Abstract
Background and purpose
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.
Patients and methods
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).
Results
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.
Conclusions
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.
Keywords: breast MRI, clinical features, external validation, germline BRCA mutation, high-risk breast cancer, machine learning, radiomics
1. Introduction
BRCA mutations represent the most clinically actionable germline alterations in breast cancer. These mutations substantially increase lifetime cancer risk and influence clinical decision-making, including genetic counseling, surveillance strategies, and systemic treatment selection, such as PARP inhibitor therapy (1–5). Accordingly, contemporary clinical guidelines emphasize the importance of identifying BRCA mutation carriers among patients with suspected hereditary breast cancer. However, genetic testing remains resource-intensive and may not be universally accessible, leading to delays or underutilization even among high-risk patients. This highlights the need for practical, noninvasive approaches to prioritize patients most likely to benefit from confirmatory testing.
Breast magnetic resonance imaging (MRI) plays a central role in the evaluation and surveillance of high-risk breast cancer patients, particularly among individuals with BRCA mutations or strong familial risk (6–16). MRI provides detailed information on tumor morphology, enhancement patterns, and the surrounding breast tissue environment. Previous studies have suggested that BRCA-associated breast cancers exhibit distinctive clinicopathologic and imaging characteristics, including higher histologic grade, differences in receptor status, and specific MRI features (7–11, 17). Radiologists routinely interpret MRI findings using standardized descriptors such as those defined in the Breast Imaging Reporting and Data System (BI-RADS). These radiologist-assessed semantic features capture clinically meaningful imaging patterns that reflect tumor biology. However, radiologist interpretation may be inherently subjective and not fully capture subtle imaging patterns associated with genetic alterations.
To complement semantic MRI interpretation, radiomics has emerged as a quantitative imaging analysis approach for extracting high-dimensional features describing tumor shape, intensity distribution, and texture patterns (6, 18). Radiomics enables systematic characterization of tumor heterogeneity and has been increasingly applied in radiogenomics studies that link imaging phenotypes to genomic alterations (6, 18). Two studies have explored MRI-based radiomics to predict BRCA mutation status or hereditary breast cancer risk (19, 20). More recently, multimodal predictive models that combine ultrasound imaging-derived radiomics features with clinicopathologic variables have been proposed to improve predictive performance for identifying germline BRCA mutation (21). However, reported performance varies widely across studies, and generalizability remains a major challenge due to differences in imaging protocols, patient populations, and modeling pipelines. Importantly, most prior work has been developed and evaluated on single-institution datasets or with limited external validation, leaving uncertainty about its robustness in real-world clinical settings. Therefore, a systematic multicenter evaluation that contrasts clinicopathologic-only, radiomics-only, and multimodal predictors under both internal and cross-center external validation is needed.
In this study, we aim to evaluate the predictive value of clinicopathologic variables, radiologist-assessed MRI semantic features, and MRI-derived radiomics features for germline BRCA mutation prediction in high-risk breast cancer patients, using data from two institutions. We develop machine learning (ML) models using each modality individually and in multimodal form and evaluate their performance using both repeated internal validation and cross-center external validation. We hypothesize that integrating clinical, semantic imaging, and radiomics features would provide complementary information and improve prediction performance and generalizability across institutions.
2. Patients and methods
2.1. Patients
This retrospective multicenter study included high-risk breast cancer patients from Center A (Korea University Guro Hospital) and Center B (Korea University Anam Hospital) who underwent preoperative breast MRI and germline BRCA testing. High-risk criteria included early age at diagnosis (under 40 years), triple-negative breast cancer (TNBC), bilateral disease, and/or a family history suggestive of hereditary breast cancer. At each center, eligible patients were initially identified from two sources: (1) Patients with complete structured clinical and histopathologic variables, (2) Patients with available preoperative breast MRI suitable for radiomic feature extraction from 2-minute contrast-enhanced subtraction images (sub-T1WI), which were generated by subtracting pre-contrast series from the second post-contrast series and T2-weighted images (T2WI). Patients with bilateral breast cancer were subsequently excluded during cohort assembly because their tumors could not be analyzed with a single-tumor ROI for radiomics extraction (Figure 1).
Figure 1.
Study flowchart describing cohort construction at each center.
As shown in Figure 1, the cohorts comprise centers A and B with 270 patients (BRCA+: 30; BRCA−: 240), and 222 patients (BRCA+: 33; BRCA−: 189). Institutional review board (IRB) approval was obtained, and informed consent was waived due to the retrospective study design.
2.2. Clinicopathologic features
Clinicopathologic variables were extracted from electronic medical records and included demographic information, histopathologic findings, and immunohistochemical markers. The clinicopathologic variables used in this study included age at diagnosis, histologic grade, lymphovascular invasion, estrogen receptor (ER) status, progesterone receptor (PR) status, Human epidermal growth factor receptor 2 (HER2) status, Ki-67 proliferation index, and molecular subtype classification (Luminal, HER2-enriched, TNBC). These variables were selected because they are routinely available in clinical practice and have been reported to be associated with hereditary breast cancer characteristics.
2.3. Radiologist-assessed MRI features
Radiologist-assessed MRI features were obtained from analysis of two breast radiologists (18 and 9 years of experience) in consensus, according to the 2013 BI-RADS MR lexicon (22). These semantic imaging features represent expert interpretation of tumor morphology and the surrounding breast tissue environment on sub-T1WI and T2WI. The collected radiologist-assessed MRI variables included breast fibroglandular tissue, background parenchymal enhancement, and determination of whether the lesion exhibited mass or non-mass enhancement (NME). For the masses, their shape, margin, and internal enhancement pattern were evaluated. On T2WI, intra-tumoral high signal intensity, peritumoral edema, indicated visually by higher than that of blood vessels or surrounding parenchymal tissues, was assessed. The presence of axillary lymph node enlargement was determined if meeting one or more of the following criteria: abnormal margin or shape of the lymph node, abnormal cortical thickness, or effacement of the fatty hilum (22). These features were used to represent clinically interpretable MRI characteristics evaluated by experienced breast radiologists.
2.4. MRI radiomics features
Two breast radiologists who were blinded to the clinicopathologic data manually drew the intratumor region-of-interest (ROI) of the tumor from the top to the bottom of the tumor for the three-dimensional segmentation of tumors with an open-source software, 3D Slicer (http://www.slicer.org) using sub-T1WI and T2WI. The ROIs manually drawn on sequential slices were rendered into a 3D volume. Radiomics features were extracted from tumor masks using PyRadiomics (version 3.0.1). To improve cross-center comparability, images were resampled to a uniform voxel spacing using interpolation, and gray-level discretization was performed using a fixed bin width. Radiomic feature definitions followed the recommendations of the Image Biomarker Standardisation Initiative (23). For the descriptive baseline tables (Tables 1, 2), one representative feature for each major family was reported to maintain readability. For the predictive machine learning analysis, all radiomics features, including shape, first-order statistics, and texture-matrix families, were used in their entirety as input to all machine learning models, without applying any dimensionality reduction or supervised feature selection.
Table 1.
Baseline characteristics in internal and external validations.
| Center A | Internal validation | External validation | ||||
|---|---|---|---|---|---|---|
| Training (n = 216) | Testing (n = 54) | P value | Center A (n = 270) | Center B (n = 222) | P value | |
| Clinicopathologic features | ||||||
| Age at diagnosis | 45.00 [39.00, 53.00] | 48.00 [37.25, 56.00] | 0.438 | 45.50 [39.00, 53.00] | 44.50 [37.00, 54.75] | 0.452 |
| Age at diagnosis ≤40 years | 78 (36.11%) | 19 (35.19%) | 1.000 | 97 (35.93%) | 91 (40.99%) | 0.264 |
| Histologic grade 3 | 100 (46.30%) | 24 (44.44%) | 0.879 | 124 (45.93%) | 92 (41.44%) | 0.361 |
| Ki-67 index ≥14% | 30 (13.89%) | 7 (12.96%) | 1.000 | 37 (13.70%) | 131 (59.01%) | <0.001 |
| Lymphovascular invasion | 45 (20.83%) | 5 (9.26%) | 0.052 | 50 (18.52%) | 46 (20.72%) | 0.569 |
| ER positivity | 151 (69.91%) | 43 (79.63%) | 0.178 | 194 (71.85%) | 153 (68.92%) | 0.488 |
| PR positivity | 153 (70.83%) | 42 (77.78%) | 0.396 | 195 (72.22%) | 153 (68.92%) | 0.428 |
| HER2 positivity | 31 (14.35%) | 12 (22.22%) | 0.210 | 43 (15.93%) | 51 (22.97%) | 0.051 |
| Molecular subtypes | ||||||
| Luminal | 162 (75.00%) | 46 (85.19%) | 0.147 | 208 (77.04%) | 163 (73.42%) | 0.400 |
| HER2-enriched | 7 (3.24%) | 2 (3.70%) | 1.000 | 9 (3.33%) | 14 (6.31%) | 0.136 |
| TNBC | 47 (21.76%) | 6 (11.11%) | 0.087 | 53 (19.63%) | 45 (20.27%) | 0.910 |
| Radiologist-assessed MRI semantic features | ||||||
| Non-mass enhancement | 8 (3.70%) | 3 (5.56%) | 0.464 | 11 (4.07%) | 13 (5.86%) | 0.404 |
| Dense fibroglandular tissue | 179 (82.87%) | 44 (81.48%) | 0.841 | 223 (82.59%) | 182 (81.98%) | 0.906 |
| Minimal or mild background parenchymal enhancement | 93 (43.06%) | 29 (53.70%) | 0.172 | 122 (45.19%) | 119 (53.60%) | 0.070 |
| Intratumoral high signal intensity on T2-weighted images | 58 (26.85%) | 15 (27.78%) | 0.866 | 73 (27.04%) | 115 (51.80%) | <0.001 |
| Peritumoral edema on T2-weighted images | 166 (76.85%) | 41 (75.93%) | 0.859 | 207 (76.67%) | 146 (65.77%) | 0.009 |
| Axillary lymph node enlargement | 75 (34.72%) | 23 (42.59%) | 0.342 | 98 (36.30%) | 117 (52.70%) | <0.001 |
| Radiomics features (sub-T1WI) | ||||||
| Shape sphericity | 0.68 [0.60, 0.74] | 0.71 [0.64, 0.74] | 0.342 | 0.68 [0.60, 0.74] | 0.67 [0.58, 0.74] | 0.244 |
| Shape voxel volume | 937.00 [389.41, 2,150.87] | 891.25 [376.61, 1,862.98] | 0.882 | 925.22 [385.39, 2,093.82] | 2,737.59 [1,177.02, 7,220.12] | <0.001 |
| First-order entropy | 0.12 [0.06, 0.27] | 0.10 [0.06, 0.29] | 0.802 | 0.11 [0.06, 0.27] | 0.11 [0.05, 0.18] | 0.073 |
| GLCM contrast | 0.02 [0.01, 0.05] | 0.01 [0.01, 0.06] | 0.795 | 0.02 [0.01, 0.05] | 0.01 [0.01, 0.03] | 0.042 |
| GLRLM run entropy | 3.50 [2.99, 3.79] | 3.39 [3.03, 3.73] | 0.772 | 3.48 [3.01, 3.78] | 3.73 [3.35, 4.07] | <0.001 |
| Radiomics features (T2WI) | ||||||
| Shape sphericity | 0.53 [0.47, 0.57] | 0.54 [0.49, 0.58] | 0.171 | 0.53 [0.47, 0.57] | 0.57 [0.52, 0.63] | <0.001 |
| Shape voxel volume | 4,411.46 [1,837.90, 10,039.85] | 4,045.94 [1,912.44, 8,293.76] | 0.618 | 4,337.64 [1,862.56, 9,696.85] | 6,951.98 [2,745.62, 20,566.83] | <0.001 |
| First-order entropy | 0.03 [−0.00, 0.11] | 0.04 [−0.00, 0.11] | 0.675 | 0.03 [−0.00, 0.11] | 0.02 [−0.00, 0.06] | 0.003 |
| GLCM contrast | 0.00 [0.00, 0.02] | 0.01 [0.00, 0.02] | 0.676 | 0.00 [0.00, 0.02] | 0.00 [0.00, 0.01] | <0.001 |
| GLRLM run entropy | 3.33 [2.90, 3.65] | 3.22 [2.95, 3.67] | 0.505 | 3.30 [2.91, 3.66] | 3.50 [3.09, 3.88] | <0.001 |
Table 2.
Univariable and multivariable analysis of germline BRCA mutation in relation to clinicopathologic, radiologist-assessed MRI semantic, and radiomic features in the training set of Center A.
| Center A | Univariable OR | P value | Multivariable OR | P value |
|---|---|---|---|---|
| Clinicopathologic features | ||||
| Histologic grade 3 | 3.19 (1.26, 8.05) | 0.014 | 2.49 (0.86, 7.21) | 0.093 |
| Ki-67 index ≥14% | 0.87 (0.24, 3.13) | 0.835 | ||
| Age at diagnosis | 1.00 (0.96, 1.04) | 0.969 | ||
| Lymphovascular invasion | 1.00 (0.35, 2.84) | 1.000 | ||
| ER positivity | 0.38 (0.16, 0.90) | 0.028 | 0.61 (0.23, 1.65) | 0.332 |
| PR positivity | 0.44 (0.19, 1.04) | 0.062 | ||
| HER2 positivity | 1.68 (0.58, 4.89) | 0.341 | ||
| Molecular subtypes | ||||
| Luminal | 0.51 (0.21, 1.24) | 0.139 | ||
| HER2-enriched | 3.40 (0.62, 18.58) | 0.158 | ||
| TNBC | 1.56 (0.61, 4.03) | 0.354 | ||
| Radiologist-assessed MRI semantic features | ||||
| Non-mass enhancement | N/A | 1.000 | ||
| Dense fibroglandular tissue | 1.04 (0.33, 3.24) | 0.949 | ||
| Minimal or mild background parenchymal enhancement | 0.94 (0.40, 2.22) | 0.884 | ||
| Intratumoral high signal intensity on T2-weighted images | 0.90 (0.34, 2.38) | 0.828 | ||
| Peritumoral edema on T2-weighted images | 3.67 (0.83, 16.17) | 0.086 | ||
| Axillary lymph node enlargement | 0.75 (0.30, 1.90) | 0.545 | ||
| Radiomics features (sub-T1WI) | ||||
| Shape sphericity | 0.51 (0.01, 26.40) | 0.738 | ||
| Shape voxel volume | 1.00 (1.00, 1.00) | 0.757 | ||
| First-order entropy | 1.72 (0.22, 13.62) | 0.606 | ||
| GLCM contrast | 2.41 (0.00, 10,492.79) | 0.837 | ||
| GLRLM run entropy | 1.88 (0.90, 3.95) | 0.093 | ||
| Radiomics features (T2WI) | ||||
| Shape sphericity | 7.54 (0.03, 2,069.95) | 0.481 | ||
| Shape voxel volume | 1.00 (1.00, 1.00) | 0.831 | ||
| First-order entropy | 0.03 (0.00, 4.72) | 0.174 | ||
| GLCM contrast | 0.00 (0.00, 3,424.50) | 0.222 | ||
| GLRLM run entropy | 1.32 (0.67, 2.58) | 0.418 | ||
2.5. Statistical analysis
Statistical analyses were performed to summarize patient characteristics and assess associations between germline BRCA mutation status and clinicopathologic variables, radiologist-assessed semantic features, and MRI-derived radiomics features. Continuous variables are presented as medians with interquartile ranges and were compared using the Wilcoxon rank-sum test, while categorical variables are presented as counts (percentages) and were compared using the chi-square test or Fisher's exact test, as appropriate. To identify features associated with germline BRCA mutation, univariate logistic regression was first performed, and variables with P < 0.05 were entered into multivariate analysis. Statistical analyses were conducted using SPSS version 20.0, R version 3.5.1, and Python version 3.7.4.
2.6. Predictive machine learning analysis
Predictive modeling was performed using six machine learning algorithms to evaluate the predictive performance of clinicopathologic variables, radiologist-assessed MRI features, and MRI-derived radiomics features. We applied logistic regression (LR), random forests (RF), support vector machines (SVMs) with radial basis function kernels, gradient-boosted models (LightGBM and XGBoost), and a multilayer perceptron (MLP). The LR model was used as a baseline linear model because of its interpretability and clinical relevance. Tree-based ensemble methods, boosting models, and the MLP were expected to capture nonlinear relationships and feature interactions. All models were implemented using standardized libraries and consistent preprocessing pipelines across feature sets to ensure a fair comparison. Because the dataset was imbalanced between BRCA mutation-positive and mutation-negative cases, class weights were applied during model training to mitigate bias toward the majority class.
Internal validation was performed to assess model stability and the average predictive performance in Center A. For each experiment, the dataset was randomly split into training (80%) and testing (20%) subsets. This process was repeated 10 times with different random seeds, generating 10 independent train-test splits. For each split, models were trained on the training set and evaluated on the corresponding test set. Performance metrics were averaged across repetitions, and standard deviations were reported to reflect variability due to data partitioning. This repeated random-split strategy was chosen to reduce dependence on a single split and provide a more robust estimate of model performance.
External validation was then performed to assess cross-institutional generalizability. Models were trained exclusively on data from Center A and evaluated on data from Center B. Model fitting was repeated with different random seeds to quantify optimization variability, such as weight initialization and stochastic training, rather than to create independent test samples.
Model performance was primarily evaluated using the area under the receiver operating characteristic curve (AUC). For all experiments, results are reported as mean AUC with corresponding standard deviation. In addition to AUROC, we computed the area under the precision–recall curve (AUPRC), which is more informative than AUROC under marked class imbalance, as well as sensitivity, specificity, and F1-score at the optimal classification cutoff determined by Youden's J statistic on the training set. AUROC was used as the primary metric for model comparison, with the remaining metrics provided as complementary indicators of calibration and operating-point behavior.
2.7. Shapley additive explanations (SHAP) analysis
To improve the interpretability of the machine learning models, SHAP was applied to quantify each feature's contribution to model predictions (24, 32). SHAP values are derived from cooperative game theory and represent the marginal contribution of each feature by considering all possible feature combinations. SHAP summary plots were generated using the test dataset to visualize the distribution of feature contributions across samples. Each point represents an individual patient, with the horizontal axis indicating the magnitude and direction of the contribution and the color representing the feature value. Positive SHAP values indicate increased probability of germline BRCA mutation, whereas negative values indicate decreased probability. To enable direct comparison, identical feature sets and preprocessing pipelines were used across models. SHAP analysis was performed on the best-balanced LR models in each validation setting, selected based on combined consideration of AUROC, AUPRC, sensitivity, and F1-score.
3. Results
3.1. Patient characteristics
Table 1 summarizes the baseline characteristics of patients in both the internal validation scenario (Center A training vs. testing split) and the external validation scenario (Center A vs. Center B). Overall, most clinicopathologic and MRI-derived variables were well balanced between the internal training and testing subsets, indicating that the random split introduced no major baseline differences.
When comparing two centers, the Ki-67 index ≥14%, intra-tumoral high signal intensity and peritumoral edema on T2WI, and axillary lymph node enlargement differed between the two centers (P < 0.05). In addition, several radiomics features showed significant differences across centers (P < 0.05). These findings suggest potential variations in patient populations and imaging characteristics across institutions, highlighting the importance of external validation to assess model generalizability.
3.2. Associations between clinicopathologic, radiologist-assessed MRI semantic features, MRI radiomics features, and BRCA mutation
Table 2 presents the results of the univariable and multivariable logistic regression analyses evaluating the associations between clinicopathologic variables, radiologist-assessed MRI semantic features, MRI radiomics features, and germline BRCA mutation status in the training cohort of Center A.
In the univariable analysis, histologic grade 3 [odds ratio (OR) 3.189, P = 0.014] and lower estrogen receptor positivity (OR 0.381, P = 0.028) were significantly associated with BRCA mutation status. Most radiologist-assessed MRI features and radiomics features were not significantly associated with BRCA mutation status in the univariable analysis.
In the multivariable analysis, none of the variables remained statistically significant. However, histologic grade 3 showed a borderline association with BRCA mutation status (OR 2.49, P = 0.093), suggesting a potential trend toward higher BRCA mutation risk in high-grade tumors.
3.3. Internal validations
Table 3 summarizes the internal validation results for BRCA mutation prediction using unimodal models and their multimodal combinations. Among the unimodal models, the clinicopathologic-based LR model showed the highest AUROC of 0.73, and the radiologist-assessed semantic feature-based models showed the lowest AUROC scores, ranging from 0.40 to 0.54. In addition, radiomics derived from sub-T1WI-based gradient-boosting models, XGBoost and LightGBM, also achieved comparable performance, with AUROCs of 0.72 and 0.71, respectively. Beyond AUROC, AUPRC values were generally low across all models (≤0.40), and cutoff-based metrics (sensitivity, specificity, F1-score) reflected the strong class imbalance, with several radiomics-only and multimodal combinations collapsing to a degenerate operating point (sensitivity or specificity near 0).
Table 3.
Diagnostic performance of 6 machine learning models in internal validations of Center A.
| Metric | Model | CL | RA | CL + RA | RD(T2WI) | RD (sub-T1WI) | CL + RA + RD (T2WI) | CL + RA + RD (sub-T1WI) |
|---|---|---|---|---|---|---|---|---|
| AUROC | LightGBM | 0.64 (0.08) | 0.40 (0.06) | 0.63 (0.07) | 0.64 (0.08) | 0.71 (0.09) | 0.62 (0.10) | 0.63 (0.07) |
| LR | 0.73 (0.08) | 0.54 (0.08) | 0.64 (0.08) | 0.59 (0.09) | 0.65 (0.09) | 0.62 (0.10) | 0.60 (0.15) | |
| MLP | 0.58 (0.16) | 0.49 (0.10) | 0.53 (0.13) | 0.64 (0.12) | 0.69 (0.08) | 0.63 (0.10) | 0.66 (0.10) | |
| RF | 0.67 (0.07) | 0.41 (0.08) | 0.63 (0.07) | 0.61 (0.09) | 0.68 (0.08) | 0.61 (0.09) | 0.67 (0.10) | |
| SVM | 0.63 (0.17) | 0.51 (0.11) | 0.51 (0.17) | 0.47 (0.20) | 0.67 (0.10) | 0.48 (0.17) | 0.57 (0.07) | |
| XGBoost | 0.66 (0.11) | 0.42 (0.07) | 0.63 (0.07) | 0.62 (0.11) | 0.72 (0.12) | 0.62 (0.11) | 0.63 (0.09) | |
| AUPRC | LightGBM | 0.21 (0.12) | 0.09 (0.01) | 0.21 (0.12) | 0.16 (0.05) | 0.26 (0.15) | 0.15 (0.03) | 0.17 (0.05) |
| LR | 0.40 (0.19) | 0.24 (0.16) | 0.32 (0.18) | 0.13 (0.03) | 0.18 (0.07) | 0.15 (0.04) | 0.14 (0.05) | |
| MLP | 0.15 (0.05) | 0.12 (0.06) | 0.17 (0.09) | 0.19 (0.07) | 0.28 (0.11) | 0.17 (0.05) | 0.17 (0.07) | |
| RF | 0.16 (0.03) | 0.10 (0.03) | 0.17 (0.10) | 0.16 (0.07) | 0.21 (0.07) | 0.18 (0.08) | 0.22 (0.11) | |
| SVM | 0.18 (0.10) | 0.16 (0.11) | 0.11 (0.03) | 0.15 (0.10) | 0.25 (0.06) | 0.16 (0.10) | 0.17 (0.10) | |
| XGBoost | 0.19 (0.10) | 0.10 (0.04) | 0.20 (0.13) | 0.15 (0.05) | 0.24 (0.12) | 0.17 (0.07) | 0.18 (0.06) | |
| Sensitivity | LightGBM | 0.22 (0.18) | 0.37 (0.20) | 0.18 (0.17) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
| LR | 0.88 (0.16) | 0.50 (0.28) | 0.63 (0.30) | 0.18 (0.28) | 0.12 (0.18) | 0.22 (0.29) | 0.13 (0.29) | |
| MLP | 0.42 (0.31) | 0.47 (0.20) | 0.20 (0.17) | 0.00 (0.00) | 0.13 (0.17) | 0.00 (0.00) | 0.03 (0.07) | |
| RF | 0.35 (0.25) | 0.37 (0.13) | 0.10 (0.18) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| SVM | 0.68 (0.30) | 0.33 (0.27) | 0.35 (0.31) | 0.05 (0.11) | 0.05 (0.11) | 0.02 (0.05) | 0.03 (0.07) | |
| XGBoost | 0.32 (0.21) | 0.33 (0.16) | 0.23 (0.25) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| Specificity | LightGBM | 0.79 (0.07) | 0.54 (0.10) | 0.89 (0.07) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) |
| LR | 0.57 (0.07) | 0.56 (0.21) | 0.60 (0.11) | 0.87 (0.18) | 0.90 (0.09) | 0.86 (0.18) | 0.91 (0.11) | |
| MLP | 0.73 (0.06) | 0.58 (0.11) | 0.83 (0.06) | 0.98 (0.03) | 0.95 (0.06) | 0.99 (0.01) | 0.95 (0.04) | |
| RF | 0.81 (0.06) | 0.54 (0.09) | 0.91 (0.09) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | |
| SVM | 0.59 (0.24) | 0.65 (0.22) | 0.65 (0.26) | 0.98 (0.05) | 0.98 (0.03) | 0.99 (0.04) | 0.96 (0.05) | |
| XGBoost | 0.81 (0.08) | 0.57 (0.07) | 0.86 (0.09) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | |
| F1 score | LightGBM | 0.15 (0.13) | 0.14 (0.07) | 0.17 (0.15) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
| LR | 0.33 (0.06) | 0.18 (0.10) | 0.25 (0.09) | 0.11 (0.12) | 0.10 (0.14) | 0.12 (0.14) | 0.07 (0.16) | |
| MLP | 0.22 (0.15) | 0.19 (0.09) | 0.15 (0.14) | 0.00 (0.00) | 0.15 (0.19) | 0.00 (0.00) | 0.03 (0.07) | |
| RF | 0.22 (0.13) | 0.15 (0.05) | 0.07 (0.12) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| SVM | 0.27 (0.08) | 0.13 (0.10) | 0.13 (0.11) | 0.05 (0.13) | 0.06 (0.13) | 0.02 (0.05) | 0.04 (0.09) | |
| XGBoost | 0.22 (0.15) | 0.14 (0.06) | 0.16 (0.15) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
Values are presented as mean (standard deviation) across 10 repeated random train–test splits in Center A. To keep the table font size consistent given the number of reported metrics, all values are rounded to two decimal places. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision–recall curve. Sensitivity, specificity, and F1-score were computed at the optimal cutoff determined by Youden's J statistic on the training set. CL, clinicopathologic features; RA, radiologist-assessed MRI semantic features; RD (T2WI), radiomics from T2-weighted images; RD (sub-T1WI), radiomics from 2-minute contrast-enhanced subtraction images.
3.4. External validation
Table 4 presents the external validation results obtained by training models on Center A and testing them on Center B. Among the unimodal models, the clinicopathologic RF model demonstrated the highest AUROC of 0.73, whereas radiologist-assessed semantic models showed lower AUROCs, ranging from 0.54 to 0.58. Radiomics-only models also demonstrated lower external performance, ranging from 0.49 to 0.62. The combined clinicopathologic and radiologist-assessed RF model achieved the highest overall external performance with an AUROC of 0.77, and the multimodal LR model integrating sub-T1WI radiomics features achieved a comparable AUROC of 0.72.
Table 4.
Diagnostic performance of 6 machine learning models in external validations of Center B.
| Metric | Model | CL | RA | CL + RA | RD (T2WI) | RD (sub-T1WI) | CL + RA + RD (T2WI) | CL + RA + RD (sub-T1WI) |
|---|---|---|---|---|---|---|---|---|
| AUROC | LightGBM | 0.66 (0.01) | 0.58 (0.01) | 0.69 (0.01) | 0.51 (0.02) | 0.61 (0.02) | 0.51 (0.02) | 0.62 (0.01) |
| LR | 0.68 (0.00) | 0.57 (0.01) | 0.69 (0.01) | 0.49 (0.01) | 0.59 (0.04) | 0.55 (0.07) | 0.72 (0.00) | |
| MLP | 0.51 (0.04) | 0.56 (0.02) | 0.64 (0.02) | 0.55 (0.03) | 0.57 (0.02) | 0.56 (0.04) | 0.60 (0.03) | |
| RF | 0.73 (0.01) | 0.55 (0.02) | 0.77 (0.01) | 0.49 (0.02) | 0.59 (0.02) | 0.49 (0.01) | 0.59 (0.01) | |
| SVM | 0.71 (0.01) | 0.54 (0.09) | 0.74 (0.03) | 0.53 (0.04) | 0.55 (0.02) | 0.53 (0.04) | 0.54 (0.03) | |
| XGBoost | 0.71 (0.02) | 0.56 (0.01) | 0.72 (0.01) | 0.52 (0.03) | 0.62 (0.01) | 0.55 (0.02) | 0.62 (0.01) | |
| AUPRC | LightGBM | 0.28 (0.01) | 0.20 (0.00) | 0.28 (0.03) | 0.15 (0.01) | 0.20 (0.01) | 0.15 (0.01) | 0.20 (0.00) |
| LR | 0.51 (0.00) | 0.25 (0.01) | 0.50 (0.05) | 0.14 (0.01) | 0.19 (0.01) | 0.17 (0.03) | 0.28 (0.00) | |
| MLP | 0.17 (0.02) | 0.20 (0.04) | 0.28 (0.01) | 0.20 (0.03) | 0.21 (0.02) | 0.20 (0.02) | 0.23 (0.02) | |
| RF | 0.27 (0.01) | 0.17 (0.02) | 0.31 (0.01) | 0.16 (0.01) | 0.22 (0.02) | 0.16 (0.01) | 0.21 (0.01) | |
| SVM | 0.32 (0.03) | 0.22 (0.08) | 0.36 (0.02) | 0.17 (0.02) | 0.19 (0.02) | 0.17 (0.01) | 0.20 (0.02) | |
| XGBoost | 0.28 (0.03) | 0.20 (0.01) | 0.33 (0.02) | 0.16 (0.01) | 0.21 (0.01) | 0.18 (0.01) | 0.22 (0.01) | |
| Sensitivity | LightGBM | 0.19 (0.03) | 0.52 (0.00) | 0.14 (0.03) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
| LR | 0.73 (0.00) | 0.48 (0.00) | 0.72 (0.02) | 0.05 (0.07) | 0.28 (0.19) | 0.15 (0.19) | 0.39 (0.00) | |
| MLP | 0.20 (0.06) | 0.49 (0.01) | 0.26 (0.03) | 0.02 (0.02) | 0.05 (0.05) | 0.01 (0.02) | 0.04 (0.03) | |
| RF | 0.51 (0.08) | 0.48 (0.01) | 0.03 (0.08) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| SVM | 0.43 (0.12) | 0.48 (0.30) | 0.46 (0.23) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| XGBoost | 0.25 (0.08) | 0.48 (0.00) | 0.16 (0.05) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| Specificity | LightGBM | 0.90 (0.02) | 0.51 (0.00) | 0.96 (0.02) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) |
| LR | 0.64 (0.00) | 0.57 (0.00) | 0.65 (0.02) | 0.93 (0.05) | 0.86 (0.11) | 0.90 (0.07) | 0.87 (0.00) | |
| MLP | 0.88 (0.02) | 0.53 (0.00) | 0.94 (0.01) | 0.99 (0.01) | 0.98 (0.02) | 1.00 (0.01) | 0.99 (0.01) | |
| RF | 0.80 (0.03) | 0.54 (0.02) | 0.98 (0.02) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | |
| SVM | 0.83 (0.06) | 0.57 (0.27) | 0.85 (0.11) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | |
| XGBoost | 0.90 (0.02) | 0.54 (0.02) | 0.95 (0.02) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | 1.00 (0.00) | |
| F1 score | LightGBM | 0.22 (0.02) | 0.24 (0.00) | 0.20 (0.03) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
| LR | 0.38 (0.00) | 0.24 (0.00) | 0.39 (0.00) | 0.05 (0.06) | 0.23 (0.12) | 0.13 (0.13) | 0.37 (0.00) | |
| MLP | 0.21 (0.05) | 0.23 (0.00) | 0.32 (0.02) | 0.03 (0.04) | 0.08 (0.08) | 0.02 (0.04) | 0.08 (0.06) | |
| RF | 0.38 (0.03) | 0.23 (0.00) | 0.03 (0.09) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| SVM | 0.36 (0.03) | 0.21 (0.11) | 0.37 (0.07) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | |
| XGBoost | 0.26 (0.06) | 0.24 (0.01) | 0.22 (0.05) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) | 0.00 (0.00) |
Values are presented as mean (standard deviation) across repeated model fits trained on Center A and evaluated on Center B. To keep the table font size consistent given the number of reported metrics, all values are rounded to two decimal places. A degenerate operating point (sensitivity ≈ 0, specificity ≈ 1, F1 ≈ 0) reflects models that classified nearly all external cases as mutation-negative under cross-center distribution shift. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision–recall curve. CL, clinicopathologic features; RA, radiologist-assessed MRI semantic features; RD (T2WI), radiomics from T2-weighted images; RD (sub-T1WI), radiomics from 2-minute contrast-enhanced subtraction images.
Consistent with the AUROC results, cutoff-based metrics confirmed the limited transferability of radiomics features: radiomics-only and several multimodal combinations frequently collapsed to a degenerate operating point in external validation, whereas the clinicopathologic and combined clinicopathologic and radiologist-assessed models retained more balanced sensitivity and specificity.
These results indicate that clinicopathologic and radiologist-assessed MRI features exhibit greater cross-institutional robustness than radiomics features, although sub-T1WI-based radiomics features may provide complementary information in specific settings.
3.5. SHAP analysis
As shown in Figure 2, SHAP analysis was performed on the best-balanced LR models in each validation setting. The LR models were selected because they achieved the highest combined performance across AUROC, AUPRC, sensitivity, and F1-score.
Figure 2.
SHAP summary plots for the best-balanced LR models in internal and external validation. (A) Internal validation: SHAP summary plot for the clinicopathologic LR model (AUROC = 0.73, sensitivity = 0.88, F1 = 0.33). Histologic grade 3 and ER positivity showed the broadest and most consistent SHAP value distributions across samples. (B) External validation: SHAP summary plot for the multimodal LR model integrating clinicopathologic, radiologist-assessed MRI, and sub-T1WI radiomics features (AUROC = 0.72, sensitivity = 0.39, F1 = 0.37). Histologic grade 3 and one sub-T1WI radiomics feature (sub_log-sigma-3-0-mm-3D_glrlm_ShortRunHighGrayLevelEmphasis) showed the broadest and most consistent contributions. In both panels, each point represents an individual patient, with the horizontal axis indicating the magnitude and direction of the SHAP contribution and the color indicating the feature value (red, high; blue, low). Positive SHAP values indicate increased predicted probability of germline BRCA mutation.
For internal validation (Figure 2A), the clinicopathologic LR model (AUROC = 0.73, sensitivity = 0.88, F1 = 0.33) was analyzed. The SHAP summary plot showed that histologic grade 3 was the most influential feature, with high grade associated with an increased predicted probability of BRCA mutation. ER positivity was the second most influential feature, with low ER expression contributing to a higher predicted probability. Remaining variables, including HER2-enriched subtype, TNBC subtype, Luminal subtype, HER2 positivity, lymphovascular invasion, age at diagnosis, PR positivity, and Ki-67 index, showed minimal SHAP contributions. This indicates that internal predictions were primarily driven by histologic grade and ER status.
For external validation (Figure 2B), the multimodal LR model integrating clinicopathologic, radiologist-assessed MRI, and sub-T1WI radiomics features (AUROC = 0.72, sensitivity = 0.39, F1 = 0.37) was analyzed. Histologic grade 3 remained a dominant feature with a clear directional effect, while one sub-T1WI radiomics feature, a GLRLM short-run high gray-level emphasis derived from a Laplacian-of-Gaussian filtered image (σ = 3.0 mm; original PyRadiomics name: sub_log-sigma-3-0-mm-3D_glrlm_ShortRunHighGrayLevelEmphasis), showed a broad and consistent SHAP distribution across samples. Additional sub-T1WI radiomics features derived from wavelet decompositions (a GLSZM small-area high gray-level emphasis, a GLCM cluster shade, and a GLDM dependence entropy) contributed more modestly. The full PyRadiomics feature names are shown in Figure 2B. The stable directional effects observed across samples reflect the additive nature of the LR model and support its robustness and interpretability for cross-center deployment.
4. Discussion and conclusion
In this multicenter study, we evaluated whether clinicopathologic variables, radiologist-assessed MRI semantic features, and MRI-derived radiomic features could predict germline BRCA mutation status in high-risk breast cancer patients. Among the unimodal models, the clinicopathologic model showed the most consistent performance, with AUCs in the 0.6–0.7 range in both internal and external validation, whereas the radiologist-assessed semantic model showed the lowest performance, with AUCs in the 0.4–0.5 range. Among the multimodal combinations, the LR model integrating clinicopathologic, radiologist-assessed MRI, and sub-T1WI radiomics features achieved an external AUROC of 0.72, comparable to the highest unimodal (clinicopathologic RF, 0.73) and bimodal (clinicopathologic + radiologist-assessed RF, 0.77) models, highlighting the value of combining complementary clinical and imaging features for cross-center deployment.
These findings suggest that baseline clinicopathologic information provides a relatively stable predictive framework, whereas qualitative MRI features alone may be insufficient to capture the imaging phenotype of BRCA-mutated tumors across institutions.
Regarding MRI semantic features, several studies have reported that BRCA mutation carriers tend to show distinctive MRI findings, including morphologic features such as round shape, circumscribed margins, rim enhancement, posterior location, and high T2 signal intensity (7–11, 17). However, in our cohort, no MRI semantic feature showed a clear association with BRCA-positive status, indicating that qualitative radiologist-assessed descriptors alone may not reliably distinguish BRCA-mutated tumors.
In contrast, radiomics-based models showed stronger performance in internal validation, particularly when derived from sub-T1WI rather than T2WI, suggesting that subtraction images may capture more informative quantitative features related to BRCA mutation status. This is consistent with previous radiomics studies showing that subtraction-based MRI features can outperform features extracted from the original image series or from T2WI, likely because subtraction images better emphasize enhancement-related tumor heterogeneity (25). Previous work using quantitative MRI for BRCA prediction has mainly focused on manually extracted texture features or CAD-based kinetic descriptors, and only a limited number of studies are available (19, 20). In our study, the use of radiomic features extends this literature by leveraging higher-dimensional imaging information that may not be fully represented by conventional texture analysis alone. Although radiomics-only models performed less consistently in external validation, their internal performance indicates that quantitative imaging features can contribute meaningful information, while also underscoring their sensitivity to inter-institutional variability in acquisition, reconstruction, and segmentation.
Complex non-linear models exhibited limited cross-center generalization. For example, sub-T1WI radiomics-based MLP, XGBoost, and LightGBM achieved internal AUROCs of 0.69, 0.72, and 0.71, but dropped to 0.57, 0.62, and 0.61 in external validation (Tables 3, 4). However, the simpler linear LR model showed the most stable cross-center performance (external multimodal AUROC = 0.72). In small and imbalanced datasets like ours, simpler models are often more reliable when applied to a new institution. For this reason, model selection should be guided by external validation rather than by internal performance alone.
From a practical standpoint, our findings also place the performance of the mpMRI-based models in context with established tools such as BRCAPRO and BOADICEA, which have reported AUCs in the range of approximately 0.72–0.80 (26–28). In that regard, the multimodal MRI-based approach showed comparable discriminative ability while offering a noninvasive, imaging-based alternative that may help prioritize patients for confirmatory BRCA testing and genetic counseling in the preoperative setting. From a clinical perspective, these findings support the potential role of noninvasive machine learning-based risk stratification for prioritizing patients who may benefit most from confirmatory BRCA testing and genetic counseling. This may be particularly useful in settings where genetic testing is costly, unavailable, or delayed. A model that combines readily available clinical information with interpretable MRI features could serve as a practical decision-support tool in the preoperative setting.
Our findings align with recent work on machine learning and explainable AI in breast imaging. Yuan et al. applied SHAP-based interpretation to predict neoadjuvant chemotherapy response and long-term outcomes in breast cancer (29). Sha et al. integrated deep learning features from mammography with SHAP values to predict DCIS recurrence after lumpectomy (30). Yuan et al. further combined deep feature extraction with MRI radiomics for survival prediction following neoadjuvant chemotherapy (31). In contrast, our study targets germline BRCA mutation prediction and provides multicenter external validation with a head-to-head comparison of unimodal, bimodal, and trimodal feature combinations under a unified interpretable framework.
This study has several limitations. First, although external validation was performed, the number of participating institutions was limited, and broader multicenter validation is needed to further establish generalizability. Second, the sample size, especially the BRCA-positive subgroup, was relatively modest, which may limit model stability and the precision of performance estimates. Third, the radiomics-only models showed limited generalizability across centers, as reflected by the significant cross-center differences in radiomics features (Table 1) and the substantial drop in external validation performance. This problem can be attributed to a domain shift due to differences in scanner protocols between the two institutions. Future studies should incorporate image harmonization techniques such as ComBat, together with prospective acquisition standardization, to further improve cross-center transferability. More broadly, complementary strategies, including image acquisition standardization, intensity normalization, and radiomic feature harmonization, can collectively address domain shift and facilitate real-world clinical translation. Supervised feature selection methods, such as LASSO, mRMR, or stability selection, within nested cross-validation could also be used to improve the parsimony and cross-center robustness of radiomics-based models.
In conclusion, this multicenter study suggests that clinicopathologic variables and radiologist-assessed MRI semantic features provide robust and transferable information for predicting germline BRCA mutation status in high-risk breast cancer patients. MRI-derived radiomic features may offer additional complementary value, but their performance is less consistent across institutions. These findings support multimodal, noninvasive risk stratification as a promising approach to identify patients who may benefit from confirmatory genetic testing and counseling.
Acknowledgments
We would like to thank the Advanced Medical Imaging Institute in the Department of Radiology, the Korea University Anam Hospital in the Republic of Korea, and researchers for providing software, datasets, and various forms of technical support.
Funding Statement
The author(s) declared that financial support was received for this work and/or its publication. This research was funded by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute, supported by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2022-KH129266).
Footnotes
Edited by: Lu Zhang, Indiana University Indianapolis, United States
Reviewed by: Jiayi Tan, Cedars Sinai Medical Center, United States
Yupeng Sha, Third Affiliated Hospital of Harbin Medical University, China
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Ethics statement
The studies involving humans were approved by the Institutional Review Board (IRB) of Korea University Anam Hospital, Korea University College of Medicine, Seoul, Republic of Korea (IRB No. 2023AN0368). The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants' legal guardians/next of kin because this was a retrospective study using de-identified data collected as part of routine clinical care. The study posed minimal risk to participants, and the waiver was approved by the Institutional Review Board in accordance with applicable regulations.
Author contributions
YC: Conceptualization, Data curation, Formal analysis, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. EO: Conceptualization, Data curation, Formal analysis, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. YH: Conceptualization, Data curation, Resources, Visualization, Writing – review & editing. KC: Conceptualization, Data curation, Resources, Supervision, Writing – review & editing. KP: Conceptualization, Data curation, Resources, Supervision, Writing – review & editing. SS: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
- 1.Balmaña J, Díez O, Rubio IT, Cardoso F. ESMO guidelines working group. BRCA in breast cancer: ESMO clinical practice guidelines. Ann Oncol. (2011) 22:vi31–4. 10.1093/annonc/mdr373 [DOI] [PubMed] [Google Scholar]
- 2.Lang GT, Shi JX, Hu X, Zhang CH, Shan L, Song CG, et al. The spectrum of BRCA mutations and characteristics of BRCA-associated breast cancers in China: screening of 2,991 patients and 1,043 controls by next-generation sequencing. Int J Cancer. (2017) 141:129–42. 10.1002/ijc.30692 [DOI] [PubMed] [Google Scholar]
- 3.Li J, Wen WX, Eklund M, Kvist A, Eriksson M, Christensen HN, et al. Prevalence of BRCA1 and BRCA2 pathogenic variants in a large, unselected breast cancer cohort. Int J Cancer. (2019) 144:1195–204. 10.1002/ijc.31841 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Dubsky P, Jackisch C, Im SA, Hunt KK, Li CF, Unger S, et al. BRCA genetic testing and counseling in breast cancer: how do we meet our patients’ needs? NPJ Breast Cancer. (2024) 10:77. 10.1038/s41523-024-00686-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Tung N, Garber JE. PARP inhibition in breast cancer: progress made and future hopes. NPJ Breast Cancer. (2022) 8:47. 10.1038/s41523-022-00411-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Grimm LJ, Mazurowski MA. Breast cancer radiogenomics: current status and future directions. Acad Radiol. (2020) 27:39–46. 10.1016/j.acra.2019.09.012 [DOI] [PubMed] [Google Scholar]
- 7.You C, Xiao Q, Zhu X, Sun Y, Di G, Liu G, et al. The clinicopathological and MRI features of patients with BRCA1/2 mutations in familial breast cancer. Gland Surg. (2021) 10:262–72. 10.21037/gs-20-596 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Gabaldó Barrios X, Sarabia Meseguer MD, Marín Vera M, Sánchez Bermúdez AI, Macías Cerrolaza JA, Sánchez Henarejos P, et al. Molecular characterization and clinical interpretation of BRCA1/BRCA2 variants in families from Murcia (south-eastern Spain) with hereditary breast and ovarian cancer: clinical-pathological features in BRCA carriers and non-carriers. Fam Cancer. (2017) 16:477–89. 10.1007/s10689-017-9985-x [DOI] [PubMed] [Google Scholar]
- 9.Kwong A, Wong LP, Wong HN, Law FB, Ng EK, Tang YH, et al. Clinical and pathological characteristics of Chinese patients with BRCA related breast cancer. Hugo J. (2009) 3:63–76. 10.1007/s11568-010-9136-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Noh JM, Han BK, Choi DH, Rhee SJ, Cho EY, Huh SJ, et al. Association between BRCA mutation status, pathological findings, and magnetic resonance imaging features in patients with breast cancer at risk for the mutation. J Breast Cancer. (2013) 16:308–14. 10.4048/jbc.2013.16.3.308 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Ang BH, Ho WK, Wijaya E, Kwan PY, Ng PS, Yoon SY, et al. Predicting the likelihood of carrying a BRCA1 or BRCA2 mutation in Asian patients with breast cancer. J Clin Oncol. (2022) 40:1542–51. 10.1200/jco.21.01647 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Kriege M, Brekelmans CT, Boetes C, Besnard PE, Zonderland HM, Obdeijn IM, et al. Efficacy of MRI and mammography for breast-cancer screening in women with a familial or genetic predisposition. N Engl J Med. (2004) 351:427–37. 10.1056/NEJMoa031759 [DOI] [PubMed] [Google Scholar]
- 13.Kuhl CK, Schmutzler RK, Leutner CC, Kempe A, Wardelmann E, Hocke A, et al. Breast MR imaging screening in 192 women proved or suspected to be carriers of a breast cancer susceptibility gene: preliminary results. Radiology. (2000) 215:267–79. 10.1148/radiology.215.1.r00ap01267 [DOI] [PubMed] [Google Scholar]
- 14.Leach MO, Boggis CR, Dixon AK, Easton DF, Eeles RA, Evans DG, et al. Screening with magnetic resonance imaging and mammography of a UK population at high familial risk of breast cancer: a prospective multicentre cohort study (MARIBS). Lancet. (2005) 365:1769–78. 10.1016/s0140-6736(05)66481-1 [DOI] [PubMed] [Google Scholar]
- 15.Sardanelli F, Podo F, D'Agnolo G, Verdecchia A, Santaquilani M, Musumeci R, et al. Multicenter comparative multimodality surveillance of women at genetic-familial high risk for breast cancer (HIBCRIT study): interim results. Radiology. (2007) 242:698–715. 10.1148/radiol.2423051965 [DOI] [PubMed] [Google Scholar]
- 16.Warner E, Plewes DB, Hill KA, Causer PA, Zubovits JT, Jong RA, et al. Surveillance of BRCA1 and BRCA2 mutation carriers with magnetic resonance imaging, ultrasound, mammography, and clinical breast examination. J Am Med Assoc. (2004) 292:1317–25. 10.1001/jama.292.11.1317 [DOI] [PubMed] [Google Scholar]
- 17.Grubstein A, Rapson Y, Benzaquen O, Rozenblatt S, Gadiel I, Atar E, et al. Comparison of background parenchymal enhancement and fibroglandular density at breast magnetic resonance imaging between BRCA gene mutation carriers and non-carriers. Clin Imaging. (2018) 51:347–51. 10.1016/j.clinimag.2018.06.010 [DOI] [PubMed] [Google Scholar]
- 18.He W, Huang W, Zhang L, Wu X, Zhang S, Zhang B. Radiogenomics: bridging the gap between imaging and genomics for precision oncology. MedComm. (2024) 5:e722. 10.1002/mco2.722 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Vasileiou G, Costa MJ, Long C, Wetzler IR, Hoyer J, Kraus C, et al. Breast MRI texture analysis for prediction of BRCA-associated genetic risk. BMC Med Imaging. (2020) 20:86. 10.1186/s12880-020-00483-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Park H, Cho KR, Lee S, Cho D, Park KH, Cho YS, et al. Prediction of germline BRCA mutations in high-risk breast cancer patients using machine learning with multiparametric breast MRI features. Sensors. (2025) 25:5500. 10.3390/s25175500 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Guo R, Yu Y, Huang Y, Lin M, Liao Y, Hu Y, et al. A nomogram model combining ultrasound-based radiomics features and clinicopathological factors to identify germline BRCA1/2 mutation in invasive breast cancer patients. Heliyon. (2023) 10:e23383. 10.1016/j.heliyon.2023.e23383 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Morris EA, Comstock CE, Lee CH. ACR BI-RADS magnetic resonance imaging. In: D'Orsi CJ, Sickles EA, Mendelson EB, editors. ACR BI-RADS Atlas: Breast Imaging Reporting and Data System, 5th ed. Reston (VA): American College of Radiology; (2013). [Google Scholar]
- 23.Zwanenburg A, Vallières M, Abdalah MA, Aerts H, Andrearczyk V, Apte A, et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. (2020) 295:328–38. 10.1148/radiol.2020191145 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: 31st Annual Conference on Neural Information Processing Systems (NIPS); 2017 (Long Beach, CA: ) (2017). [Google Scholar]
- 25.Fan M, Wu X, Yu J, Liu Y, Wang K, Xue T, et al. Multiparametric MRI radiomics fusion for predicting the response and shrinkage pattern to neoadjuvant chemotherapy in breast cancer. Front Oncol. (2023) 13:1057841. 10.3389/fonc.2023.1057841 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Antoniou AC, Cunningham AP, Peto J, Evans DG, Lalloo F, Narod SA, et al. The BOADICEA model of genetic susceptibility to breast and ovarian cancers: updates and extensions. Br J Cancer. (2008) 98:1457–66. 10.1038/sj.bjc.6604305 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Parmigiani G, Berry DA, Aguilar O. Determining carrier probabilities for breast cancer-susceptibility genes BRCA1 and BRCA2. Am J Hum Genet. (1998) 62(1):145–158. 10.1086/301670 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Schneegans SM, Rosenberger A, Engel U, Sander M, Emons G, Shoukier M. Validation of three BRCA1/2 mutation-carrier probability models myriad, BRCAPRO and BOADICEA in a population-based series of 183 German families. Fam Cancer. (2012) 11:181–8. 10.1007/s10689-011-9498-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Yuan Q, Ye R, Qian Y, Yu H, Zhou Y, Cui X, et al. Machine learning and SHAP value interpretation for predicting the response to neoadjuvant chemotherapy and long-term clinical outcomes in Chinese female breast cancer. Ann Med. (2025) 57:2541316. 10.1080/07853890.2025.2541316 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Sha Y, Yuan Q, Du Y, Yang S, Niu M, Liang X, et al. Integrating deep learning features from mammography with SHAP values for a machine learning model predicting over 5-year recurrence of breast ductal carcinoma in situ post-lumpectomy. Front Immunol. (2025) 16:1681072. 10.3389/fimmu.2025.1681072 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Yuan Q, Hong Z, Ye R, Yang P, Lin J, Jiang X, et al. Integrating deep feature extraction and MRI radiomics for survival prediction in breast cancer after neoadjuvant chemotherapy. Acad Radiol. (2026) 33:872–88. 10.1016/j.acra.2025.10.050 [DOI] [PubMed] [Google Scholar]
- 32.Lundberg SM, Erion GG, Lee SI. Consistent individualized feature attribution for tree ensembles. ArXiv (2018). 10.48550/arXiv.1802.03888 [DOI]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.