Multiparametric MRI-based radiomics for differentiating HER2-negative, HER2-low, and HER2-positive breast cancer.

BACKGROUND

Breast cancer stratification increasingly extends beyond conventional TNM staging and binary human epidermal growth factor receptor 2 (HER2) classification. This shift is driven in part by the therapeutic relevance of HER2-low disease. We investigated whether multiparametric MRI-based radiomics could support pairwise discrimination among HER2-negative, HER2-low, and HER2-positive breast cancers.

METHODS

This retrospective single-center study included 157 patients who underwent breast MRI on two 3.0-T scanners. Radiomic features were extracted from T2-weighted, contrast-enhanced T1-weighted, and subtraction T1-weighted images. Seven classifiers were compared after a stratified 8:2 training-test split. Model development, five-fold cross-validation, and model selection were confined to the training set, and the selected logistic regression model was evaluated on the untouched test set.

RESULTS

Logistic regression had the highest mean training AUC across the three pairwise tasks (0.778) and was selected as the common final classifier. Test-set AUCs were 0.922 (95% CI, 0.750-1.000) for HER2-negative versus HER2-positive, 0.756 (95% CI, 0.529-0.938) for HER2-low versus HER2-positive, and 0.705 (95% CI, 0.493-0.893) for HER2-negative versus HER2-low classification.

CONCLUSIONS

Training-set model selection followed by independent test-set evaluation identified the strongest discrimination between HER2-negative and HER2-positive tumors. Performance was less reliable when HER2-low tumors were involved, particularly for HER2-negative versus HER2-low classification. These findings are exploratory and require independent multicenter validation.