Adapting to the Unknown: Continual Learning for Unsupervised Anomaly Detection in Medical Imaging

Medical imaging is paramount in modern healthcare, providing essential visual data for screening, diagnosis, and treatment planning across numerous disciplines. The effectiveness of this process is increasingly supported by artificial intelligence, particularly in the domain of Visual Anomaly Detection (VAD). VAD models are being actively researched for medical imagery to automatically identify subtle, novel, or unexpected pathological changes that may evade human detection. However, clinical data streams are inherently non-stationary; the distribution of pathologies, patient profiles, and imaging techniques evolves constantly. Conventional deep learning models, trained on fixed datasets, rapidly suffer from catastrophic forgetting—losing past expertise when updated with new information. To address this fundamental limitation, Continual Learning (CL) offers a promising paradigm, enabling models to incrementally acquire and retain knowledge over time. This thesis proposes to study the VAD problem specifically within the context of medical imaging under this continual setting. Our research investigates various Continual Learning methodologies and their efficacy in training Unsupervised Anomaly Detection (UAD) models that must adapt to an evolving data stream while minimizing the erosion of prior diagnostic capabilities. The overall objective is to design a resilient framework capable of lifelong learning and deployment. The successful implementation of this research will contribute to the development of robust, adaptive AI tools that ensure sustained high performance and reliability for critical diagnostic support in dynamic clinical environments.