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

Introduction

Prostate cancer (PCa) is the most common cancer among men, and the incidence of low-risk disease (ie, International Society of Urological Pathology [ISUP] grade < 2) is high (1). Men with low-risk PCa have a substantial risk of overdiagnosis and overtreatment, making active surveillance an increasingly adopted approach for this group (2,3). In addition to proper inclusion criteria, effective follow-up is equally vital. The ideal follow-up strategy aims to detect high-risk disease early while minimizing unnecessary examinations (4). Current strategies to reduce unnecessary biopsies include the use of an MRI-first approach, with selective biopsies for patients with MRI examinations showing findings positive for PCa (5). A combination of MRI and prostate-specific antigen (PSA) density has shown potential to further reduce the need for biopsies (6).

Risk calculators could play an important role in management of PCa by estimating personalized future risk levels that guide the timing and intensity of subsequent follow-up examinations (7). Although several established clinical tools exist for the primary diagnosis of PCa, such as the Prostate Cancer Prevention Trial (PCPT) version 2 risk calculator (8), only a few studies have explored tools for future risk of progression. One promising tool for estimating PCa risk is the European Randomized Study of Screening for Prostate Cancer (ERSPC) future-risk calculator, which assesses PCa risk after 4 years (9). Unfortunately, currently available risk calculators have limited reliability because of their dependence on clinical variables with high variability.

Although including MRI information into risk models could improve the accuracy and reliability of risk assessment tools, this is challenging because of the lack of image processing capabilities in clinical tools (10,11). Prostate Imaging Reporting and Data System (PI-RADS) (12) scores could serve as an indirect way of including MRI information, but these scores also demonstrate considerable interreader variation (13,14), which may hinder the reliability of risk estimates.

Deep learning (DL) models based on MRI features have the potential to enhance follow-up strategies by training them to capture patterns relevant to the risk of progression, and these models may offer more precise, patient-specific risk estimates for disease progression. Recent studies have demonstrated that DL imaging predictors in other cancer types, including breast cancer (15) and pancreatic cancer (16), are predictive of disease progression. However, to our knowledge, no previous study has explored this approach in the setting of PCa.

Therefore, the present study aimed to develop and validate a DL-based risk model that predicts the time to progression to clinically significant PCa (csPCa, defined as ISUP grade > 1) in patients with low-risk PCa, based on MRI and clinical parameters. We also compared this model with the aforementioned ERSPC future-risk and PCPT risk calculators and PI-RADS classifications.

Materials and Methods

Study Design and Sample

This retrospective study included patients from two health care institutions in the Netherlands: Radboud University Medical Center (RUMC), Nijmegen, and Netherlands Cancer Institute (NKI), Amsterdam. The need for informed consent was waived by the respective institutional review board (RUMC: IRB 2016–3045; NKI: IRB 22–159).

Patients meeting the following criteria were included: (a) underwent prostate MRI (referred to as “baseline MRI”); (b) were suspected of having PCa because of elevated PSA levels, abnormal digital rectal examination findings, or lower urinary tract symptoms; and (c) underwent at least one subsequent biopsy or follow-up MRI 6 months after the baseline MRI. The exclusion criteria were as follows: (a) biopsy-confirmed csPCa before or within 6 months after the baseline MRI; (b) previous treatment for prostate cancer; (c) poor quality of the baseline MRI scan; (d) missing axial T2-weighted or diffusion-weighted imaging; and (e) positive (defined as PI-RADS score ≥ 3) follow-up MRI findings without histopathologic confirmation. Relevant clinical parameters, including PSA, PSA density, prostate volume, patient age, and baseline ISUP grade group, were extracted from patient files.

Multiple MRI examinations for a patient meeting inclusion and exclusion criteria were treated as distinct observations, with each additional scan serving as a new baseline for risk prediction. Figure 1 presents flow diagrams showing patient selection in both the internal and external cohorts, in accordance with Standards for Reporting of Diagnostic Accuracy Studies (17).

Figure 1:

Download as PowerPoint

The study used the ISUP grade group derived from biopsy, prostatectomy, or transurethral resection of the prostate as the reference standard (1). PCa progression was defined as the detection of csPCa at follow-up histopathologic assessment. For patients without documented progression, data points were censored on the date of the last negative radiologic or histopathologic assessment before the cutoff date. Follow-up was truncated at 5 years after MRI.

Imaging Protocol

All patients underwent biparametric or multiparametric MRI of the prostate. For each MRI examination, axial T2-weighted imaging, monoexponential apparent diffusion coefficient maps, and diffusion-weighted images with b value of 1400 sec/mm² or greater were extracted. Dynamic contrast-enhanced imaging was excluded from this study because of its limited role in current prostate MRI guidelines (18,19). Detailed information regarding imaging parameters can be found in Table S1.

Results

Study Sample

After exclusion criteria were applied, 1607 MRI examinations from 1143 male patients (median age, 64 years; IQR, 59–68 years) were included as baseline scans for risk assessment. Progression to csPCa was observed in 113 examinations (13.6%) in the internal cohort and 154 examinations (19.8%) in the external cohort. Median follow-up was 2.1 years (IQR, 1.2–3.2 years) and 2.5 years (IQR, 1.4–4.2 years), respectively. Baseline characteristics are presented in Table 1.

Table 1: Overview of the Patient Characteristics for the Internal and External Cohorts

Time to Progression to PCa with ISUP Grade 2 or Higher

On the internal testing dataset, DL-based risk scores significantly predicted PCa progression (HR, 1.97 [95% CI: 1.61, 2.41]; P < .001). The C-index for DL was 0.68 (95% CI: 0.63, 0.74) (Table 2), which was higher than that for ERSPC (C-index, 0.59 [95% CI: 0.52, 0.65]; P < .001) and PCPT (C-index, 0.59 [95% CI: 0.52, 0.65]; P < .001). Kaplan–Meier curves for the internal testing show the time to progression over 5 years of follow-up (Fig 2).

Table 2: Univariable and Multivariable Cox Regression Results and C-Indexes for All Predictors

Figure 2:

Download as PowerPoint

During external testing, DL scores significantly predicted PCa progression (HR, 1.32 [95% CI: 1.14, 1.55]; P < .001). The C-index was 0.56 (95% CI: 0.51, 0.61) (Table 2). No evidence of differences was found when we compared the predictive performance of DL risk scores to ERSPC (C-index, 0.61 [95% CI: 0.56, 0.66]; P = .11) and PCPT (C-index, 0.56 [95% CI: 0.51, 0.61]; P = .95). PI-RADS significantly outperformed DL scores (C-index, 0.63 [95% CI: 0.59, 0.68]; P = .01). Kaplan–Meier curves for the external testing show the time to progression over 5 years of follow-up (Fig 3).

Figure 3:

Download as PowerPoint

In the multivariable Cox regression analysis, both DL risk scores (HR, 1.66 [95% CI: 1.32, 2.09]; P < .001) and PI-RADS scores were significant predictors of PCa progression. Higher PI-RADS scores were associated with an increased risk: PI-RADS score of 3 to 5 had HRs ranging from 2.18 to 7.96 (all P ≤ .048). In the external cohort, all models except PCPT were significantly associated with progression, with DL risk scores (HR, 1.28 [95% CI: 1.07, 1.53]; P = .006) and ERSPC (HR, 1.60 [95% CI: 1.09, 2.36]; P = .02) showing significant associations, along with PI-RADS scores of 4 and 5 (HRs, 2.44 and 3.14; P ≤ .007).

Optimized Follow-up Schedules

Figure 4 illustrates the optimized follow-up times for the internal evaluation. DL scores identified a patient stratum with the 48% lowest DL risk scores (DL risk group ≤ 2), who could delay follow-up for 3.5 years with less than 10% risk of missing progression to csPCa. Conversely, patients assigned to DL risk group of 5 or higher reached a greater than 10% risk of missed csPCa within 1 year.

Figure 4:

Download as PowerPoint

On external testing, in the 15% of patients with the lowest DL-predicted risk (DL risk group ≤ 1), follow-up could be delayed for 2 years with less than 10% risk of missing progression to csPCa, and for 3 years at a risk of 13% (Fig 4). Patients assigned to DL risk group 2 or higher reached 10% risk after 2 years.

Across both cohorts, the lowest DL risk group had progression risks for csPCa of 3% or less, 8% or less, and 18% or less after 1-, 2-, and 4-year follow-up, respectively. Conversely, patients in DL risk group 5 reached a 29% or greater risk of csPCa progression within 3 years.

Discussion

We introduced a DL model based on clinical and MRI characteristics to predict the risk of csPCa progression and demonstrated potentially useful follow-up schedules for patients with low-risk PCa based on their predicted risk scores. The DL model achieved a promising performance on both internal (C-index, 0.68 [95% CI: 0.63, 0.74]; HR, 1.97 [95% CI: 1.61, 2.41; P < .001]) and external (C-index, 0.56 [95% CI: 0.51, 0.61]; HR, 1.32 [95% CI: 1.14, 1.55; P < .001]) testing datasets in predicting disease progression in comparison to established clinical risk calculators and PI-RADS. In our analysis, follow-up schedules based on DL risk scores would have allowed a substantial proportion of patients to forego follow-up for 3 years with 13% or lower risk of missing disease progression. External testing further demonstrated the model’s prognostic value and showed its generalizability across institutions. Our proposed method could help to reduce the overuse of biopsies and MRI by guiding follow-up frequency in low-risk PCa surveillance. To encourage further research, we provide public access to the code and trained DL models.

Our DL model showed robust prognostic value across internal and external cohorts. In contrast, existing clinical calculators (ERSPC and PCPT) predicted PCa progression only in the external cohort. Despite similar C-indexes among the predictors in external data, the DL risk scores exhibited stronger effect sizes for predicting PCa progression compared with the clinical calculators. This suggests that DL-based risk estimates could enhance the robustness of risk assessments across health care centers compared with existing tools. Although PI-RADS scores showed larger effect sizes for PCa progression than DL scores, their performance may have been inflated because of their inclusion in the reference standard for negative follow-up, known as verification bias.

DL scores also predicted the risk of progression to higher ISUP grading (ie, PCa with ISUP grade ≥ 3). This result is particularly relevant because some active surveillance protocols may allow patients who have PCa with an ISUP grade of 2 to enroll under specific conditions (eg, the absence of cribriform growth [4]) or use less strict criteria (eg, ISUP grade ≥ 3) to trigger active treatment (29,30). Furthermore, surveillance inclusion and termination criteria are still subject to change. The model’s ability to estimate the risk of progression to PCa with ISUP grade 3 or higher suggests the utility of DL risk scores across diverse surveillance protocols.

The proposed DL model may also have value for optimizing patient selection by identifying patients at higher risk of developing aggressive disease. For example, our DL model identified patients at 29% or greater risk of csPCa progression within 3 years. Whether patients with such a high risk of progression can be considered suitable candidates for active surveillance should be decided on an individual basis. Similarly, patients with PI-RADS grade 5 lesions were at considerably higher risk of csPCa detection at follow-up in both cohorts (34% and 16% after 1.5 years, respectively). Currently, none of the published surveillance guidelines incorporate MRI information in their eligibility criteria (4,30,31). Thus, there may be an opportunity to refine inclusion criteria and enhance overall patient outcomes.

We presented optimized follow-up times at a 10% risk level as an illustrative example, inspired by established guidelines for the follow-up of pulmonary nodules detected in lung cancer screening (7). Currently, no such agreed-upon cutoff exists for low-risk PCa management. By providing validated risk levels for different risk calculators, this research promotes discussions on the determination of achievable risk thresholds in low-risk PCa follow-up.

Our results are supported by previous works. In 2020, Wang et al (11) evaluated the PI-RADS score as a marker for PCa progression in 344 patients enrolled in active surveillance in a single center and detected csPCa in 30% of patients classified as low-risk 4 years after baseline MRI. Likewise, a systematic review by Grivas et al in 2022 (32) revealed that higher reported MRI suspicion scores (ie, PI-RADS or Likert scores) were associated with an elevated csPCa detection rate at repeat biopsy following an initial negative biopsy result. We identified only one previous study that used artificial intelligence for risk estimation (Jia et al, 2022 [10]), which showed that MRI radiomics predicted progression-free survival in 191 patients with localized PCa over 3 years. Although these previous works provided evidence for the use of MRI for PCa risk estimation, they were limited by small cohort sizes and lacked comparisons with existing clinical tools. The present study used a substantially larger patient sample and validated the results in a multicenter setting.

Our DL risk estimation may be adapted to clinical purposes beyond PCa progression. Most risk models are based on statistical regression and thus cannot incorporate imaging information directly without manually reducing the image data to features, such as through radiomics extraction or by using PI-RADS scores. We show that imaging data and clinical values can be combined in a multimodal DL model to generate useful risk predictions.

Our study had limitations. First, this study was retrospective and was thus subject to inherent biases associated with the collection of historical data. Second, there was considerable heterogeneity between the baseline pathology grades of the internal and external cohorts, which can be attributed to differences in institutional standards. In general, RUMC avoided biopsies in patients without any visible lesions at MRI, resulting in a considerable proportion of patients included on the basis of negative MRI results. In contrast, NKI typically conducted biopsies before the initial MRI being performed. However, the model’s consistent performance across both cohorts demonstrates robustness to diverse clinical settings. Third, the broad time span for patient inclusion may have led to variability in MRI scan quality, reflecting technologic advancements. Fourth, although some assumptions for proportional hazards models were not met, these deviations can be considered minor, given that the primary aim was to assess predictor significance. Despite a suboptimal fit, all affected predictors showed statistical significance, and their P values would likely have been more noteworthy with correct modeling (28). Nonetheless, caution is advised in interpreting HRs from these models because they may produce erroneous predictions in certain intervals. Finally, our study sample and outcome differed from the intended population and outcome of the PCPT risk calculator. A lack of publicly available future risk models motivated our decision to include this widely used model as a benchmark. In addition, its inclusion provides insights into the applicability of primary diagnostic tools for predicting the time to csPCa progression.

In conclusion, a DL-based prediction model was developed using MRI and clinical parameters and accurately predicted progression to csPCa up to 5 years after the MRI. DL-based risk estimates may help to individualize follow-up times to the risk level of the patient. Prospective studies with longer follow-up times are needed to confirm these results.