Development and Internal Validation of TRUS Radiomics Models for Patient- and Needle-Level Prostate Cancer Prediction in Biopsy-Naive Men

Background: Current prostate cancer assessment uses clinical findings, PSA, and mp-MRI, while TRUS mainly provides real-time biopsy guidance. We developed zone-specific TRUS radiomics models for patient-level and planned needle-level PCa prediction.

Methods: This retrospective single-center study included 247 biopsy-naive men undergoing freehand transperineal 10-core systematic biopsy from August 2023 to April 2024. Each patient contributed two TZ cores and eight PZ cores. A blinded sonographer delineated whole-zone and pre-firing needle-target ROIs, visually matched to post-firing images acquired seconds later. From each ROI, 513 radiomic features were extracted, sparsely selected, and classified with a two-layer RNN. Models were evaluated using an 80:20 split. A patient or core was positive when pathology showed GS >=6.

Results: The cohort included 116 patients with PCa and 131 with benign findings; 545 of 2,470 cores were positive. Without PSA, the combined TZ+PZ patient-level model had the highest patient-level AUC (0.828), whereas needle-level AUCs were 0.699, 0.698, and 0.680 for TZ, PZ, and TZ+PZ models. With PSA, the combined patient-level model reached an AUC of 0.876, and the needle-level TZ+PSA model reached 0.826.

Conclusions: The findings support further investigation of zone-specific TRUS radiomics, particularly for patient-level prediction. Needle-level results are exploratory and do not establish real-time biopsy guidance. External validation, confidence intervals, comparator models, prospective registration, and patient-grouped evaluation are required before clinical use.