Habitat imaging from DCE-MRI accurately differentiates breast cancer subtypes in two centers

A stacking model integrating DCE-MRI habitat, radiomics, and clinical features differentiated luminal from non-luminal breast cancer with AUC 0.840 (external test set n=120), outperforming habitat-only (0.830) and radiomics-only (0.805) models.

• Multicenter retrospective study of 396 patients, with external validation from a second center (n=120).
• The habitat model (AUC 0.830) itself significantly outperformed the clinical model (p<0.05).
• K-means clustering on DCE-MRI was used to generate habitat features; a stacking ensemble of habitat-LGBM, radiomics-LGBM, and clinical-XGBoost achieved the best performance.

Automated summary

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