Lifelong Machine Learning for Biomedical Image Classification

This dissertation explores a Lifelong Learning framework for AI in healthcare, addressing three key challenges: aligning validation with clinical needs, transferring knowledge across data-scarce environments, and adapting to changing clinical contexts. The approach introduces three metacognitive loops for continuous improvement throughout the AI lifecycle. A structured interview process helps capture the “problem fingerprint” of biomedical applications to better align performance measures with clinical objectives. A similarity measure (bKLD) is proposed to facilitate knowledge transfer between institutions while preserving patient privacy. Additionally, a five-step workflow is developed to adapt models to new deployment environments using unlabeled data. The work is evaluated extensively through experiments on heterogeneous image classification tasks  and structured input from international experts. This research contributes to understanding how AI systems might better adapt and evolve within healthcare environments.