Prostate cancer is the most common cancer in men, and a substantial proportion of diagnosed cases are low-risk disease -- slow-growing tumors that may never cause symptoms or require treatment. For these patients, active surveillance (monitoring without immediate treatment) has become a widely adopted strategy to avoid overtreatment while staying ready to act if the cancer progresses.
A critical challenge in active surveillance is knowing when to intensify monitoring and when it is safe to space out follow-up visits. Too much monitoring leads to unnecessary biopsies with their associated risks of pain, bleeding, and infection. Too little risks missing cancer progression to a stage where treatment outcomes are worse.
Existing risk calculators like the ERSPC future-risk calculator (European Randomized Study of Screening for Prostate Cancer) and the PCPT risk calculator (Prostate Cancer Prevention Trial) rely entirely on clinical variables such as PSA and patient age. These tools have limited reliability because they cannot incorporate the rich spatial and tissue information visible on MRI scans.
The PI-RADS scoring system provides a structured way for radiologists to report suspicious findings on prostate MRI, but PI-RADS scores show considerable variation between readers and cannot directly translate complex imaging patterns into a patient-specific probability of future progression. A deep learning model that learns directly from MRI images could provide more objective, image-informed risk estimates.
This retrospective study enrolled 1,143 male patients (1,607 MRI examinations) from two Dutch medical centers: Radboud University Medical Center (RUMC) in Nijmegen and the Netherlands Cancer Institute (NKI) in Amsterdam. Scans were collected between January 2012 and May 2022. RUMC served as the internal cohort for model development and internal testing; NKI served as the external validation cohort.
All included patients were negative for clinically significant prostate cancer (csPCa) -- defined as ISUP grade 2 or higher -- at their baseline MRI, and subsequently underwent at least one follow-up biopsy or MRI at least 6 months later. Progression was defined as detection of csPCa at follow-up histopathology. Patients without documented progression were censored at the date of their last negative assessment, with follow-up truncated at 5 years.
Clinical variables collected for each patient included age, PSA level, PSA density (PSA divided by prostate volume), and prostate volume. Imaging data consisted of axial T2-weighted images, apparent diffusion coefficient (ADC) maps, and diffusion-weighted images. Dynamic contrast-enhanced imaging was excluded, making the model compatible with bi-parametric MRI protocols that avoid contrast injection.
Progression to csPCa occurred in 113 examinations (13.6%) in the internal cohort and 154 examinations (19.8%) in the external cohort. The two cohorts had notable differences in clinical practice: RUMC typically avoided biopsies in patients with MRI-negative findings, while NKI routinely performed biopsies before MRI, creating meaningful heterogeneity that tested the model's generalizability.
The deep learning model has two sequential components. The first is a U-Net, pre-trained to automatically identify regions suspicious for any prostate cancer (ISUP grade 1 or higher) in baseline MRI scans. This U-Net generates voxel-level heatmaps representing the spatial probability of cancer-containing tissue throughout the prostate gland, capturing where and how intensely the imaging suggests abnormal tissue.
The second component is a DL classifier that takes the U-Net's heatmaps together with the four clinical variables (age, PSA, PSA density, prostate volume) and the time interval to follow-up as inputs. This classifier was trained to predict, at the patient level, whether csPCa would be detected by the end of the follow-up period. The combined architecture enables the model to integrate both spatial imaging features and patient-level clinical context.
The model was optimized by minimizing binary cross-entropy loss. During inference, it generates a single continuous DL risk score per patient, computed at a standardized 4-year horizon to facilitate long-term risk estimation. Patients were then grouped into five DL risk groups (analogous to the PI-RADS 1-5 scale) using institution-specific thresholds calibrated on a 5% calibration subset, accounting for differences in scanner parameters and institutional practices.
Five-fold cross-validation was used for internal testing at RUMC, and external testing used predictions averaged across the five trained folds applied to NKI data. All code and trained model weights were made publicly available to encourage further research and enable comparison with future methods.
On internal testing, DL risk scores were strongly predictive of csPCa progression with a hazard ratio (HR) of 1.97 (95% CI: 1.61 to 2.41; P less than 0.001), meaning each one standard deviation increase in the DL score was associated with roughly a doubling of progression risk. The Harrell C-index was 0.68 (95% CI: 0.63 to 0.74), outperforming both the ERSPC calculator (C-index 0.59) and the PCPT calculator (C-index 0.59), with differences reaching statistical significance (both P less than 0.001).
On external testing at NKI, DL scores remained significantly predictive (HR 1.32; 95% CI: 1.14 to 1.55; P less than 0.001; C-index 0.56), confirming the model's generalizability to a different institution. However, PI-RADS significantly outperformed the DL model in the external cohort (C-index 0.63 vs. 0.56; P = 0.01). The authors note this difference may partly reflect verification bias: PI-RADS scores were used to determine which patients received follow-up biopsies in NKI, which could artificially inflate PI-RADS's apparent predictive power in that cohort.
In multivariable Cox regression including both DL scores and PI-RADS simultaneously, DL scores remained independently significant in both cohorts (internal HR 1.66, P less than 0.001; external HR 1.28, P = 0.006). This means DL risk scores provide prognostic information above and beyond what PI-RADS alone captures, and the two measures are complementary rather than redundant.
The model also predicted risk of progression to higher-grade cancers (ISUP grade 3 or above), with C-indexes of 0.70 (internal) and 0.57 (external) -- suggesting the model's predictive ability is not limited to any one progression threshold, which is important given that different active surveillance protocols use different ISUP grade cutoffs for triggering active treatment.
A key clinical application of the model is guiding when patients need their next follow-up MRI or biopsy. The researchers derived optimized follow-up schedules from Kaplan-Meier curves for each DL risk group, inspired by follow-up guidelines for pulmonary nodules detected in lung cancer screening -- an established framework that specifies follow-up intervals based on a defined maximum acceptable risk of missing a progression event.
Using a 10% risk threshold for missing csPCa progression as the decision criterion, patients in DL risk group 1 or 2 (approximately 48% of internal cohort patients) could safely delay follow-up for 3.5 years without exceeding this threshold. In the external cohort, the 15% of patients in the lowest DL risk group could delay follow-up for 2 to 3 years. This means a substantial proportion of patients currently undergoing annual monitoring could potentially have follow-up intervals doubled or tripled without meaningful risk of missed progression.
Conversely, patients in DL risk group 5 -- the highest-risk stratum -- reached a 29% or greater risk of csPCa progression within just 3 years and a greater than 10% risk of missing progression within 1 year. These high-risk patients may warrant accelerated follow-up or early consideration of treatment rather than continued surveillance.
The model also identified approximately 20% of patients with very low absolute progression risks: 3% or less after 1 year, 8% or less after 2 years, and 18% or less after 4 years. Identifying this subgroup with high confidence could meaningfully reduce the cumulative burden of unnecessary biopsies and MRI scans in the active surveillance population.
The core advantage of the proposed approach over existing clinical calculators is the ability to incorporate MRI imaging data directly, without first reducing it to summary statistics like PI-RADS scores. Most clinical risk models are built from statistical regression and can only use manually extracted features. This deep learning framework processes the full spatial pattern of the MRI-derived heatmaps, capturing subtle tissue characteristics that summary scores necessarily discard.
The model's performance was consistent across two institutions with different clinical practices and patient populations, supporting its practical robustness. This is notable because the two cohorts differed substantially in how often biopsies were performed before MRI and in the distribution of baseline ISUP grades -- differences that could introduce bias in simpler models.
The authors highlight that current active surveillance guidelines from major professional societies do not incorporate MRI findings in eligibility or follow-up criteria. This model suggests there is an opportunity to refine those protocols by incorporating both baseline MRI patterns and clinical parameters, moving toward truly personalized surveillance strategies rather than one-size-fits-all annual monitoring intervals.
This study provides proof-of-concept that a multimodal deep learning model combining MRI imaging features and standard clinical parameters can predict progression from low-risk to clinically significant prostate cancer across a 5-year horizon. The model's ability to stratify patients into meaningfully different risk groups could help clinicians individualize surveillance intensity, reducing unnecessary testing in low-risk patients while intensifying monitoring for high-risk individuals.
Key limitations include the retrospective design, the relatively short median follow-up of about 2 years for many patients (which limits observation of late progressors), and the use of biparametric MRI in some patients but multiparametric MRI in others. The broad time span of patient inclusion also introduced variability in MRI scan quality as imaging technology evolved over the decade of enrollment.
The performance gap between internal (C-index 0.68) and external (C-index 0.56) testing indicates the model is more powerful in the center where it was trained, with some performance loss during external transfer -- a common challenge in medical imaging AI. Prospective multicenter validation with standardized imaging protocols and longer follow-up will be essential before this model can be incorporated into clinical surveillance guidelines.