Improved Histology Image Classification Under Label Noise Via Feature Aggregating Memory Banks

Medical images are often subjectively labeled or weakly supervised, which introduces both closed-set and open-set label noise. Deep learning (DL) models underperform when the quality of supervisory labels degrades. We propose a unified framework for robust training of DL models that takes into account both closed-set and open-set noise, whereas the latter has been largely overlooked. We use a contrastive learning framework and feature aggregating memory banks to identify and increase the emphasis on clean training samples. Based on two publicly available datasets we demonstrate improved results for both synthetically introduced and naturally occurring (unknown) label noise.