Tumor Region Segmentation on Histopathology Whole Slide Images Using Uncertainty-Aware Contrastive Learning

Abstract

In this thesis, I propose a novel contrastive learning technique that effectively learns histopathology whole slide image tumor segmentation using scribble annotation and derived self-label information. Since a careful cell-level examination leads to accurate treatments, it is a critical procedure in clinical diagnoses that pathologists have close observations on tissue samples. However, this visual inspection process is extremely tiring and error-prone even for experienced professionals due to the size and the information density of histopathology whole slide images. Thus, it would be very useful and expandable if the model could effectively learn from scribbles from medical experts. The recently proposed semi- or weakly-supervised learning method has shown the potential to significantly reduce the complexity of diagnosis and labeling on medical data. Although many of the previous methods are promising, it is still hard to find a straightforward scribble-supervised learning framework for tumor segmentation on histopathology data. Inspired by the recent related work, \texttt{Scribble2Label}, the target data domain of the previous work is extended and pixel latent information is introduced that takes into account the uncertainty and leverages scribble annotations maximally to boost the model training. The model trained by the proposed method learns not only by comparing outputs to the given scribbles but also by measuring the relative distance of pixel latent from them. I demonstrate the performance enhancement with the proposed method compared to the bottom baseline, which is the previous work I had inspired from, and the upper baseline trained with full annotations. Furthermore, I show the ablation studies comparing robustness according to the amount of scribbles by several sample rates.