|
Sophisticated non-linear image processing techniques based on machine learning, such as denoising and superresolution, are being aggressively incorporated into the clinical practice of radiology. There is a need for equally sophisticated methods to evaluate the performance of these new algorithms, particularly in cases where adaptive processing might behave differently in bench testing phantoms than in patients. This issue has been the subject of intense research in recent years in photography due to the emergence of cell-phone cameras with built-in processing. In this work, we introduce methods used to evaluate image resolution in photography, as defined by the international standards ISO 12233 and IEEE 1858, and describe how some of these methods can be adapted to radiography. We propose the use of a completely in silico pipeline for the evaluation of post-processing software devices, to enable the use of noise-free, high-resolution ground-truth data in the evaluation. A digital twin of a mammography device, based on the open-source VICTRE MC-GPU x-ray simulator which accurately replicates the noise and loss of resolution in clinical mammograms, was used as a test bed to study different evaluation methods. Three upgrades to the simulation software were implemented: a charge-sharing mechanism to model pixel correlations, fast primary-only x-ray tracking, and insertion of a 2D image of arbitrary resolution inside a voxelized 3D geometry. These new capabilities were used to insert complex image quality evaluation charts, such as the sine-wave Siemens star, slanted-edge, and “dead leaves” pattern, inside simulated mammograms. An innovative adaptation of the “dead leaves” pattern to mammography is introduced. The proposed method uses a noise-free, high-resolution projection of a VICTRE breast phantom as a reference to evaluate the spatial frequency response of the post-processed simulated mammograms. The advantage of this new test pattern is that the effective resolution is calculated directly on the mammogram itself, and deep-learning algorithms are likely to process the pattern in a similar way as a patient image. The introduced computer modeling and simulation methods might provide valuable information to complement bench testing results in the evaluation of image processing algorithms used in radiology.
|
