An Apparent Diffusion Coefficient-Based Radiomics Tool for Clinically Significant Prostate Cancer Detection in...

The clinical management of Prostate Imaging Reporting and Data System (PI-RADS) 3 lesions remains controversial due to their equivocal likelihood of clinically significant prostate cancer (csPCa). This study aimed to develop and validate a prediction model for csPCa detection to assist in clinical decision-making for these indeterminate lesions.

We retrospectively collected 443 patients from Hospital 1 (May 2017 to July 2023) for model development. The model was subsequently validated externally using a prospective cohort (n = 206) from the same institution and a retrospective cohort (n = 197) from five subcenters (August 2023 to February 2025). All lesions were confirmed by magnetic resonance imaging-transrectal ultrasound targeted biopsy. Regions of interest were delineated and reviewed by two radiologists in consensus. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and additional metrics derived from ROC curves, precision-recall (PR) curves, calibration plots, and decision-curve analysis (DCA) in both training and validation sets.

Among 846 patients (mean age 66.2 ± 7.5 years), 24.6% had csPCa. The radiomics model achieved an AUC of 0.939 (95% confidence interval [CI]: 0.914-0.963) in the training set and 0.782 (95% CI: 0.731-0.832) in the validation set. The area under the PR curve was 0.883 (training set) and 0.480 (validation set). The Brier score for model calibration was 0.126 (training set) and 0.152 (validation set). Per DCA, the model provided a net benefit for determining csPCa in the validation set at thresholds of 0.3-0.6.

The apparent diffusion coefficient-based radiomics model effectively detected csPCa in PI-RADS 3 lesions in the training set and demonstrated moderate performance in the external validation set.