ADC Map Radiomics for Non-Invasive Detection of BRCA1/2 Mutation Status in Prostate Cancer: A Machine Learning-Based Radiogenomics Study

<title>Abstract</title> <p> <bold>Purpose</bold> Conventional multiparametric magnetic resonance imaging (mpMRI) parameters have limited ability to identify prostate cancer (PCa) patients harboring BRCA1/2 mutations. We investigated whether radiomic features extracted from apparent diffusion coefficient (ADC) maps can non-invasively discriminate BRCA-mutated from doubly negative (BRCA−) PCa patients. <bold>Materials and Methods</bold> This prospective cohort study with retrospective imaging analysis included 107 PCa patients (58 BRCA-positive [BRCA+], confirmed by germline and/or somatic testing; 49 doubly negative controls [BRCA−]) at a quaternary referral center (2021–2024). Patients with a negative result in one genetic pathway who were not tested in the other were excluded. Conventional mpMRI parameters (lesion size, mean ADC, zonal location, PI-RADS category, prostate volume) were compared between groups. Radiomic features were extracted from ADC maps using PyRadiomics (107 features; 33 retained after morphological robustness filtering). Twenty-five machine-learning algorithm families were evaluated with stratified k-fold cross-validation. The primary endpoint was area under the precision-recall curve (PR-AUC); secondary endpoint was ROC-AUC. <bold>Results</bold> No significant differences were found in conventional mpMRI parameters between BRCA + and BRCA− groups (mean ADC: 0.631 ± 0.153 vs. 0.675 ± 0.177 × 10⁻³ mm²/s, p = 0.19; lesion size, PI-RADS, zonal location: all p &gt; 0.10). For radiomic analysis (n = 98; 51 BRCA+, 47 BRCA−), the best overall model was CatBoost applied to robustness-filtered radiomic features (PR-AUC = 0.6936; ROC-AUC = 0.6612). Restricting analysis to lesions ≥ 500 voxels (n = 31) improved performance (PR-AUC = 0.778; ROC-AUC = 0.615). <bold>Conclusion</bold> Conventional mpMRI parameters cannot distinguish BRCA-mutated from non-mutated PCa. ADC map radiomics combined with machine learning demonstrated intermediate yet promising discriminatory performance, suggesting that quantitative imaging features may capture biologically relevant information associated with BRCA mutational status. Larger multicenter validation studies are warranted. <bold>Clinical trial number</bold> : not applicable. </p>