Radiomics and Machine Learning approaches for breast MRI data interpretation

by Luana Conte (Dipartimento di Fisica e Chimica dell’Universita' di Palermo)

Wednesday 23 Jul 2025, 15:00 → 16:00 Europe/Rome

250 (INFN - Pisa)

Abstract: 

Background. Distinguishing between in situ and invasive breast cancer prior to surgery is essential for guiding treatment decisions and surgical planning. Although Radiomics and Machine Learning (ML) have demonstrated potential in enhancing diagnostic accuracy using breast MRI, their specific application to this classification task remains relatively underexplored. This study aimed to assess the performance of various ML algorithms trained on Radiomic features derived from dynamic contrast-enhanced MRI (DCE–MRI), combined with basic clinical data, to classify breast cancer lesions as either in situ or invasive.

Methods. A retrospective dataset of 71 post-contrast DCE–MRI scans was analyzed, including 24 in situ and 47 invasive cases. Tumor regions were manually segmented, and Radiomic features were extracted using the PyRadiomics library. A small set of clinical variables was also incorporated. Multiple ML classifiers were tested using a Leave-One-Out Cross-Validation (LOOCV) approach. Feature selection was performed through three distinct techniques: Minimum Redundancy Maximum Relevance (MRMR), mutual information, and an AUC-based ranking method that progressively added features based on individual ROC AUC performance. Axial 3D rotation was applied for data augmentation.

Results. Among the models tested, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) achieved the highest performance, with maximum accuracy reaching 79% and an AUC of 0.81. SVM and RF offered higher precision and specificity, whereas KNN and XGBoost demonstrated stronger recall and F1-scores, particularly in detecting in situ cases. Notably, KNN showed the most balanced trade-off between sensitivity and specificity, even without data augmentation.

Conclusion. The results indicate that Radiomic analysis of DCE–MRI, when combined with robust ML models, can effectively support the preoperative classification of in situ versus invasive breast cancer. This approach performs well even with limited datasets and could enhance clinical decision-making. Further validation on larger cohorts and integration with complementary imaging or clinical data is encouraged.