Towards Efficient and Robust Deep Learning for Medical Image Analysis

Medical imaging underpins modern healthcare by enabling non-invasive visualization for diagnosis, detection, treatment planning, and intervention. With rapid advances in deep learning (DL), substantial progress has been achieved in medical image analysis. However, large-scale clinical deployment remains constrained by two fundamental requirements: efficiency (computational, data, and system scalability) and robustness (reliable performance under heterogeneous, noisy, and imperfect real-world conditions). Addressing both is essential for translating DL models into clinically deployable systems. This thesis develops a unified framework for efficient and robust DL across multiple medical imaging applications. First, pulmonary tree modeling from 3D CT is studied, where structure-aware and implicit representations enable efficient localization and topology repairing of disconnected airways and vessels, supported by a large-scale benchmark of corrupted trees. Second, fMRI-based autism spectrum disorder classification is investigated using compact spatio-temporal representations that avoid costly full 4D modeling, achieving scalable and stable generalization. Third, federated learning under model heterogeneity is addressed via an aggregation-free peer-to-peer framework with similarity-based knowledge distillation, improving communication efficiency, system scalability, and robustness to client heterogeneity and adversarial behaviors. Finally, spatial transcriptomics prediction from H&E whole-slide images is explored through lightweight fusion of spot-level and contextual features, enabling efficient and robust gene expression prediction across slides and subjects. Overall, this thesis advances a coherent design philosophy centered on structure-aware modeling and robustness-oriented learning under real-world clinical constraints, strengthening both the methodological foundations and practical deployability of DL for medical image analysis.