Prostate-specific antigen density (PSAd) is a valuable predictor of prostate cancer in patients undergoing transrectal ultrasound (TRUS) guided prostate biopsy. This study aims to assess the predictive efficacy of PSAd when prostate volume is measured using different methods. We apply deep learning-based autosegmentation to calculate prostate volume and compare its performance to traditional methods with ultrasound and MRI-based ellipsoid volume estimates.

A total of 256 patients who underwent MR/TRUS fusion targeted prostate biopsies between 2020 and 2024 were retrospectively analyzed. 6 patients with a history of transurethral resection of the prostate or those treated with 5α-reductase inhibitors or testosterone were excluded. All patients had preoperative MRI revealing PI-RADS 3, 4, or 5 lesions. Prostate volume was measured using three methods: (1) ultrasound-derived ellipsoid formula, (2) MRI-derived ellipsoid formula, (3) deep learning-based autosegmentation. PSAd was calculated as the ratio of preoperative PSA to the prostate volume determined by each method. The predictive accuracy of each method for biopsy-proven prostate cancer was compared by receiver operating characteristic (ROC) curves and area under the curve (AUC).

The 250 patients were stratified into PI-RADS 3 (n=60), PI-RADS 4 (n=73), and PI-RADS 5 (n=117), with positive biopsy rates of 41.7%, 74.0%, and 92.3%. PSA were 8.26 ± 3.83, 7.54 ± 3.77, and 16.28 ± 23.48 ng/mL across the groups. ROC analysis showed AUCs for PI-RADS 3 lesions of 0.59 (ultrasound), 0.60 (MRI), and 0.62 (autosegmentation). For PI-RADS 4 lesions, AUCs were 0.73, 0.74, and 0.75, and for PI-RADS 5 lesions, they were 0.71, 0.73, and 0.74. Autosegmentation consistently yielded higher AUCs across the groups, indicating improved predictive accuracy for biopsy outcomes compared to traditional methods. For PI-RADS 3 lesions, the optimal threshold for PSAd (autosegmentation) is 0.12 ng/mL², with sensitivity of 84.0% and specificity of 45.9%. For PI-RADS 4 lesions, the threshold is 0.15 ng/mL², achieving sensitivity of 65.5% and specificity of 73.9%. For PI-RADS 5 lesions, the threshold is 0.23 ng/mL², yielding sensitivity of 58.7% and specificity of 80.0%.

Deep learning-based autosegmentation offers a more accurate method for calculating prostate volume, leading to more precise PSAd assessment. This study shows potential for improving detection in MR/TRUS fusion-targeted biopsies, especially for PI-RADS 4 and 5 lesions.