MB-DSMIL-CL-PL: Scalable Weakly Supervised Ovarian Cancer Subtype Classification and Localisation Using Contrastive and Prototype Learning with Frozen Patch Features

TL;DR

This work tackles scalable weakly supervised classification and localisation of ovarian cancer histological subtypes in whole-slide images using precomputed patch features. It introduces MB-DSMIL-CL-PL, which combines a multi-branch DSMIL with feature-space contrastive learning and prototype-based pseudo-labelling to achieve multiclass performance without end-to-end feature extraction. The method yields substantial gains over DSMIL, with an instance-level macro F1 improvement of approximately $0.212$ absolute (roughly $70.3\%$ relative) and a slide-level macro F1 improvement of about $0.103$ absolute (roughly $15.3\%$ relative), along with notable AUC gains for localisation and slide classification. The approach preserves scalability by relying on frozen patch features while enhancing discriminative power through class-specific attention, contrastive learning, and prototypes, with demonstrated effectiveness on the DROV ovarian cancer dataset. Future directions include multi-resolution extensions and broader applicability to other histopathology tasks such as CAMELYON.

Abstract

The study of histopathological subtypes is valuable for the personalisation of effective treatment strategies for ovarian cancer. However, increasing diagnostic workloads present a challenge for UK pathology departments, leading to the rise in AI approaches. While traditional approaches in this field have relied on pre-computed, frozen image features, recent advances have shifted towards end-to-end feature extraction, providing an improvement in accuracy but at the expense of significantly reduced scalability during training and time-consuming experimentation. In this paper, we propose a new approach for subtype classification and localisation in ovarian cancer histopathology images using contrastive and prototype learning with pre-computed, frozen features via feature-space augmentations. Compared to DSMIL, our method achieves an improvement of 70.4\% and 15.3\% in F1 score for instance- and slide-level classification, respectively, along with AUC gains of 16.9\% for instance localisation and 2.3\% for slide classification, while maintaining the use of frozen patch features.

Figures (6)

• Figure 1: Architecture of the MB-DSMIL-CL contrastive learning. Instance classifier predictions are not used for the MoCo warmup epochs
• Figure 2: Class prototype learning with precomputed instance features in MB-DSMIL-CL-PL.
• Figure 3: Qualitative comparison for an MBT slide (test set). The ground-truth image is identical across methods and shown in the left column for reference. Rows show instance predictions, predictions filtered by GT vs. Normal, and attention maps (scaled using Min-Max normalisation).
• Figure 4: Qualitative comparison for a HGSC slide (test set). The ground-truth image is identical across methods and shown in the left column for reference.
• Figure 5: Qualitative comparison for an SBT slide (test set). The ground-truth image is identical across methods and shown in the left column for reference.