Detecting key instances within sparsely populated positive bags through self-supervised one-class representation learning, applied to the analysis of cytology images towards early detection of cancer – Swarnadip Chatterjee

Classification of whole slide pathological images using slide-level labels is a scenario where deep multiple-instance learning is predominantly applied. When confronted with images containing a very large number of instances, such as in the context of whole slide cytology images where each instance (patch) is containing roughly one single cell, the algorithm faces challenges in pinpointing key instances when they are limited in number. In this seminar, I will present our evaluation of the efficacy of utilizing representations learned exclusively from patches extracted from normal slides for making decisions at the instance level. The goal is to achieve interpretable decisions at the slide level for whole slide cytology images without overlooking any key instances. Specifically, we explore the effectiveness of a self-supervised contrastive learning framework known as SimCLR in a one-class-classifier setup, evaluating its capability for domain generalization from a limited number of normal slides. We evaluate the proposed approach on one publicly available cytology dataset and one oral cancer dataset collected in collaboration with Folktandvården Stockholms län AB. I will also present work on exploring the role of contextual information, towards improving deep CNN based oral cancer screening on whole slide cytology images.