Each year, approximately 1.6 million men worldwide are diagnosed with prostate cancer, and the most common surgical treatment is radical prostatectomy -- complete removal of the prostate. While surgery is often effective, roughly 20-40% of patients experience biochemical cancer recurrence (BCR) within 5 years, detected by a rise in PSA (prostate-specific antigen) levels in the blood.
BCR is a strong risk factor for subsequent metastasis and death from prostate cancer. Identifying which patients will relapse after surgery is critical for clinical decision-making: high-risk patients could benefit from more frequent PSA monitoring, advanced imaging like PSMA-PET scans to detect hidden spread, or early systemic therapy.
Existing prediction tools, such as the CAPRA-S score, combine clinical variables like PSA level, Gleason grade, and surgical margin status to estimate recurrence risk. Genomic tests like the Decipher test analyze gene expression in tumor tissue. However, these tools have not achieved widespread adoption due to cost barriers and limited accuracy.
A key insight motivating this study is that routine H&E-stained tissue slides -- already produced for every prostate cancer surgery -- may contain visual information about recurrence risk that goes beyond what pathologists formally record. Unlocking this hidden information with AI could improve prediction without additional cost or testing.
The researchers developed PathCLR (Pathology Contrastive LeaRning), a two-stage machine learning pipeline that predicts 5-year BCR from tissue microarray (TMA) images combined with clinical data. TMAs are small standardized tissue cores punched from tumor samples and embedded on glass slides for scanning.
The first stage uses SimCLR, a self-supervised contrastive learning algorithm that learns image features without any outcome labels. SimCLR is trained by creating two different augmented views of the same tissue patch (through cropping, color distortion, and blurring) and learning to recognize them as the same image. This teaches the model to extract meaningful features from tissue appearance.
The second stage combines the learned tissue image features with seven clinicopathological variables -- including PSA level, Gleason grade (primary and secondary), surgical margin status, extracapsular extension, lymph node involvement, and seminal vesicle invasion -- in a supervised neural network classifier that predicts BCR as a binary yes/no outcome.
A key design feature is that PathCLR makes a personalized prediction for each individual patient rather than grouping patients by risk category. For patients with multiple tissue cores, the model applies a disjunction rule: if any single core is predicted as a recurrence case, the patient is classified as high-risk, capturing heterogeneity across biopsy sites.
Two independent prostate cancer TMA datasets were used: the Cooperative Prostate Cancer Tissue Resource (CPCTR) with 374 patients (189 with BCR), and the Johns Hopkins University (JHU) dataset with 646 patients (451 with BCR). Both datasets used a matched case-control design, pairing BCR cases with non-recurrence controls matched on age, race, and Gleason grade.
Importantly, both datasets were specifically designed to be challenging: the majority of patients had Gleason sum score 7 (either 3+4 or 4+3), the most common intermediate-risk grade group where recurrence prediction is hardest. This contrasts with many prior studies that included easier-to-distinguish low and high grade cases.
Before feeding images into the model, the pipeline uses HoVer-Net, a state-of-the-art nuclear segmentation model, to automatically detect and outline epithelial cells (the cells that become cancerous) in each TMA core. From each core, 200 patches of size 128x128 pixels are extracted, centered on detected epithelial cell positions, to ensure the model focuses on the most diagnostically relevant tissue regions.
The fine-tuned HoVer-Net model achieved a panoptic quality score of 0.675, exceeding the top submission from the MoNuSAC2020 nuclear segmentation challenge (0.611), demonstrating reliable epithelial cell detection as the foundation for feature extraction.
Using 10-fold cross-validation, PathCLR achieved an F1 score of 0.61 on CPCTR and 0.85 on JHU when combining tissue image features with clinical variables. The F1 score balances precision and recall, making it appropriate for datasets with unequal BCR/non-BCR ratios.
Critically, combining tissue image features with clinical data (Setting 3) improved accuracy by at least 5% over clinical features alone (Setting 1), and this improvement was statistically significant on both datasets (P = .005 for JHU; P = .04 for CPCTR). This confirms that the tissue slides genuinely contain predictive information beyond what pathologists formally report.
Using clinical features alone (Setting 1) performed barely better than simply predicting BCR for all patients -- demonstrating that in matched datasets where cases and controls share similar Gleason grades and other variables, standard clinical scores like CAPRA-S essentially fail. The CAPRA-S score alone achieved accuracy of 53.22% on CPCTR, close to random chance.
PathCLR also outperformed a fully supervised CNN trained directly on whole tissue cores (by about 2% on both datasets) and outperformed a prior benchmark that used hand-crafted gland morphology features (histotyping). The semi-supervised approach, by first learning image representations without labels, proved more effective than direct supervised learning.
The study's matched case-control design intentionally created a hard problem: patients with and without BCR share similar Gleason grades and other standard clinical variables. This explains why clinical features alone barely exceeded random prediction -- there is a ceiling on what reported variables can capture.
The improvement from adding tissue image features suggests that H&E slides contain microscopic visual signals about future recurrence that are not captured by current grading systems. These signals may reflect subtle differences in nuclear morphology, cellular density, stromal organization, or other features that pathologists do not formally quantify.
PathCLR's individualized yes/no prediction is clinically distinct from hazard ratio approaches that only rank patients relative to each other. Hazard ratios answer the question of whether patient A will recur before patient B, but cannot directly tell a clinician whether a specific patient needs additional treatment. PathCLR addresses this directly.
The authors note that cross-dataset generalization was not attempted due to major differences between CPCTR and JHU in clinical feature ranges, image resolution (40x vs. 20x), BCR rates, and staging distributions. Each model was trained and evaluated within its own dataset to ensure fair assessment.
A significant advantage of PathCLR is that it relies entirely on routine H&E-stained tissue slides that are already produced for every prostate cancer surgery. No additional staining, genomic testing, or expert annotation is required. The cost per slide scan is approximately $4, and scanning takes about one minute per slide.
The entire BCR prediction for a patient -- including epithelial cell detection and recurrence probability estimation across up to four tissue cores -- completes in under 7.3 seconds of computational time. This makes it practical for integration into standard pathology workflows without adding meaningful delay.
For patients identified as high recurrence risk, the early signal could trigger PSMA-targeted PET imaging to detect occult metastases or local spread, as demonstrated in recent clinical trials like CONDOR where 64% of high-risk patients changed their management strategy based on PET results.
PathCLR requires no manual tumor annotation, overcoming a major practical barrier that limited earlier computational pathology approaches. The self-supervised learning component extracts tissue features automatically, making the system feasible for deployment in any clinical center with access to digital pathology scanning.
PathCLR represents an important step toward personalized prostate cancer surveillance after surgery. By combining the hidden prognostic information in tissue images with clinical variables, it achieves meaningfully better BCR prediction than either source alone on two independent datasets.
The study establishes that the semi-supervised approach -- first learning visual representations without outcome labels, then training a classifier -- outperforms both direct supervised learning and clinical-only models in this setting, validating contrastive self-supervised learning as a valuable strategy for medical imaging with limited labeled data.
Future work should address the gap in cross-dataset generalizability, expand validation to larger and more diverse patient cohorts, and explore whether additional tissue stains or imaging modalities could further improve prediction accuracy.