from different perspectives of the AEGs – such as information theory, spectral analysis, and nonlinear dynamics. The three PCSs with the highest variances were selected and used to perform an unsupervised classification with k-means. The electrophysiological characteristics of the identified groups were further investigated using six AEG markers: sample entropy (SampEn), wave similarity (WS), cycle length (CL), CARTO’s interval confidence level (ICL), peak-to-peak (PP) amplitude and determinism (DET). The three most relevant PCSs explained 81.3% of the variance (56% PCS1, 14.6% PCS2 and 10.7% PCS3). Six AEG groups were identified from the PCSs (F=485, P<0.0001). Group 1 (15%) had highest SampEn, group 4 (16%) highest ICL, and group 5 (13%) highest WS. Markers suggest mild fractionation on group 2 (30%). Group 3 (14%) are AEGs with low ICL and lowest PP. Finally, group 6 (12%) clustered AEGs with highest PP and DET, and with lowest SampEn. The six clusters of AEGs found in this work are in line with recent findings, highlighting the complex pathophysiology of persAF. These could provide a more complete characterization of persAF substrate during ablation target identification in future clinical studies.

Myocardial stimulation using electric field (E) is used not only clinically (e.g. defibrillation protocols) but also in in vitro experiments. Normally, to stimulate heart cells in vitro it is necessary that the cell is positioned in the same direction as the stimulus, in order to lower the intensity, and that is not always possible. In protocols that require multidirectional stimuli, the lack of space to position the electrodes is a major limitation, knowing that two electrodes are required for each stimulus direction, which makes it not always viable. This paper presents a multidirectional electrical stimulator that is capable of determining the stimulation threshold of a rat heart cell, using only two isolated channels, applying 3 stimuli in 3 directions by the principle of superposition. The stimulator is implemented through a microcontroller, responsible for processing the user interface data and controlling the intensity of the stimuli using PWM. The stimulator reaches voltages up to ±30 V. When comparing the intended voltage with the achieved voltage in the stimulator, we have a maximum error of 3%. The stimulator fulfills its objective which allows future experiments with isolated cardiac cells.

The increase in deaths from cardiovascular diseases in extreme temperature events, particularly during cold waves, is a phenomenon already reported in the literature. One of its mechanisms is the increased likelihood of atherosclerotic plaque formation due to higher concentrations of low-density lipoprotein (LDL-C) in the blood at periods of lower temperatures. This study adopts a data science approach to check evidences of this mechanism in the population of Campinas, a city in the southeast region of Brazil, with over 1 million inhabitants. We integrated climate and health datasets and processed over 1,677,424 LDL-C exam results in combination with minimum and maximum daily air temperature data in the city in eleven years (2008-2018). The data were stratified into sex and age groups, and we analyzed the difference in distributions of LDL-C levels for those exposed to cold waves versus control days. Cold waves were defined as at least three consecutive days with minimum and maximum temperatures below their 10th percentiles, considering a 30-year climate normal (1961- 1990). In particular, we analyzed the effect of cold waves on LDL-C levels above reference value. Our analyses identified nine cold waves in the period and statistically relevant effects on exam results. The amount LDL-C exams with levels above reference value was 3.32% greater for women between 20 and 65 years old (lag 2), 9.27%, 7.39% and 4.06% for women over 65 (lags 0, 2 and 7, respectively), and 11.45% for men over 65 (lag 4).

Reduction of the speckle noise and other artifacts, inherent to the ultrasound imaging technique, and the increase of the resolution of reconstructed images are important features to improve image quality, assisting interpretation and increasing assertiveness of medical diagnosis. In this paper, we design a computational study to compare the results of conventional and spatial compounding imaging methods, using frames in B-Mode and/or Harmonic Imaging. A 3D phantom was simulated, using Matlab software in conjunction with the k-Wave toolbox, consisting of 4 targets with higher scattering density. The assessment of results is made by applying two graphical metrics: composite signal-to-noise ratio (SNRc) and composite contrast-to-noise ratio (CNRc). The results are satisfactory, and for the regions of interest, there is an improvement up to 143% of SNRc and 140% of CNRc metrics.

In this paper we present a model-based Field Programmable Gate Array (FPGA) design flow for ultrasound imaging using the delay-and-sum (DAS) beamforming technique. The beamformer model is based on DSP Builder toolbox in the Matlab/Simulink environment, used for fast prototype design and automatic hardware description language code generation for Intel FPGAs. The design was synthesized for an Intel Stratix IV FPGA and implemented on a Terasic DE4- 230 board, including the following digital signal processing steps: digital filtering, focusing delay adjustment, apodization, coherent summation, envelope detection and logarithmic compression. To evaluate our model, we used raw radiofrequency data from a phantom acquired using a 128-element convex transducer (2-5 MHz) with an active aperture of 8 elements. A Nios II processor was integrated into the FPGA to send and receive data from the model and transmit them to the host computer for off-line scan conversion and B-mode image generation. The results show that the proposed strategy is efficient for rapid prototyping of FPGA devices in ultrasound imaging applications. The overall FPGA resources utilization was less than 15%, indicating that other digital signal processing algorithms, such as interpolation, decimation, and scan conversion, can be implemented using this design methodology in both medical and non-destructive imaging applications.

Open ultrasound platforms foster the development of novel image processing techniques due to its flexibility in terms of both software and hardware updates. There is a growing need for rapid prototyping and modelling platforms can be used to develop high-level applications, such as MATLAB/Simulink. Although there are several model-based methods for ultrasound imaging described in the literature, most of them focus on specific digital signal processing stages, and there is little assessment of the simulation environment of Model Composer and System Generator for Xilinx Field-programmable Gate Arrays (FPGA). The purpose of this article is to implement the steps of an ultrasound reception beamforming using the delay and sum technique for B-mode imaging, and then comparing the results between a reference MATLAB script and the results from Xilinx based-model implemented in Simulink environment. The simulation uses 12-bit ultrasound data sampled at 40 MHz from a phantom acquired with a 128-element transducer with an aperture of 8 elements. The design was synthesized for a Xilinx Zynq 7000 FPGA. The qualitative results of the generated B-mode imagens are in excellent agreement, with a CNR error of 0.01, and a processed envelope with NRMSE of 0.085%. In conclusion, the use of tools such as Model Composer and system generator proved to be adequate for rapid prototyping and yields very accurate results for FPGA simulations when compared to MATLAB calculations.

Focused ultrasound has been widely investigated in medical applications for tissue ablation and, more recently, drug delivery and neuromodulation. Those techniques require high focusing precision in the brain, where treatment is commonly planned with the help of numerical simulations. Treatment planning is patient-specific based on computed tomography (CT) images and/or magnetic resonance images. In this study, a threshold segmentation process based on the density and sound speed values obtained from the CT data conversion was tested in thermoacoustic simulations using the k-Wave Matlab Toolbox. Acoustic and thermal mappings generated by focused ultrasound sources provided information about treatment targeting and safety, including temperature mappings covering the skull and brain tissues.

We describe an automatic pipeline for processing ultrasound images of the carotid artery, consisting of image type classification, carotid artery localization, segmentation, feature descriptor extraction, and plaque stability classification. The aim is to distinguish between stable (safe) and progressive (dangerous) atherosclerotic plaques from a single standard ultrasound transversal or longitudinal B-mode examination. The processing pipeline uses modern deep CNN techniques, while the descriptors are based on geometry and wavelets to characterize texture. When testing on a large dataset of 28718 images from 413 patients, we found that our automatically calculated descriptors are statistically significantly different between the two classes with a very high significance level, p < 10⁻³ . We have also created a random forest-based classifier to distinguish between progressive and stable plaques, although its accuracy remains low (61 ∼ 62%).