Content area
Artificial intelligence (AI) is transforming the healthcare landscape, offering the promise of earlier diagnoses, more personalized treatments, and improved patient outcomes. However, despite its tremendous potential, deploying AI in real-world clinical settings remains fraught with challenges. Models must operate reliably with limited annotated data, generalize across diverse imaging modalities and patient populations, function efficiently within computationally constrained environments, and remain robust against noise and adversarial perturbations. Overcoming these barriers requires moving beyond narrowly data-driven systems toward AI frameworks that are both technically sophisticated and broadly adaptable to the complexity of clinical practice.
This dissertation introduces a cohesive set of AI-driven methodologies designed to meet these demands. At the center of this work is the development of strategies for medical image reconstruction, semantic segmentation, multimodal knowledge integration, and adversarial robustness. I propose Teach-Former, a multi-teacher knowledge distillation framework that enables lightweight models to absorb rich spatial and contextual representations from multiple large networks, achieving efficient and accurate segmentation across multimodal medical images. To address the limitations of low-resolution imaging, particularly in resource-constrained environments such as space-flight medicine, I introduce Swin-FSR, a Swin Transformer-based super-resolution model that reconstructs high-fidelity fundus images while preserving fine anatomical detail. Recognizing the importance of volumetric imaging, I further developed Swin-VFTR, a transformer architecture capable of segmenting irregular fluid accumulations in 3D optical coherence tomography (OCT) volumes with enhanced precision. Beyond imaging tasks, I tackle the vulnerability of physiological signal analysis through the design of ECG-Adv-GAN and ECG-ATK-GAN, conditional generative adversarial networks that generate realistic adversarial examples and fortify the robustness of ECG arrhythmia classification models. To frame these efforts within a broader security perspective, we developed a game-theoretical framework that models adversarial interactions between attackers and defenders, offering a principled approach to designing robust machine learning systems.
Extensive evaluations across diverse, real-world datasets validate the effectiveness of the proposed approaches, demonstrating improvements in segmentation accuracy, super resolution quality, adversarial defense, and computational efficiency compared to state-of-the-art baselines. Importantly, the models presented here are not confined to data-driven optimization within medical imaging alone. By focusing on generalizable learning strategies such as knowledge transfer, multimodal integration, and robustness against uncertainty, this work lays a foundation for adapting AI solutions to other domains where data is limited, interpretability is essential, and robustness against uncertainty is crucial. Overall, the contributions of this dissertation move us closer to building safe, scalable, and trustworthy AI systems capable of making a meaningful impact across a wide range of scientific and societal challenges.