Stacking fusion of DCE-MRI modalities classifies luminal vs non-luminal breast cancer: multicenter, interpretable study

A stacking fusion model combining clinical, radiomics, habitat, and deep learning features from DCE-MRI achieved AUC 0.854 on external test for classifying luminal vs non-luminal breast cancer, with sensitivity 0.839. Deep learning contributed most (48.6% SHAP).

• External validation set (n=120) from a different center.
• SHAP ranked DL-ResNet50 as top contributor (48.6%), and its removal caused the largest AUC decline (-0.033), while removing Radiomics slightly increased AUC (+0.003).
• The stacking model significantly outperformed Clinical-MLP (P<0.001) and DL-ResNet50 (P=0.048) alone.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.