AbdoMReg: A Deep Learning Framework for Abdominal MR Deformable Image Registration

Nowadays, novel imaging techniques benefit various clinical applications ranging from patient diagnosis and follow-up to real-time medical imaging integration to guide critical surgical procedures. The imaging tool fundamental for all these processes is image registration, which seeks to find the optimal spatial transformation that best aligns the anatomical structures imaged. Currently, Deep Learning-based Image Registration (DLIR) approaches, like VoxelMorph, use U-Net-based architectures to substitute the iterative methods warping steps, with the single-shot registration function learning, from affine/rigid pre-aligned images, thus expanding the applications to time-sensitive clinical scenarios. Despite this, most common approaches apply image registration to the brain, which usually deals with localized deformations, being confined to deformed in the rigid bony structure of the skull. However, other organ transformations, such as the abdominal ones, could show wider deformation ranges and vaster anatomical differences between patients than the brain ones. This aspect may lead the affine/rigid pre-alignment to not be sufficiently near the solution, causing the subsequent deformable matching procedure to converge poorly. Furthermore, loss functions are adopted regardless of image content and permissible deformations, thus producing suboptimal transformations classes. In this context, this thesis proposes AbdoMReg: an unsupervised deformable DLIR framework based on a U-Net architecture specialized for abdominal Magnetic Resonance (MR) images, adopting the VoxelMorph framework as a backbone to make it reliable to more complex deformations than the brain ones. AbdoMReg proposes data engineering tools to provide normalized and synthetic data generation, monomodal and multimodal intrapatient and monomodal interpatient image registrations, novel loss function terms to correct the interpatient larger organs displacements and inference tools to estimate the registration of new image pairs with or without pre-alignment of an organ of interest. Compared to state-of-the-art iterative and DLIR methods, AbdoMReg leads up to 10% Dice Similarity Coefficient (DSC) improvement, when dealing with the high variability of abdominal deformations and organs' misalignments, as well as 40 times faster registration performances than iterative methods, thus potentially expanding the medical application scenario even to the most critical use-cases.