Information Fusion of Magnetic Resonance Images and Mammographic Scans for Improved Diagnostic Management of Breast Cancer

Abstract

Medical imaging is critical to non-invasive diagnosis and treatment of a wide spectrum of medical conditions. However, different modalities of medical imaging employ/apply di erent contrast mechanisms and, consequently, provide different depictions of bodily anatomy. As a result, there is a frequent problem where the same pathology can be detected by one type of medical imaging while being missed by others. This problem brings forward the importance of the development of image processing tools for integrating the information provided by different imaging modalities via the process of information fusion. One particularly important example of clinical application of such tools is in the diagnostic management of breast cancer, which is a prevailing cause of cancer-related mortality in women. Currently, the diagnosis of breast cancer relies mainly on X-ray mammography and Magnetic Resonance Imaging (MRI), which are both important throughout different stages of detection, localization, and treatment of the disease. The sensitivity of mammography, however, is known to be limited in the case of relatively dense breasts, while contrast enhanced MRI tends to yield frequent 'false alarms' due to its high sensitivity. Given this situation, it is critical to find reliable ways of fusing the mammography and MRI scans in order to improve the sensitivity of the former while boosting the specificity of the latter. Unfortunately, fusing the above types of medical images is known to be a difficult computational problem. Indeed, while MRI scans are usually volumetric (i.e., 3-D), digital mammograms are always planar (2-D). Moreover, mammograms are invariably acquired under the force of compression paddles, thus making the breast anatomy undergo sizeable deformations. In the case of MRI, on the other hand, the breast is rarely constrained and imaged in a pendulous state. Finally, X-ray mammography and MRI exploit two completely di erent physical mechanisms, which produce distinct diagnostic contrasts which are related in a non-trivial way. Under such conditions, the success of information fusion depends on one's ability to establish spatial correspondences between mammograms and their related MRI volumes in a cross-modal cross-dimensional (CMCD) setting in the presence of spatial deformations (+SD). Solving the problem of information fusion in the CMCD+SD setting is a very challenging analytical/computational problem, still in need of efficient solutions.

In the literature, there is a lack of a generic and consistent solution to the problem of fusing mammograms and breast MRIs and using their complementary information. Most of the existing MRI to mammogram registration techniques are based on a biomechanical approach which builds a speci c model for each patient to simulate the effect of mammographic compression. The biomechanical model is not optimal as it ignores the common characteristics of breast deformation across different cases. Breast deformation is essentially the planarization of a 3-D volume between two paddles, which is common in all patients. Regardless of the size, shape, or internal con guration of the breast tissue, one can predict the major part of the deformation only by considering the geometry of the breast tissue. In contrast with complex standard methods relying on patient-speci c biomechanical modeling, we developed a new and relatively simple approach to estimate the deformation and nd the correspondences. We consider the total deformation to consist of two components: a large-magnitude global deformation due to mammographic compression and a residual deformation of relatively smaller amplitude. We propose a much simpler way of predicting the global deformation which compares favorably to FEM in terms of its accuracy. The residual deformation, on the other hand, is recovered in a variational framework using an elastic transformation model. The proposed algorithm provides us with a computational pipeline that takes breast MRIs and mammograms as inputs and returns the spatial transformation which establishes the correspondences between them. This spatial transformation can be applied in different applications, e.g., producing 'MRI-enhanced' mammograms (which is capable of improving the quality of surgical care) and correlating between different types of mammograms. We investigate the performance of our proposed pipeline on the application of enhancing mammograms by means of MRIs and we have shown improvements over the state of the art.