整合全玻片图像与临床特征以实现更优的前列腺癌风险分层

Background: Prostate cancer remains one of the most common malignancies in men, with significant heterogeneity in clinical outcomes. Early and accurate diagnosis and risk stratification is crucial for effective treatment and improved outcomes. Although Gleason score is an established risk stratification tool, it does not fully explain outcome variability. We present a multimodal deep-learning framework that integrates attention-based whole-slide image (WSI) features with a clinical variable (Age) to predict survival in patients with prostate adenocarcinoma. Design: Clinical data and WSI for 500 patients with prostate adenocarcinoma were obtained from TCGA-PRAD. Clinical data were parsed for recurrence or progression (RoP) events and age. RoP was used as the outcome label, and age was used as a patient feature. Diagnostic WSI tumor regions were annotated and tiled at 224×224 pixels using QuPath (Bankhead Sci Rep 2017). Tile embeddings were extracted with UNI2-h (Mahmood Nat Med 2025). An end-to-end machine learning pipeline was developed to aggregate tile embeddings into slide-level embeddings using Clustering-Constrained Attention Multiple Instance Learning (CLAM) and predict RoP (Lu Nat Biomed 2021). A second pipeline used a Gleason score-trained binary cross entropy (BCE) model to predict RoP in the same patients. Three models were evaluated: WSI-only CLAM model, WSI+age CLAM model, and Gleason-only BCE model. All models were trained and evaluated using 5-fold stratified cross-validation. Each fold was repeated 3 times with independent random initialization. AUROC was averaged across all folds and runs per model. Standard deviation was also computed. Model performance was compared using Friedman's test, with post hoc pairwise comparisons by Wilcoxon signed-rank test. Results: The Gleason BCE model underperformed (AUROC 0.67 ± 0.07). Both WSI-based models outperformed the Gleason BCE model, achieving AUROC 0.76 ± 0.02 (without age), and AUROC 0.76 ± 0.01 (with age). AUROCs were significantly different by Friedman's test (p = 0.0224) and showed a trend toward significance in post-hoc comparison between WSI and Gleason-based approaches (p = 0.06). Conclusion: AI models using WSI-derived features, with or without basic clinical context, outperformed traditional Gleason score for recurrence risk prediction in prostate cancer. These findings support the integration of digital pathology and AI into routine prognostic assessment for prostate cancer. Future work will focus on including additional patients, clinical variables, and multi-class prognostic labels to expand prognostic prediction beyond RoP.