Multi-scale ensemble model for dMMR prediction from histopathological images of colorectal cancer

Colorectal cancer, the second most fatal malignancy globally, burdens public healthcare systems. AI-assisted cancer diagnostics could enable significant cost savings. This study presents a multi-scale ensemble model for DNA mismatch repair deficiency (dMMR) detection from Whole Slide Images (WSIs). dMMR is a clinically important feature, traditionally identified through labor- and time-intensive DNA analysis. The dMMR prediction capability of non-tumorous regions was also evaluated, but it showed limited potential. Therefore, tumorous regions were utilized. The model, comprising two convolutional neural network (CNN) branches and an XGBoost layer, was trained on 1,228 WSIs. It achieved an F ₁ score of 0.863 (sensitivity 0.852) on internal testing, and F ₁ scores of 0.770 (sensitivity 0.868) and 0.743 (sensitivity 0.951) on external test sets of 1,010 and 457 WSIs, respectively. The results indicate that a multi-scale approach can be an effective strategy when developing digital pathology algorithms.