Towards Proactive Medicine Using AI: Detecting and Addressing Medical Imaging Data Shift

Glaucoma remains a leading cause of irreversible blindness worldwide, currently affecting over 80 million people. Recent advancements in deep learning (DL) technologies and ophthalmic retinal imaging have significantly improved computer-aided glaucoma diagnostics. A key factor in translating these advancements into clinical practice is the ability of DL models to generalize across diverse datasets while maintaining reliable performance. Equally important is their capacity to avoid unwarranted overconfidence in diagnoses. Failure to achieve these objectives poses substantial risks, potentially leading to undiagnosed or misdiagnosed cases that jeopardize patient health. Despite DL advancements in automated glaucoma diagnosis, existing approaches often struggle to maintain performance when confronted with data drift or shifts, leading to potentially harmful and unreliable diagnoses. This thesis first examines the performance of state-of-the-art deep learning methods for glaucoma classification when faced with inevitable data shifts in real-world medical imaging. It then introduces innovative techniques for detecting samples that significantly deviate from training data, thereby enhancing the trustworthiness of existing methods and empowering users to critically evaluate the reliability of DL predictions for informed decision-making in glaucoma. Ultimately, this work presents two distinct solutions aimed at addressing data shift challenges in deep learning predictions, significantly improving the performance of AI-assisted glaucoma diagnosis and potentially advancing the safe integration of these technologies into clinical practice.