Deep multitask learning for accelerating Magnetic Resonance Imaging

Dimitrios Karkalousos

Keywords: Deep Learning; Multitask learning, Medical Image Analysis; Image reconstruction; Image segmentation; Magnetic Resonance Imaging

Abstract: This dissertation advances the acceleration of Magnetic Resonance Imaging (MRI) through novel Artificial Intelligence approaches combining Deep Learning (DL) and MultiTask Learning (MTL). The first contribution introduces the Cascades of Independently Recurrent Inference Machines (CIRIM), a physics-informed DL network achieving superior performance compared to a conventional reconstruction method, i.e., compressed sensing, in both speed and reconstruction quality. Extensive evaluation across large-scale challenges and clinical applications demonstrated CIRIM's robustness in preserving pathological features in neurological conditions while enhancing image quality through efficient denoising. The second contribution establishes that reconstruction and analysis tasks, specifically segmentation, can be effectively combined through MTL. By performing these tasks simultaneously and sharing learned representations, MTL improves performance while eliminating computational overhead of single-task execution. The MTL framework for joint reconstruction and segmentation showed particular promise in detecting multiple sclerosis lesions while maintaining reconstruction fidelity. The final contribution presents the Advanced Toolbox for Multitask Medical Imaging Consistency (ATOMMIC), implementing Deep MTL (DMTL) for accelerating MRI. ATOMMIC uniquely harmonizes complex-valued and real-valued data processing across tasks, enabling standardized evaluation of various models across multiple tasks. Our findings suggest that DMTL can significantly reduce MRI reconstruction times while enhancing image quality and analysis capabilities. While initial clinical results are promising, larger multi-center studies across diverse patient populations and pathologies are needed to validate DMTL's robustness and generalizability. This dissertation thus presents both novel theoretical advances and practical implementations of DMTL methods for accelerating MRI, with potential to transform clinical workflows through faster and more accurate diagnostic assessments.

Submission Number: 156