Weakly supervised contrastive learning infers molecular subtypes and recurrence risk of breast cancer from pathology images
02:06 Share

Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2023.04.13.536813v1?rss=1

Authors: Liu, H., Zhang, Y., Luo, J.

Abstract: The integration of deep learning with ultra-high resolution digital pathology has led to significant improvements in the speed and accuracy of tumor diagnosis, while also demonstrating substantial potential to infer genetic mutations and gene expression levels. However, the application of computational pathology in predicting molecular subtypes, prognosis, and recurrence risk remains limited in breast cancer. To address this issue, our paper proposes a weakly supervised contrastive learning method. This method first performs contrastive learning pre-training on a large-scale unlabeled pathological image to extract tile-level features. The tile-level features are then aggregated using a gated attention mechanism to obtain slide-level features. During the downstream tasks, the feature extractor is fine-tuned using weak supervisory signals. To confirm the effectiveness of this method, we conducted extensive experiments, including tumor diagnosis, gene expression level prediction, molecular typing, recurrence, and drug response prediction, as well as a prognostic risk score. Our findings indicate that our method can accurately identify tumor regions using only weakly supervisory signals and performs exceptionally well in the prediction of breast cancer-related gene expression levels, as well as multiple clinically relevant tasks. Additionally, the tile-level attention scores learned in downstream tasks are used to visualize spatial heat maps that are highly consistent with pathologist annotations and spatial transcriptome. These results demonstrate that the deep learning-based model effectively establishes the high-order genotype-phenotype associations, thereby enhances the potential of digital pathology in clinical applications.

Copy rights belong to original authors. Visit the link for more info

Podcast created by Paper Player, LLC