Ensemble-MIL: deep learning-based ensemble framework for biomarker prediction from histopathological images in colorectal cancer

Recent studies have explored histopathological whole slide images (WSIs) for predicting colorectal cancer (CRC) biomarkers, aiming to create cost-effective and efficient diagnostic tools. However, achieving strong predictive performance and generalizability across datasets remains a challenge. Here, we introduce a deep learning-based ensemble framework, Ensemble-MIL, designed to robustly predict key CRC biomarkers, including BRAF V600E, KRAS mutations, and MSI-H status, with improved cross-dataset performance. We employed two independent CRC datasets: TCGA-COAD for model training and internal evaluation, and CPTAC-COAD as an external test set to assess generalizability. All WSIs were preprocessed and divided into small image patches. A tumor detection model was applied to identify tumor regions, and patch-level; features were extracted via SimCLR, a contrastive learning method. These features were utilized to train three multiple instance learning (MIL) models: Att-MIL, Tran-MIL, and GNN-MIL. The models were then integrated into the final Ensemble-MIL framework. In internal testing with TCGA-COAD, the proposed method achieved area under the curve (AUC) scores of 0.90, 0.87, and 0.64 for MSI-H, BRAF, and KRAS, respectively. In external testing on CPTAC-COAD, it achieved AUCs of 0.78, 0.76, and 0.61 for MSI-H, BRAF, and KRAS, respectively, outperforming previous results. This framework offers a scalable and effective solution for image-based biomarker screening and demonstrates strong potential for clinical application, particularly in resource-limited settings. The code is available at https://github.com/chenh2lab/Ensemble-MIL.