Automated Discovery and Visualization of Discriminating Structural Markers from MRI Using Transport-Based Morphometry
Imaging studies are a vital part of accurate medical care for every part of the body. Recent advances in magnetic resonance imaging (MRI) technology hold promise to uncover structural changes underlying diseases commonly considered medical mysteries. Unfortunately, the changes are often subtle, spatially distributed and complex, escaping detection by visual inspection. The key challenge is to disambiguate meaningful information from heterogeneous normal variations. The assistance of computer-aided techniques is sought to answer the following questions: are there significant morphologic differences that differentiate these patients? If so, what is the nature of these differences? Traditional computer-aided techniques that extract a set of pre-specified descriptors (i.e. volume, thickness, etc.) or compare images pixelwise can test directed hypotheses about regional differences, but do not consider important spatially diffuse effects. Furthermore, current approaches for assessing statistical differences do not allow recovery or visualization of the images underlying the associations, hindering physical interpretation. The goal of this research is to develop a new framework for automated discovery and visualization of discriminating structural markers from MRI data. We utilize the mathematics of optimal mass transport (OT) to quantify tissue spatial distribution and transform data to enhance separability of images. Our approach for Transport- Based Morphometry (TBM) defines a fully invertible transformation, allowing statistically significant discriminant differences identified in the transform domain to be visualized as morphology changes through inverse transformation. This thesis extends the mathematics of TBM to enable its first application to 3D radiology data, and applies the new TBM framework for classification and regression tasks in a variety of open clinical and research problems. Our results based on knee cartilage images indicate that TBM enables detection of osteoarthritis with 86.2% accuracy three years in advance of symptoms. TBM also enables identification of 16p11.2 duplication, deletion, and control carriers with 96% accuracy based on structural appearance of the brain alone. Patterns of injury underlying post-concussive cognitive deficits are identified with the TBM technique. The new technique also finds that aerobic fitness significantly affects brain tissue distribution in older adults. These are significant improvements over prior detection methods. Moreover, unlike previous approaches, our method allows structural markers to be visualized.