Learning Patient-Level Visual Representations from Pathology Foundation Models for Survival Prediction from Whole-Slide Images

Whole-slide histopathology images (WSIs) capture rich, multi-scale morphological information reflecting tumor heterogeneity and disease progression. However, learning robust patient-level representations from gigapixel WSIs remains a fundamental computer vision challenge, requiring aggregation of thousands of local image features into a single prognostic representation. While pathology foundation models (FMs) have demonstrated strong transfer learning across downstream tasks such as tissue classification and segmentation, their ability to learn transferable representations for survival prediction remains underexplored. In this work, we develop a pathology foundation model-derived visual biomarker for predicting biochemical recurrence (BCR) following radical prostatectomy using routine H&E whole-slide images. Our framework first applies patch-level cancer classification to identify tumor-containing regions and construct a whole-slide cancer segmentation map. The resulting map is post-processed using morphological closing to generate a refined tumor mask, from which tumor-derived patches are selected for prognostic representation learning. The framework then leverages the pretrained UNI pathology foundation model as a frozen visual encoder to extract high-dimensional embeddings from 224 × 224 px image patches at 20× magnification within these cancerous regions. Prior to feature extraction, patches are stain-normalized using the Macenko method to mitigate color variability across slides. Survival modeling uses 1.5M tumor-derived patches from 429 radical prostatectomy patients in the training cohort and 500K patches from 46 patients in the independent evaluation cohort. Patch-level embeddings are aggregated into patient-level visual representations using mean pooling followed by an attention-based aggregation module, and a Cox proportional hazards model is trained to predict time-to-recurrence. The model is developed on the 429-patient training cohort and evaluated without retraining on the independent 46-patient cohort, which has matched whole-slide pathology, clinical variables, and Decipher genomic scores. The resulting FM-derived visual biomarker is compared against established prognostic tools, including UCSF-CAPRA, the MSKCC preoperative nomogram, and the Decipher genomic assay. On the independent evaluation cohort, the proposed visual biomarker achieved a concordance index (C-index) of 0.828, outperforming UCSF-CAPRA (0.797) and the MSKCC preoperative nomogram (0.812), while approaching the Decipher genomic assay (0.861). These findings demonstrate that pretrained pathology foundation models learn transferable morphological representations that encode prognostic tissue characteristics directly from routine H&E pathology images. Our results highlight the potential of pathology foundation models as scalable visual representation learners for survival analysis from gigapixel pathology images, demonstrating that foundation model-derived features can support clinically relevant prognostic modeling without task-specific feature engineering or genomic profiling. More broadly, this work demonstrates the potential of pathology foundation models to translate gigapixel tissue morphology into clinically meaningful patient-level prognostic signals.