Deep Learning Models for Brain Magnetic Resonance Imaging Synthesis

Abstract

Magnetic Resonance Imaging (MRI) is a powerful medical imaging technique to reconstruct detailed images of the internal structure of the human body. MRI is particularly suitable for brain imaging due to its advantages in non-invasively providing high-contrast images of soft tissue and visualizing neural pathways. It offers a comprehensive approach to studying the brain by applying various modalities that reveal information from different aspects, including structural, diffusion, and functional imaging. Image synthesis is a special imaging task that synthesizes images from input images to contain the desired contents. Brain MRI-related imaging synthesis tasks are particularly challenging due to the diverse characteristics of different MRI modalities and challenges in real-world clinical evaluation. Therefore, it is necessary to develop deep learning-based image synthesis techniques that are tailored to the corresponding brain MRI modality, addressing the unique requirements of each specific task. This thesis introduces novel deep learning models to address various synthesis challenges in different brain MRI modalities. Two specific issues, the presence of focal pathologies and suboptimal acquisition protocols, are addressed with several MRI modalities including T1-weighted imaging, Diffusion-weighted imaging (DWI), as well as its two different representation models including diffusion tensor imaging (DTI) and fibre orientation distribution (FOD). The proposed methods achieve state-of-the-art performance on the corresponding synthesis tasks from both computer science and clinical perspectives. The experimental results demonstrate that the proposed synthesis frameworks can improve the performance of downstream tasks such as tissue classification, tractography, and connectome analysis. The extensive evaluations further show the clinical impact of the proposed methods, and the feasibility of applying these methods in clinical practice is investigated and discussed.