Prediction of germline BRCA mutation using clinicopathologic, MRI semantic, and radiomics features in high-risk...

BRCA mutations are strongly associated with hereditary breast cancer and have important implications for personalized treatment; however, genetic testing may be costly. This highlights the need for noninvasive, practical approaches to prioritize patients most likely to benefit from confirmatory testing. This study evaluated the predictive value of clinicopathologic features, radiologist-assessed magnetic resonance imaging (MRI) semantic features, MRI-derived radiomics features, and their multimodal integration for identifying germline BRCA mutation status.

This retrospective multicenter study included high-risk breast cancer patients from two institutions (Center A and Center B) who underwent preoperative breast MRI and germline BRCA testing. Three types of predictors were used: clinicopathologic features, radiologist-assessed MRI semantic features, and MRI-derived radiomic features. Radiomics features were extracted from tumor masks on 2-minute contrast-enhanced subtraction images and T2-weighted images of preoperative MRI using a standardized, open-source PyRadiomics pipeline. Six machine learning models, including logistic regression (LR), random forest (RF), support vector machine (SVM), gradient-boosted models (LightGBM and XGBoost), and multilayer perceptron (MLP), were performed using unimodal models and their multimodal combinations. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) in both internal validation (10 repeated random train-test splits in Center A) and external cross-center validation (training on Center A and testing on Center B).

A total of 492 patients were included (Center A, n = 270; Center B, n = 222). In internal validation, clinicopathologic-based LR model (AUC = 0.73) and radiomics (sub-T1WI)-based gradient-boosting models (LightGBM and XGBoost) achieved the highest performance (AUC = 0.71 and 0.72, respectively). In external validation, the clinicopathologic-based RF model (AUC = 0.73) and the combined clinicopathologic and radiologist-assessed MRI-based RF model (AUC = 0.77) achieved the highest performance. The multimodal features integrating clinicopathologic, radiologist-assessed MRI, and sub-T1WI radiomics features-based LR model (AUC = 0.72) achieved comparable performance.

A non-invasive machine learning model integrating clinicopathologic, radiologist-assessed MRI features, and radiomic signatures provides complementary predictive information for germline BRCA mutation status in high-risk breast cancer patients.