Coupling Self-Supervised and Supervised Contrastive Learning for Multiple Classification of Cervical Cytological Whole Slide Images

Abstract

Cervical cytologic whole slide image (WSI) multiple classificaton (grading) is a challenging task. Current studies typically ignore the unbalanced data distribution and require multi-class annotations to learn cell features for WSI grading, which largely suffers from label noise. In this paper, we design a three-stage framework to solve these problems. The first stage uses a binary detector and classifier to screen abnormal cells from the gigapixel WSI. By focusing on binary tasks, we alleviate the effects of label noise and data imbalance. To explore the intrinsic characteristics of cervical cells, we use self-supervised learning to acquire comprehensive cell features for subsequent analysis. In the third stage, we propose a well-designed supervised contrastive learning (SCL) framework for WSI grading. To handle the data-imbalance problem, we pre-compute the optimal positions of class centers which are uniformly distributed on the feature space. During training, we perform SCL whilst matching WSIs to their corresponding class centers, which fosters a class-balanced feature space for WSI representations. Extensive experiments on a large-scale dataset demonstrate that our method achieves state-of-the-art performance.