Multimodal Deep Learning for Medical Imaging: A Survey and A New Approach to Brain Tumor Segmentation with Incomplete Data

Multimodal MRI is crucial for brain tumor segmentation, but its clinical use is hampered by the "missing modality problem," where incomplete data degrades model performance and deployment. This thesis introduces the Grouped Modality Distillation Transformer (GMD-Trans), a novel, fully supervised framework designed to be inherently robust to this challenge. The GMD-Trans architecture uses a pure 3D Vision Transformer (ViT) backbone for global context modeling and a dual-stream encoder for synergistic modality groups. Features are integrated via a cross-attention mixer (IG-CAM). Robustness is achieved through a teacher-student knowledge distillation (KD) scheme guided by the mathematically stable Hölder Divergence to ensure performance even when key modalities are absent. Evaluated on the BraTS 2021 benchmark with randomly missing modalities, GMD-Trans achieves a state-of-the-art Dice score of 82.1% for the Tumor Core (TC), surpassing strong baselines. Ablation studies confirm the efficacy of the proposed methods. This specialized success, however, reveals a performance trade-off, with lower accuracy on the Enhancing Tumor (ET) region. GMD-Trans provides a powerful and efficient solution for robustly segmenting the main tumor body from incomplete data using a fully supervised paradigm, without needing complex pre-training or data synthesis. This work advances the development of dependable AI tools for real-world neuro-oncology.