Advancing Adaptive and Generalizable Deep Learning for Reliable Medical Computer Vision

Abstract

Biomedical imaging represents a sophisticated and indispensable facet of modern healthcare and scientific research, which could offer an unparalleled insight into the human body’s internal architecture and dynamic processes. The interpretation and analysis of those intricate, non-linear, and high-dimensional measurements used to hugely depend on the specialized knowledge and skills of medical/physiological professionals. Recently, artificial intelligence (AI), is introduced to the healthcare sector and stands out as a revolutionary technique that could significantly enhance the accuracy, efficiency, and breadth of clinical diagnostics and fundamental biomedical research. Despite their transformative potential, these data-driven approaches are frequently criticized for their limited adaptability and generalizability. This deficiency in adaptive capacity can hinder their effectiveness across diverse imaging modalities or patient populations, resulting in variability in predictive outcomes. To tackle these challenges, in this thesis, we present a new framework that combines several novel methods towards building adaptive and generalizable AI systems for driving unbiased biomedical investigations and clinical decision-making based on molecular, pathological, and radiological inspections and measurements. Through extensive evaluations on a broad spectrum of cross-domain settings under miscellaneous data distribution shifts, the suggested method is demonstrated to outperform the state-of-the-art methods by a substantial margin.