Prognosis Prediction in Bladder Cancer Pathological Images Based on Nuclear Structure Encoding

Background: Bladder cancer is identified as a common malignancy in the male urinary system. Muscle-invasive bladder cancer (MIBC) is known for rapid disease progression and poor prognosis. Traditional pathological evaluation using tissue slides is challenged by subjectivity and inter-observer variability, whereby precise prognostic tools are required for personalized treatment plans. Methods: A deep learning framework integrating graph neural networks and knowledge distillation is designed to predict bladder cancer prognosis using hematoxylin and eosin (H\&E) stained whole slide images (WSI). An MIBC cohort from TCGA (N=387) is utilized to create datasets for nuclear classification and prognosis assessment. A multi-scale feature fusion and knowledge distillation module is designed, where the pre-trained Segment Anything Model (SAM) is employed for nuclear feature extraction. These features are transmitted to Vision Transformer (ViT) through knowledge distillation. A graph neural network framework based on an attention mechanism is constructed, where nuclear morphological features are mapped to graph nodes and spatial relationships between nuclei are explored. Results: The effectiveness of using WSI images to support MIBC treatment decision-making is significantly improved by the proposed method, as demonstrated by experimental results. Conclusions: Accurate MIBC prognosis classification is achieved by the proposed method through effectively capturing nuclear morphological characteristics and their spatial distribution patterns, demonstrating its capability in precise prognostic stratification.