Search papers, labs, and topics across Lattice.
This paper introduces MB-DSMIL-CL-PL, a novel weakly supervised method for ovarian cancer subtype classification and localization that leverages contrastive and prototype learning with frozen patch features to improve scalability. The method uses feature-space augmentations to enhance the learning process with pre-computed features. Results show significant improvements over DSMIL, achieving a 70.4% and 15.3% increase in F1 score for instance- and slide-level classification, respectively, and AUC gains of 16.9% for instance localization and 2.3% for slide classification.
Frozen image features can still achieve state-of-the-art weakly supervised classification and localization performance in ovarian cancer histopathology, thanks to contrastive and prototype learning.
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.
