This study represents a multicenter, retrospective analytical investigation aimed at evaluating BRCA mutation status among patients diagnosed with HGSOC. This study was conducted in accordance with the Declaration of Helsinki. The Ethics Committee of the Affiliated Hospital of Qingdao University and Qingzhou People’s Hospital approved this retrospective study and waived informed consent. Patients were enrolled from August 2018 to July 2022. The inclusion criteria specified that participants must have: 1) histologically confirmed HGSOC based on surgical resection; 2) pre-treatment CT images acquired within two weeks prior to the initiation of treatment; 3) available BRCA mutation status; and 4) no prior treatment with chemotherapy, radiotherapy, or targeted therapies prior to imaging acquisition. The exclusion criteria were categorized into technical and clinical factors: 1) technical limitations, including poor contrast enhancement or significant imaging artifacts that precluded high-quality radiomic feature extraction; 2) clinical and biological factors, age < 30 years (as young patients may exhibit distinct tumor biology and clinical presentations); and 3) lesion-specific factors, a maximum tumor diameter of < 2 cm, as small lesions may not provide sufficient volume to reliably capture the spatial heterogeneity required for habitat analysis. After applying these criteria, a total of 228 patients were included in the study, comprising a training cohort of 168 patients from the Affiliated Hospital of Qingdao University and a test cohort of 60 patients from Qingzhou People’s Hospital.

The training cohort was exposed to comprehensive pelvic CT scans using a diverse array of five high-end CT scanners: the Canon AquilionONE, the GE Healthcare Discovery CT750 HD, the GE Optima CT670, the Philips iCT 256, and the Siemens Somatom Definition Flash. The test cohort used two systems: the Philips Brilliance CT 128s and the Siemens Somatom Definition Flash. Notably, significant variations in imaging parameters were observed across scanners, including a range of tube currents from 100 to 300 mA, a consistent fixed voltage of 120 kV, multiplex values spanning 0.599 to 0.984, and a uniform slice thickness of 5 mm. Additionally, the rotation times varied between 0.42 and 0.6 seconds. To enhance image quality, intravenous administration of Iohexol (300 mg iodine/mL) was employed at typical volumes of 85–100 mL, with controlled infusion rates of 2.0–3.0 mL/s. Post-contrast imaging was conducted at 30 seconds (arterial phase), 60 seconds (venous phase), and 90–120 seconds (delayed phase) to capture comprehensive vascular and hemodynamic data.

To address discrepancies in voxel dimensions and window settings across different CT systems, an extensive preprocessing step was implemented. All images underwent re-sampling to a uniform voxel size of 1 × 1 × 1 mm³ using nearest-neighbor interpolation, ensuring preservation of voxel intensity values for subsequent radiomics analysis. Furthermore, standardized window parameters of 350 Hounsfield Unit (HU) (window level) and 35 HU (window width) were applied to ensure consistency across datasets.

Relevant clinical data were collected from hospital information system database, including clinical features such as age, presence or absence of hypertension, presence or absence of diabetes and history of cancer (family history of breast/ovarian cancer or previous history of breast cancer). Tumor-related biomarkers, such as serum cancer antigen 125 (CA125) levels (≤700 U/mL or >700 U/mL) and human epididymis protein 4 (HE4) levels (≤400 pmol/L or >400 pmol/L), were also recorded (19). Imaging characteristics were systematically assessed by a radiologist with >10 years of experience in abdominal imaging. Tumor-related features from CT scans included tumor size (maximum diameter in mm), solid component size (largest dimension in mm), peritoneal disease (PD) pattern (nodular or infiltrative), ascites (presence or absence of ascitic fluid), mesenteric involvement (whether the tumor extended into the mesentery), supradiaphragmatic lymphadenopathy (lymph node involvement above the diaphragm), and bilateral ovarian involvement (whether the tumor was confined to one ovary or involved both ovaries).

Univariate logistic regression was first used to identify clinical and imaging features associated with BRCA status (p < 0.05). Significant features were then included in multivariate logistic regression to identify key predictors of BRCA status (p < 0.05).

The radiomics workflow is shown in Figure 1. Tumor regions of interest (ROIs) were manually annotated on axial CT images by two experienced radiologists using ITK-SNAP software (version 3.8.0). Radiomic features were extracted using PyRadiomics (version 3.0) with Python 3.7.12.

A total of 1,834 radiomic features were extracted from the entire tumor, including first-order statistics, shape features, and texture metrics from the Gray Level Co-occurrence Matrix, Gray Level Run Length Matrix, Gray Level Difference Matrix, Gray Level Size Zone Matrix, and Normalized Gray Level Total Density Matrix.

To evaluate spatial heterogeneity, the ROIs were divided into five sub-regions, or “habitats,” using k-means clustering based on local features such as original_firstorder_Entropy, original_glcm_DifferenceEntropy, original_glcm_JointEnergy, original_glcm_JointEntropy, original_ngtdm_Contrast, and Hu values. Radiomic features were then extracted for each habitat, labeled Habitat 1 to Habitat 5.

Radiomic features were standardized using Z-score normalization to adjust for variations in imaging acquisition parameters, ensuring consistency and reproducibility before proceeding with further analysis and model development.

Feature selection was conducted through a multi-step process to identify the most relevant features for model development. First, intra-class correlation coefficients (ICCs) were calculated to assess reproducibility. Thirty CT images were randomly selected, with tumor regions independently segmented by Reader 1 and Reader 2 to evaluate inter-observer variability, and by Reader 1 for intra-observer variability (initial and repeated segmentation after two weeks). Features with ICCs below 0.8 were excluded to ensure high reproducibility. Statistical analysis was performed using a two-tailed u-test, retaining features with a P-value < 0.05. Spearman’s rank correlation coefficient was computed to identify highly correlated features, with pairs showing a correlation > 0.9 being reduced to a single representative feature. Remaining features were further selected using the least absolute shrinkage and selection operator (LASSO) regression, optimized by 10-fold cross-validation and the 1-SE criterion to determine the optimal regularization parameter (λ). Features with nonzero coefficients were retained in the final model.

We developed predictive models to assess BRCA mutation status using five machine learning algorithms: logistic regression (LR), support vector machine (SVM), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP). The models were trained and validated on the training cohort and the test cohort. A comprehensive feature set was employed, including selected clinical features, whole-tumor radiomic features, and habitat features (habitat1-5). Additionally, combined habitat radiomic features were incorporated to enhance model performance.

The optimal model was selected based on the highest average area under the curve (AUC) across five machine learning algorithms in the training cohort. This multi-algorithmic approach ensures a more robust model selection, as it evaluates performance across different algorithms, providing a balanced and stable assessment of the model’s predictive capabilities.

To validate the optimal model, several advanced evaluation techniques were applied. Receiver operating characteristic (ROC) curves and AUC values were compared for each algorithm (LR, SVM, LightGBM, XGBoost, MLP). Model performance was assessed using metrics like AUC, sensitivity, specificity, positive predictive value, negative predictive value, precision, recall, and accuracy. The DeLong test was used to compare AUC values. Calibration curves evaluated the agreement between predicted and actual outcomes. Finally, decision curve analysis (DCA) assessed clinical utility by evaluating net benefit at various threshold probabilities, highlighting the model’s impact on clinical decision-making and patient outcomes.

The analysis was conducted using SPSS (version 20), custom Python code (version 3.7.12) on Onekey v.4.10.27 platform. Descriptive statistics were employed to summarize continuous and categorical variables, with t-tests and chi-square tests applied for group comparisons. Machine learning models, including LR models, SVM, and MLP, were developed using scikit-learn (version 1.0.2). Advanced gradient boosting models, including LightGBM and XGBoost, were developed using their respective libraries: lightgbm and xgboost.

The clinical characteristics of patients in the training and test cohorts were presented in Table 1. No statistically significant differences were observed between the two groups. Univariate and multivariate logistic regression analyses identified solid size (p=0.041, OR = 1.007, 95% CI = 1.001-1.013) and history of tumor (p<0.01, OR = 11.479, 95% CI = 3.904-33.751) as independent predictors for BRCA status, as detailed in Table 2.

In this study, we conducted an extensive analysis of 40 single-modality models to predict BRCA status, incorporating clinical, radiomic, and different habitats data (habitat1, habitat2, habitat3, habitat4, habitat5, and combined habitat), along with five machine learning classifiers: LR, SVM, LightGBM, XGBoost, and MLP. The heatmap in Figure 2 displays the AUC values for these models. Owing to its strong performance in training and its sustained AUC in the test cohorts, the combined habitat model was selected as the superior method. It demonstrated the highest predictive stability, surpassing the performance of any single-habitat or whole-tumor radiomic models. (Table 3; Figure 3).

In the combined habitat model, we evaluated five machine learning algorithms using ROC curves and the DeLong test (Figure 4). While XGBoost achieved the highest AUC in the training group, there is a significant drop in its performance in the test group, suggesting possible overfitting. This pattern is also observed in the LightGBM and MLP models, which performed well on the training cohort but showed a decline in the test cohort. In contrast, SVM, while not the top performer in the training cohort, exhibited more consistent performance across both cohorts, with an AUC of 0.952 in training and 0.841 in test. LR also demonstrated relatively stable results but with lower performance compared to SVM. Considering overfitting and generalization, the SVM model appears to be the most reliable for assessing BRCA mutation status, given its balanced performance across both training and test cohorts.

In the combined habitat model, five different machine learning models evaluated on the training and test cohorts, the calibration curves revealed varying degrees of model calibration (Figure 5). In the training cohort, all models show relatively good calibration, with LR and SVM displaying the closest alignment to the “perfectly calibrated” line, indicating that their predicted probabilities closely match the observed outcomes. In contrast, LightGBM and MLP appear to show some deviation from the ideal calibration line, suggesting less accurate calibration. For the test cohort, the calibration curves of all five models are similarly distant from the “perfectly calibrated” line, indicating that their calibration performance is more comparable in the test cohort.

We present the DCA performance of five machine learning models in the combined habitat model (Figure 6). In the training cohort XGBoost performs best, followed by SVM and LR, while LightGBM and MLP show lower net benefits. In the test cohort, LR and SVM have the highest net benefits, with XGBoost showing reduced performance, suggesting overfitting. Overall, SVM and LR are the most reliable models, offering consistent net benefits in both cohorts, making them the preferred choices for clinical use.