cs.CVMar 15, 2026

Deep Learning From Routine Histology Improves Risk Stratification for Biochemical Recurrence in Prostate Cancer

Authors: Clément Grisi, Khrystyna Faryna, Nefise Uysal, Vittorio Agosti, Enrico Munari, Solène-Florence Kammerer-Jacquet, Paulo Guilherme de Oliveira Salles, Yuri Tolkach, 

Organizations: Department of Pathology, Radboud University Medical Center, Nijmegen, The Netherlands · Department of Molecular and Translational Medicine, University of Brescia, Brescia, Italy · Surgical Pathology Unit, Verona University Hospital Trust, Verona, Italy · Department of Pathology, CHU de Rennes, Rennes, France · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Accurate prediction of biochemical recurrence (BCR) after radical prostatectomy is critical for guiding adjuvant treatment and surveillance decisions in prostate cancer. However, existing clinicopathological risk models reduce complex morphology to relatively coarse descriptors, leaving substantial prognostic information embedded in routine histopathology underexplored. We present a deep learning-based biomarker that predicts continuous, patient-specific risk of BCR directly from H&E-stained whole-slide prostatectomy specimens. Trained end-to-end on time-to-event outcomes and evaluated across four independent international cohorts, our model demonstrates robust generalization across institutions and patient populations. When integrated with the CAPRA-S clinical risk score, the deep learning risk score consistently improved discrimination for BCR, increasing concordance indices from 0.725-0.772 to 0.749-0.788 across cohorts. To support clinical interpretability, outcome-grounded analyses revealed subtle histomorphological patterns associated with recurrence risk that are not captured by conventional clinicopathological risk scores. This multicohort study demonstrates that deep learning applied to routine prostate histopathology can deliver reproducible and clinically generalizable biomarkers that augment postoperative risk stratification, with potential to support personalized management of prostate cancer in real-world clinical settings.

Figures & tables

CohortLocationYears of surgery# Patients# SlidesScanner
RUMCNijmegen, The Netherlands1992–201260817233DHistech P1000
PLCOUnited States1993–20017231971Leica
IMPBelo Horizonte, Brazil20164212091Motic Easy Scan Infinity 60N
UHCCologne, Germany2015–20203301028Hamamatsu NanoZoomer S360
TCGA-PRADMultiple locations2000–2013403449Multiple scanners
Table 1: Overview of datasets . Summary of the main characteristics of each cohort used in this study, including collection period, number of patients, number of associated digital pathology slides, and scanner system used for digitization. TCGA-PRAD comprises cases from multiple contributing institutions, predominantly from the United States.
Figure 1: Representative whole-slide histopathology from external test cohorts. Low-resolution thumbnails of H&E-stained prostatectomies from the four external test cohorts used in this study. The examples illustrate cohort-specific differences in tissue appearance and color profiles arising from variations in staining protocols, slide preparation, and slide acquisition. All slides were processed using the same preprocessing pipeline prior to model inference.
EncoderArchitectureParametersPretraining corpusPretraining method
Prost40MViT-Small22M40M prostatectomy patches ∼\sim 2k WSIsDINO
UNIViT-Large307M100M patches from >> 100k WSIsDINOv2
Virchow2ViT-Huge632M1.9B patches from 3.1M WSIsDINOv2
H-optimus-0ViT-giant1.1B273M patches from 500k WSIsDINOv2
Table 2: Overview of vision encoders . Summary of the four vision encoders evaluated in this study, including model architecture, number of parameters, pretraining corpus size and source, and pretraining method.
Figure 2: Integration of multiple histopathology slides into a unified patient-level prediction framework. (A) For patients with multiple slides, segmented tissue regions are cropped using the corresponding tissue masks and stitched into a single larger packed slide. This packing process minimizes empty background space while preserving the relative arrangement of tissue, enabling patient-level analysis in a unified coordinate frame. (B) The packed slide is tiled into non-overlapping square regions of 20482048 pixels at 0.500.50 mpp. Regions with less than 11 % tissue coverage, estimated from the segmentation mask, are excluded. (C) Each retained region is then unrolled into non-overlapping 256256 pixel tiles, which are processed by a frozen tile encoder to generate tile-level embeddings. For each region, the sequence of tile embeddings is aggregated into a region-level embedding by a first Transformer. The sequence of region-level embeddings for the entire case is then aggregated by a second Transformer into a single patient-level representation, which is projected onto the target classes via a fully connected (FC) layer. The two Transformers and the FC layer are jointly trained for biochemical recurrence risk prediction. Together, these steps enable efficient processing of large whole-slide images, integrating spatially distributed information from multiple slides into a unified patient-level prediction.
Table 3: Cross-validation performance of deep learning models. Five-fold cross-validation c-index (mean ± std) of models trained on different data splits and evaluated on held-out test cohorts. Best values for each cohort are shown in bold.
Figure 3: Pairwise statistical comparisons of encoder performance across test cohorts. Heatmaps depict differences in concordance index between models trained on the RUMC-only splits and evaluated on four independent test cohorts ( RUMC , PLCO , IMP , UHC ). Each cell indicates the difference in performance between the encoder in the row and the encoder in the column. Positive values (green) reflect higher performance for the encoder in the row relative to the encoder in the column, while negative values (red) indicate the opposite. Significant differences are highlighted in bold.
Figure 4: Effect of training data enrichment on model performance across test cohorts. The heatmap depicts the change in concordance index ( Δ\Delta ) for each model when trained on the combined RUMC+TCGA dataset compared to training on RUMC-only . Rows correspond to encoders and columns to test cohorts. Positive values (green) indicate improved performance with training data enrichment, while negative values (red) indicate decreased performance. Significant differences are highlighted in bold.
Figure 5: Pairwise statistical comparisons of encoder performance across test cohorts. Heatmaps depict differences in concordance index between models trained on the RUMC+TCGA splits and evaluated on four independent test cohorts ( RUMC , PLCO , IMP , UHC ). Each cell indicates the difference in performance between the encoder in the row and the encoder in the column. Positive values (green) reflect higher performance for the encoder in the row relative to the encoder in the column, while negative values (red) indicate the opposite. Significant differences are highlighted in bold.
CohortEncoderEnsembleJointHR (95% CI)p-value
RUMCProst40M0.7600.7881.80 (1.27-2.56)0.001
UNI0.7090.7721.61 (1.12-2.29)0.009
Virchow20.7670.7741.61 (1.14-2.26)0.006
H-optimus-00.6710.7441.44 (1.01-2.06)0.043
PLCOProst40M0.7110.7791.32 (1.15-1.52)<< 0.001
UNI0.6740.7671.29 (1.12-1.49)<< 0.001
Table 4: Ensemble and joint modeling performance of models trained on RUMC+TCGA . Summary of the prognostic performance of deep learning models evaluated on the four held-out test cohorts. For each cohort, the ensemble c-index reflects the standalone performance of the deep learning model and is computed by averaging patient-level risk scores from the five RUMC+TCGA cross-validation folds. The joint c-index is obtained by fitting a multivariable CoxPH model using both the ensemble risk score and the CAPRA-S score as covariates. For joint models, the hazard ratio (HR) with 9595 % confidence interval (CI) and the Wald test pp -value indicate the independent contribution of the ensemble risk score after adjusting for CAPRA-S. Bold values indicate the highest c-index observed for each cohort.
Figure 6: Factorized attention maps reveal regions the model considers informative in prostatectomy slides. Factorized attention heatmaps derived from the Virchow2 -based model and overlaid on H&E-stained slides. Each panel shows the whole-slide image with regions of high model attention, alongside a magnified view of the most salient region. (A) True positive case: a patient who experienced biochemical recurrence within a short time and was assigned a high risk score by the model. (B) True negative case: a patient who remained recurrence-free with long follow-up and was assigned a low risk score by the model.
Figure 7: t-SNE plot of model-refined tile representations reveals recurrence-associated morphology. Across the RUMC test cohort, we selected tiles with the strongest positive or negative contributions to the recurrence risk estimated by the Virchow2 -based model. Their model-refined embeddings are projected into two dimensions using t-SNE. Color encodes each tile’s signed contribution to the final prediction, with red indicating tiles associated with higher recurrence risk and blue indicating tiles associated with lower risk, highlighting gradients of prognostic relevance within clusters. Grey outlines denote the representative clusters subsequently reviewed by an expert pathologist.
Figure 8: ISUP grade distribution across cohorts. Stacked bar plots show the proportion of patients in each cohort used in this study, grouped by ISUP grade. Each bar represents a cohort, with colors indicating ISUP grades 1 (green) to 5 (red). Differences in grade distributions reflect variation in recruitment periods, inclusion criteria, and treatment guidelines.
Figure 9: Time-to-event distributions across cohorts. Stacked histograms show times from radical prostatectomy to biochemical recurrence or last follow-up in each cohort used in this study. The horizontal axis represents time in years, and the vertical axis shows the proportion of patients within each cohort, subdivided into BCR (orange) and censored (green) cases. Consistent bin edges are used to facilitate comparison across cohorts. Cohort-specific differences in follow-up times reflect variations in recruitment periods, surveillance protocols, and clinical practices.
VariableValuePoints VariableValuePoints
PSA0–60 pGS2–60
 6.01–101  3+41
 10.01–202  4+32
 >> 203  8–103
SMNegative0 ECENo0
 Positive2  Yes1
Table 5: Scoring system used for calculating the CAPRA-S score following radical prostatectomy. Each variable contributes a fixed number of points: preoperative prostate-specific antigen (PSA) level in ng/mL, prostatectomy Gleason score (pGS), status of the surgical margin (SM), seminal vesicle invasion (SVI), extracapsular extension (ECE), and lymph node involvement (LNI). Points are summed to yield the CAPRA-S score.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

CohortModel ComparisonΔ\Delta95% CIqq
RUMCProst40M - UNI-0.017[-0.107, 0.080]0.852
 Prost40M - Virchow2-0.031[-0.128, 0.075]0.852
 Prost40M - H-optimus-0-0.003[-0.103, 0.079]0.977
 UNI - Virchow2-0.014[-0.070, 0.042]0.852
 UNI - H-optimus-0+0.014[-0.069, 0.086]0.852
 Virchow2 - H-optimus-0+0.029[-0.051, 0.102]0.852
Appendix
Table 6: Pairwise comparison of encoders trained on RUMC-only data. The table reports the difference in concordance index ( Δ\Delta ) between pairs of encoders when evaluated on each test cohort. For each comparison, Δ\Delta values are shown together with 95% bootstrap confidence intervals (CI) and false discovery rate (FDR) adjusted qq values. Positive Δ\Delta values indicate higher performance of the encoder listed first in the comparison, whereas negative values indicate higher performance of the encoder listed second. Values that remain significant after correction ( q<0.05q<0.05 ) are shown in bold.
CohortModel ComparisonΔ\Delta95% CIqq
RUMCProst40M - UNI+0.051[-0.021, 0.141]0.283
 Prost40M - Virchow2-0.007[-0.124, 0.115]0.882
 Prost40M - H-optimus-0+0.088[0.005, 0.183]0.151
 UNI - Virchow2-0.058[-0.153, 0.024]0.283
 UNI - H-optimus-0+0.037[-0.056, 0.137]0.548
 Virchow2 - H-optimus-0+0.096[-0.000, 0.201]0.151
Appendix
Table 7: Pairwise comparison of encoders trained on RUMC+TCGA data. The table reports the difference in concordance index ( Δ\Delta ) between pairs of encoders when evaluated on each test cohort. For each comparison, Δ\Delta values are shown together with 95% bootstrap confidence intervals (CI) and false discovery rate (FDR) adjusted qq values. Positive Δ\Delta values indicate higher performance of the encoder listed first in the comparison, whereas negative values indicate higher performance of the encoder listed second. Values that remain significant after correction ( q<0.05q<0.05 ) are shown in bold.
CohortEncoderΔ\Delta95% CIppqq
RUMCProst40M+0.074[0.023, 0.136]0.0020.008
 UNI+0.006[-0.060, 0.067]0.8660.990
 Virchow2+0.050[-0.019, 0.139]0.1900.380
 H-optimus-0-0.017[-0.112, 0.075]0.7080.871
PLCOProst40M+0.107[0.052, 0.161]<< 0.001<< 0.001
 UNI-0.020[-0.064, 0.021]0.3530.565
Appendix
Table 8: Effect of training data enrichment on model performance. The table shows the change in concordance index ( Δ\Delta ) between models trained on the combined RUMC+TCGA splits and those trained on RUMC-only splits. For each test cohort and encoder, Δ\Delta values are reported alongside 95% bootstrap CI, the exact two-sided pp values, and FDR-adjusted qq values. Positive Δ\Delta values indicate higher performance after training data enrichment, whereas negative values indicate lower performance. Values that remain significant after FDR correction ( q<0.05q<0.05 ) are highlighted in bold.
Test setEncoderEnsembleJointHR (95% CI)p-value
RUMCProst40M0.6860.7541.50 (1.05-2.15)0.028
UNI0.7030.7391.31 (0.93-1.85)0.122
Virchow20.7170.7591.50 (1.04-2.17)0.030
H-optimus-00.6880.7331.17 (0.86-1.60)0.327
PLCOProst40M0.6040.7661.07 (0.98-1.17)0.121
UNI0.6940.7691.26 (1.11-1.43)<< 0.001
Appendix
Table 9: Ensemble and joint modeling performance of models trained on RUMC-only . Summary of the prognostic performance of deep learning models evaluated on the four held-out test cohorts. For each cohort, the ensemble c-index reflects the standalone performance of the deep learning model and is computed by averaging patient-level risk scores from the five RUMC-only cross-validation folds. The joint c-index is obtained by fitting a multivariable CoxPH model using both the ensemble risk score and the CAPRA-S score as covariates. For joint models, the hazard ratio (HR) with 9595 % confidence interval (CI) and the Wald test pp -value indicate the independent contribution of the ensemble risk score after adjusting for CAPRA-S. Bold values indicate the highest c-index observed for each cohort.
Figure 10: ISUP-stratified predictive performance on the UHC cohort before and after training set enrichment. Ensemble c-index of models trained on RUMC-only versus RUMC+TCGA , stratified by ISUP grade within the UHC cohort. Results exclude grade 1 (all censored, preventing c-index calculation) and grade 4 (small sample size). Models trained on RUMC-only show performance skewed toward ISUP 2, reflecting the grade distribution in the training set. Adding TCGA-PRAD cases reduces this bias, improving performance in higher-grade cases (ISUP 3 and 5) and yielding a flatter profile across grades, consistent with better calibration of recurrence risk across grades.
Figure 11: Representative tiles from risk-increasing cluster 1. This cluster is dominated by adverse histology, including extensive cribriform structures and fused glands forming anastomosing cords, aligning with the model’s interpretation of this cluster as increasing risk of recurrence.
Figure 12: Representative tiles from risk-increasing cluster 2. Tiles predominantly contain high-grade carcinoma, including single infiltrating cells characteristic of Gleason pattern 5 and fused pattern 4 glands forming anastomosing cords. These morphologies are strongly prognostically adverse and align with large positive contributions to risk of recurrence.
Figure 13: Representative tiles from risk-increasing cluster 3. Tiles primarily depict fibromuscular stroma rather than carcinoma, suggesting that the model may partially attribute recurrence risk based on patient- or slide-specific non-tumor features.
Figure 14: Representative tiles from risk-lowering cluster 1. Most tiles display predominantly benign tissue with only occasional proximity to carcinoma.
Figure 15: Representative tiles from risk-lowering cluster 2. While some tiles show cribriform morphology which should confer higher risk, the overall Gleason pattern 4 complexity seen in other tiles is more consistent with a favorable histological appearance.
More in this topic

Clinical Risk Prediction

Clinical risk prediction estimates patients’ future health outcomes from medical records, physiological signals, imaging, and other health data.

See all 214 papers

Explore similar work

CardsList

  1. GD-MIL: Grade-Disentangled Multiple Instance Learning for Multimodal Biochemical Recurrence Prediction in Prostate Cancer

    Jun 8, 2026Dasari Naga RajuAdversarial Representation LearningCancer Survival Prediction

  2. Efficient AI-Driven Multi-Section Whole Slide Image Analysis for Biochemical Recurrence Prediction in Prostate Cancer

    Mar 17, 2026Yesung Cho, Dongmyung Shin, Sujeong Hong +5Clinical Outcome PredictionWhole-Slide Image Analysis

    OmixAI Co. Ltd., Republic of Korea, 02636. · Oncocross Co. Ltd., Republic of Korea, 05836. · Department of Urology, School of Medicine, Pusan National University, Republic of Korea, 50612 +1

  3. CHIMERA Challenge: Biochemical Recurrence Prediction in Prostate Cancer Patients using multimodal datasets

    Aug 21, 2026Robert N. Spaans, Catherine Chia, Tongjie Wang +7Clinical Outcome PredictionMedical Image Benchmarks

    Radboud University Medical Center, Department of Pathology, Research Institute for Medical Innovation, Nijmegen, The Netherlands · Canisius Wilhelmina Hospital, Department of Urology, Nijmegen, The Netherlands · Oncode Institute, Utrecht, The Netherlands +6