Detecting Sparse Key Instances in Whole Slide Cytology Images via Self-Supervised One-Class Representation Learning – Swarnadip Chatterjee

The detection of sparse key instances in whole slide cytology images poses significant challenges due to extreme class imbalance and the complexity of cellular structures. In this talk, I will present our approach that leverages self-supervised one-class representation learning to address these challenges effectively. By training exclusively on patches from normal slides, we develop robust, interpretable instance-level representations that support accurate slide-level decisions without missing key abnormalities. I will discuss our experiments on our oral cancer whole slide cytology and bone marrow cytomorphology datasets, highlighting methods like Deep SVDD, ItS2CLR, and Contrastive Self-Supervised One-class Representation Learning. Our results demonstrate superior performance in identifying rare abnormal instances through strong and weak augmentation strategies. Additionally, we introduce a ranking-based recommendation system designed to assist cytologists, ensuring high recall rates for critical instances. This approach paves the way for more reliable, interpretable, and efficient diagnostic support in digital cytology.

About Swarnadip Chatterjee