Predicting the risk of prostate cancer metastasis at diagnosis in the VA Health Care System from H&E slides of the prostate needle biopsies.

Abstract

128 Background: Men diagnosed with high-grade prostate cancer harbor increased rates of occult metastatic disease. We propose that features of the cancer in H&E stained tissue sections from prostate needle biopsies can be used to predict metastatic disease at diagnosis. Methods: We analyzed 168 high-risk cases from the Greater Los Angeles VA (GLA VA), equally divided into non-metastatic with at least five years of follow-up after diagnosis (M0) or synchronous metastatic cases (M1), and 38 cases with metachronous metastases (M0-P). From each case we analyzed all available images of H&E biopsies. Biases in the data were mitigated through supervised stratification and novel color augmentation strategies. Features were extracted from high-grade cancer tiles using 9 Foundation models. These embeddings served as input to a DTFT multiple instance learning (MIL) aggregator model that was trained de-novo to provide a metastatic risk (MR) score. The MR scores were combined with race, age and PSA clinical variables as a covariate in a linear regression or in a super learner to examine the contribution of each variable for the M1 prediction. Performance was assessed using 5-fold nested cross-validation. Results: The AUCs for prediction of metastasis at diagnosis ranged from 0.82 – 0.85. Addition of PSA and race further increased the AUC. A principle component analysis revealed explainable differences in the embeddings of M0 and M1 cases. Tiles used for M0/M1 discrimination by the aggregator model were further examined by analyzing the model’s attention on cancer and stroma regions. While the majority of the attention was in the cancer region in M0 cases, the attention in M1 cases was primarily in the stroma. Conclusions: We demonstrate that digital pathology slides from diagnostic prostate needle biopsies contain information for prostate cancer staging at diagnosis in a small, real-world cohort at VA. Using state-of-the-art Foundation models as feature extractors improves the accuracy of staging, in particular when the number of cases for training is small. Explainability of the feature embedding and classification tasks allows to examine what the models are learning and increases the confidence in the AI-generated results. Accuracy of metastasis prediction in VA cohort. Foundation model architecture MIL model MIL Model *AUC *Balanced Accuracy CONCH 80.86 ± 7.91 76.48 ± 5.77 HIBOU-B 85.82 ± 6.10 81.07 ± 6.20 MoCo v3 ResNetS0 80.36 ± 7.44 78.76 ± 5.81 Phikon 85.45 ± 5.70 80.59 ± 4.85 Phikon-v2 85.26 ± 6.25 80.83 ± 3.59 GigaPath (Prov-GigaPath) 85.17 ± 6.31 80.30 ± 4.21 UNI 86.03 ± 5.44 81.48 ± 4.33 UNl2-H 85.50 ± 7.45 82.02 ± 6.30 Virchow (1280d) 84.99 ± 6.51 83.12 ± 5.91 Virchow2 (1280d) 84.11 ± 6.87 78.64 ± 7.26 *Classification performance AUC and balanced accuracy (mean and std, %) using case-level bags on final tiles with a fixed feature setting and dino cls keep=4. Best result per column in bold.