Attention Aware Continual Learning in Digital Pathology Contexts

There has been a recent strategic push globally to integrate digital pathology analysis into national health programs, which has been further encouraged by the emergence of large pathology foundation models which promise to make analysis of such data possible at scale in future. However, large scale clinical deployment of such systems remain a challenge in settings where datasets arrive over extended time, and there are requirements of adaptation to new classes of data over time. Also, in many clinical and pharmaceutical settings, retention of past data is often complicated by privacy issues, legal requirements, or storage constraints. Existing deep architectures are susceptible to decline in performance when trained on new tasks over time, in a phenomenon termed as catastrophic forgetting, addressing which is one of the key challenges towards enabling robust learning systems. This is particularly compounded in digital pathology as often new classes will have heterogeneous samples in limited numbers which need to be learned robustly in an explainable and trustworthy manner, as required by emerging legislation worldwide that call for AI safety and transparency in high risk applications like healthcare. We mitigate catastrophic forgetting by designing enhanced representation learning for histology image analysis, which not only improves continual learning performance under non-stationary data environments, but does so in an interpretable, and hence trustworthy manner.