Development and Validation of a Deep Learning Model Based on MRI and Clinical Characteristics to Predict Risk of Prostate Cancer Progression

Abstract

Purpose To validate a deep learning (DL) model for predicting the risk of prostate cancer (PCa) progression based on MRI and clinical parameters and compare it with established models. Materials and Methods This retrospective study included 1607 MRI scans of 1143 male patients (median age, 64 years; IQR, 59-68 years) undergoing MRI for suspicion of clinically significant PCa (csPCa) (International Society of Urological Pathology grade > 1) between January 2012 and May 2022 who were negative for csPCa at baseline MRI. A DL model was developed using baseline MRI and clinical parameters (age, prostate-specific antigen [PSA] level, PSA density, and prostate volume) to predict the time to PCa progression (defined as csPCa diagnosis at follow-up). Internal and external testing was performed. The model's ability to predict progression to csPCa was assessed by Cox regression analyses. Predictive performance of the DL model up to 5 years after baseline MRI in comparison with the European Randomized Study of Screening for Prostate Cancer (ERSPC) future-risk calculator, Prostate Cancer Prevention Trial (PCPT) risk calculator, and Prostate Imaging Reporting and Data System (PI-RADS) was assessed using the Harrell C-index. Optimized follow-up intervals were derived from Kaplan-Meier curves. Results DL scores predicted csPCa progression (internal cohort: hazard ratio [HR], 1.97 [95% CI: 1.61, 2.41; P < .001]; external cohort: HR, 1.32 [95% CI: 1.14, 1.55; P < .001]). The model identified a subgroup of patients (approximately 20%) with risks for csPCa of 3% or less, 8% or less, and 18% or less after 1-, 2-, and 4-year follow-up, respectively. DL scores had a C-index of 0.68 (95% CI: 0.63, 0.74) at internal testing and 0.56 (95% CI: 0.51, 0.61) at external testing, outperforming ERSPC and PCPT (both P < .001) at internal testing. Conclusion The DL model accurately predicted PCa progression and provided improved risk estimations, demonstrating its ability to aid in personalized follow-up for low-risk PCa. Keywords: MRI, Prostate Cancer, Deep Learning Supplemental material is available for this article. ©RSNA, 2025.

Keywords: Deep Learning; MRI; Prostate Cancer.

Graphical abstract

Figure 1:

Flow diagrams show patient selection in the (A) Radboud University Medical Center and (B) Netherlands Cancer Institute cohorts. csPCa = clinically significant prostate cancer, PI-RADS = Prostate Imaging Reporting and Data System, TUR = transurethral resection of the prostate.

Figure 2:

Kaplan–Meier curves for time to prostate cancer progression in the internal (Radboud University Medical Center) cohort, stratified by (A) deep learning (DL) predictions, (B) Prostate Imaging Reporting and Data System (PI-RADS), (C) European Randomized Study of Screening for Prostate Cancer (ERSPC) future-risk calculator scores, and (D) Prostate Cancer Prevention Trial (PCPT) risk calculator scores.

Figure 3:

Kaplan–Meier curves for time to prostate cancer progression in the external (Netherlands Cancer Institute) cohort, stratified by (A) deep learning (DL) predictions, (B) Prostate Imaging Reporting and Data System (PI-RADS), (C) European Randomized Study of Screening for Prostate Cancer (ERSPC) future-risk calculator scores, and (D) Prostate Cancer Prevention Trial (PCPT) risk calculator scores.

Figure 4:

Heatmaps show risk of clinically significant prostate cancer (csPCa) progression (International Society of Urological Pathology grade ≥ 2) at follow-up in (A) internal cross-validation and (B) external testing cohorts, stratified by deep learning (DL) risk scores and other prediction models. Models include the European Randomized Study of Screening for Prostate Cancer (ERSPC) future-risk calculator, Prostate Cancer Prevention Trial (PCPT) risk calculator, and Prostate Imaging Reporting and Data System (PI-RADS).