Pathological graph self-supervised learning for clear-cell renal cell carcinoma survival prediction

Whole-slide images (WSIs) are pivotal for diagnosing clear-cell renal cell carcinoma (ccRCC), but their complex tissue topology and tumor-microenvironment cues challenge conventional pipelines. We propose a Pathological Graph Self-Supervised Learning (PGSL) method for ccRCC survival prediction. PGSL performs large-scale graph self-supervised pretraining on over 1000 ccRCC WSIs using a masked link-prediction task that recovers local topology and tumor-microenvironment semantics by randomly masking nodes and predicting their connectivity. In inference, the pretrained model induces a weighted adjacency that provides topology-aware guidance for slide-level computation. This learned graph, together with initial patch features, is processed by a graph neural network to integrate global WSI context and produce patient-level risk scores. We evaluate PGSL on two private cohorts and the public TCGA-KIRC dataset, where it consistently outperforms multiple-instance learning and graph baselines across survival tasks. The gains stem from data-adaptive graph induction that replaces heuristic constructions and from effective aggregation over the induced topology. PGSL demonstrates robust generalization and delivers interpretable attention patterns aligned with prognostic pathology. These results highlight PGSL as an effective and scalable framework for pathological graph analysis and ccRCC survival stratification.