Background Breast cancer (BC) prognosis and treatment are closely related to the molecular subtype. The study aimed to evaluate radiomics combined with artificial intelligence (AI) as a promising non-invasive approach to preoperatively distinguish Luminal from non-Luminal BC.

Methods This retrospective pilot study included patients with invasive BC who underwent multiparametric breast MRI with diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) sequences (January 2020-January 2022). Radiomic analysis was conducted using the IBSI-compliant Trace4Research™ platform, incorporating manual segmentation, voxel resampling, feature extraction, and feature reduction. Three machine learning (ML) architectures—Random Forest, Support Vector Machine (SVM), and k-Nearest Neighbors (kNN)—were trained using nested 10-fold cross-validation. Model performance was assessed using ROC-AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

Results The study involved 73 patients (37 non-Luminal, 36 Luminal), with additional 24 cases for independent testing. Among various ML models, the SVM classifier demonstrated optimal performance, achieving a ROC-AUC of 75%, accuracy of 73%, sensitivity of 70%, specificity of 75%, PPV of 74%, and NPV of 71%. External validation, albeit limited, indicated promising potential for clinical application. Statistical analysis identified six significant radiomic predictors contributing effectively to subtype differentiation.

Conclusion Despite moderate performance, the results highlighted the significant clinical potential of using radiomics and machine learning analysis of DWI/ADC sequences to non-invasively classify BC molecular subtypes, in order to improve preoperative stratification and personalized treatment. Further validation in larger, multi-center studies, and integration with clinical or genomic data, could enhance diagnostic accuracy and clinical applicability.