Title: Mitigating Representation Shift in Unicentric Data through Logit Perturbation for Robust Histology Image Segmentation

Abstract: Semantic segmentation is a crucial task in computational pathology (CPath), but deep learning (DL) methods often struggle to generalise well to different domains due to colour variations and domain shifts. Multicentric data acquisition can help mitigate this challenge, but it is often difficult to obtain large-scale pathology data. In this study, we propose a framework incorporating logit perturbations, a form of logit augmentation that leverages predictive uncertainty, stain normalisation and class weights to improve segmentation performance when trained on limited unicentric annotated data. Our experiments demonstrate that the proposed framework boosts the baseline Dice-score by 6% when trained with unicentric centre data.

Code:

URL: https://openreview.net/forum?id=lfcgIxHsvG