Abstract
RATIONALE AND OBJECTIVES: To develop a pathology-derived radiomics signature for detecting clinically significant prostate cancer (csPCa) and to evaluate its performance using lesion diameter-based simplified segmentations. MATERIALS AND METHODS: In this retrospective single-center study, 175 participants (120 radical prostatectomy cases; 55 controls) underwent biparametric MRI during 2013-2022. Whole-mount histopathology was registered to MRI using a patient-specific, mold-based 3D pipeline to generate lesion-level ground truth. Six radiologists from different institutions marked lesion diameters per Prostate Imaging Reporting & Data System (PI-RADS) v2.1, blinded to pathology; automated circular segmentations were generated from these measurements and expanded (±1 slice). Features (PyRadiomics) were preprocessed, filtered, and benchmarked via nested cross-validation. Recursive feature elimination produced a 10-feature pathology-derived radiomics signature. Models (Signature, prostate-specific antigen density [PSAD], PI-RADS, and their combinations) were trained/evaluated using a soft-voting ensemble (logistic regression, random forest, and XGBoost) with patient-level grouping; thresholds were optimized using Youden's J. DeLong's and McNemar's tests were used for model comparisons. RESULTS: PSAD+Signature achieved 0.75 area under the curve (AUC) and 68% (211/312) accuracy. PI-RADS+PSAD achieved 0.77 AUC and 74% (230/312) accuracy. Signature-only achieved 0.66 AUC and 62% (194/312) accuracy. A higher AUC for PSAD+Signature versus Signature (|ΔAUC|=0.093, p=0.012) and for the tripartite model versus Signature (|ΔAUC| = 0.11, p=0.007) was found. PSAD+Signature and PI-RADS+PSAD had similar accuracy (p=0.06). CONCLUSION: A histopathology-trained radiomics signature demonstrated moderate standalone performance for lesion-level csPCa detection. When combined with PSAD, diagnostic performance improved and approached that of PI-RADS+PSAD, which achieved the highest absolute accuracy. The PSAD+Signature framework offers a simplified, spatially localized approach that may complement existing PI-RADS-based assessment while maintaining low implementation complexity.