Development of a radiomic model using machine learning for breast lesion classification and breast cancer subtyping

Background Breast cancer is a heterogeneous disease requiring accurate diagnosis and classification for optimal treatment. Conventional imaging methods, such as mammography and ultrasound, have limitations in differentiating benign from malignant lesions and in classifying molecular subtypes. Magnetic Resonance Imaging (MRI), combined with radiomics and machine learning (ML), offers a promising non-invasive approach to improving diagnostic accuracy and tumor subtyping. Purpose This study aimed to develop and validate radiomic models for distinguishing between benign and malignant breast lesions and classifying malignant lesions into molecular subtypes (Luminal A, Luminal B, HER2-positive, and Triple-Negative). Materials and Methods This research was conducted on a dataset of 347 patients, including both retrospective and prospectively collected cases. Radiomic features were extracted from manually segmented lesions using the TRACE4Research™ platform. Five ML models were developed: Logistic Regression, Random Forest, k-Nearest Neighbors, Support Vector Machine, and Multi-Layer Perceptron. Internal validation was performed on a prospective cohort (n=47), while external validation included an independent dataset (n=50). Model performance was evaluated using ROC-AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Results The Benign vs. Malignant model demonstrated excellent performance, achieving a ROC-AUC of 88%. The model maintained a high ROC-AUC of 86% in external validation. The Luminal vs. Non-Luminal model performed well in internal validation (ROC-AUC 84%) but showed a decline in specificity from 74% (internal) to 63% (external validation). The Luminal A vs. Luminal B model exhibited limited discriminatory power (ROC-AUC 74%), indicating significant challenges in distinguishing these subtypes radiologically. The HER2+ vs. Triple-Negative model retained acceptable reliability, with ROC-AUC 81%. Conclusion Radiomic models, particularly the Benign vs. Malignant classifier, showed high diagnostic accuracy, reinforcing the potential of radiomics in non-invasive breast cancer detection. This approach could support clinical decision-making by providing additional quantitative information, optimizing patient management, and minimizing unnecessary interventions. However, molecular subtyping models exhibited a progressive decline in performance, emphasizing the complexity of distinguishing tumor subtypes based solely on imaging. Future research should explore larger, multi-center datasets and integrate radiomics with molecular biomarkers to enhance precision oncology in breast cancer management.