Diagnosing breast cancer subtypes using MRI radiomics and machine learning: A systematic review

Objective: This systematic review aims to explore the role of MRI-based radiomics combined with machine learning (ML) algorithms in accurately diagnosing and classifying breast cancer subtypes. It emphasizes the performance of advanced ML models, especially deep learning (DL), in comparison with traditional methods like Support Vector Machines (SVM) and Random Forests. The study focuses on how these models improve diagnostic accuracy for subtypes such as Luminal A/B, HER2-enriched, and Triple-Negative Breast Cancer (TNBC). Methods: A comprehensive search was conducted on October 11, 2024, using PubMed, Scopus, and Web of Science with the keywords “MRI AND Radiomics AND breast cancer.” Only peer-reviewed articles in English were included. Citation tracking via PubMed and Google Scholar ensured complete coverage. Studies involving breast cancer patients, using MRI radiomics (e.g., DCE-MRI, DWI) for subtype classification, and applying supervised ML models were selected. Data from 42 eligible studies were extracted, including radiomic features (texture, shape) and performance metrics (accuracy, sensitivity, AUC). Statistical analysis was conducted with Python libraries, focusing on pooled performance metrics and model generalizability. Results: The review reveals that DL models perform better than traditional ML approaches, particularly in identifying aggressive subtypes like TNBC, achieving an AUC of 0.93 compared to 0.85 for SVM. Texture features emerged as the most significant predictors, contributing 40% to the overall accuracy of the models. Across all subtypes, DL consistently outperformed traditional methods, showing higher accuracy, sensitivity, and specificity. Using multiple MRI sequences further enhanced model performance. Conclusion: This review demonstrates that combining MRI-based radiomics with ML can greatly enhance non-invasive breast cancer diagnosis. DL models, with their ability to capture complex tumor characteristics, offer the most potential for clinical use. However, challenges such as the need for standardized MRI protocols and the limited interpretability of some models must be addressed before these tools can be widely adopted. Future research should focus on large, multicenter studies and the integration of radiomic, genomic, and clinical data to develop more comprehensive precision oncology solutions.