Habitat imaging based on DCE-MRI for differentiating luminal and non-luminal subtypes of breast cancer: a two-center study

Objective To explore the application value of habitat imaging based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for differentiating luminal and non-luminal subtypes of breast cancer (BC). Methods Retrospective data from 396 BC patients across two centers were collected. The data from Center 1 were split into a training set of 220 patients and an internal validation set of 56 patients, while Center 2 provided an external test set of 120 patients. Multivariable analysis was performed to identify independent risk factors for developing the clinical model. K-means algorithm was used to perform clustering on DCE-MRI. After feature extraction and selection, eight machine learning algorithms were utilized to build traditional radiomics model, habitat model, and clinical model. A stacking fusion strategy was employed to integrate the traditional radiomics model, habitat model, and clinical model for identifying luminal and non-luminal subtypes patients. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curve and decision curve analysis (DCA). Shapley Additive Explanations (SHAP) was performed for model interpretability. Results For the discrimination of breast cancer luminal and non-luminal subtypes, the stacking model yielded the largest area under the curve value (AUC = 0.840), followed by the habitat model and the conventional radiomics model (AUC = 0.830, AUC = 0.805, respectively), all of which were significantly better than the clinical model (p Conclusion Habitat imaging exhibits promising potential to differentiate luminal and non-luminal breast cancer subtypes. The stacking model integrating habitat-LGBM, traditional radiomics-LGBM and clinical-XGBoost may provide favorable predictive performance.