CRASH-MIX: cancer region-based aggregation and slide histopathology mixing for enhanced active learning in gigapixel whole-slide image classification

Abstract

Whole-slide imaging (WSI) is a digital technique that enables high-resolution scanning of entire histological slides, thereby creating digitized representations of tissue samples for subsequent analysis. However, this process confronts significant challenges, including the gigapixel size of images and the demand for high-quality annotated data, which often burden pathologists with laborious manual annotation. Additionally, most methods necessitate preprocessing tens of thousands of patches, leading to prolonged training time. In response to these challenges, we introduce CRASH-MIX, an efficient framework designed to eliminate the need for detailed annotations and significantly reduce the number of required slide labels while maintaining high performance. Our approach builds upon the robust foundation provided by the Hierarchical Image Pyramid Transformer (HIPT) and incorporates active learning strategies for more proficient WSI classification. Specifically, CRASH-MIX modifies the vanilla Barlow Twins method by introducing CutMix, thus emphasizing the significance of mixing different slides together to generate more positive pairs during pretraining. To the best of our knowledge, we are the first to demonstrate that fully finetuning slide-level pretrained models significantly enhances performance compared to training from scratch under the same settings. When paired with Manifold Mixup in downstream classification, our proposed framework shows an average improvement of 0.047 across six existing acquisition functions under five different WSI sample budgets. This result is confirmed against a full set of 129 WSI test samples on the Camelyon16 dataset and compared to the modified HIPT in the active learning setting. Our innovative approach greatly enhances the efficiency and effectiveness of WSI classification, particularly under conditions of limited labeled datasets. By reducing reliance on manual labeling through the utilization of unlabeled WSIs and accelerating training time, our approach not only provides considerable benefits for clinical practice but also holds substantial potential to revolutionize the field of digital histopathology.