Cross-Modality Feature Learning for Three-Dimensional Brain Image Synthesis

Multi-modality medical imaging is increasingly used for comprehensive assessment of complex diseases in either diagnostic examinations or as part of medical research trials. Different imaging modalities provide complementary information about living tissues. However, multi-modal examinations are not always possible due to adversary factors such as patient discomfort, increased cost, prolonged scanning time and scanner unavailability. In addition, in large imaging studies, incomplete records are not uncommon owing to image artifacts, data corruption or data loss, which compromise the potential of multi-modal acquisitions. Moreover, independently of how well an imaging system is, the performance of the imaging equipment usually comes to a certain limit through different physical devices. Additional interferences arise (particularly for medical imaging systems), for example, limited acquisition times, sophisticated and costly equipment and patients with severe medical conditions, which also cause image degradation. The acquisitions can be considered as the degraded version of the original high-quality images. In this dissertation, we explore the problems of image super-resolution and cross-modality synthesis for one Magnetic Resonance Imaging (MRI) modality from an image of another MRI modality of the same subject using an image synthesis framework for reconstructing the missing/complex modality data. We develop models and techniques that allow us to connect the domain of source modality data and the domain of target modality data, enabling transformation between elements of the two domains. In particular, we first introduce the models that project both source modality data and target modality data into a common multi-modality feature space in a supervised setting. This common space then allows us to connect cross-modality features that depict a relationship between each other, and we can impose the learned association function that synthesizes any target modality image. Moreover, we develop a weakly-supervised method that takes a few registered multi-modality image pairs as training data and generates the desired modality data without being constrained a large number of multi-modality images collection of well-processed (\textit{e.g.}, skull-stripped and strictly registered) brain data. Finally, we propose an approach that provides a generic way of learning a dual mapping between source and target domains while considering both visually high-fidelity synthesis and task-practicability. We demonstrate that this model can be used to take any arbitrary modality and efficiently synthesize the desirable modality data in an unsupervised manner. We show that these proposed models advance the state-of-the-art on image super-resolution and cross-modality synthesis tasks that need jointly processing of multi-modality images and that we can design the algorithms in ways to generate the practically beneficial data to medical image analysis.