Enhancing prognostic precision in bladder cancer: AI-driven tumor microenvironment analysis from H&E images.

Abstract

3017 Background: Bladder cancer (BC) represents a significant healthcare burden. Despite advancements in diagnostics and treatment, the survival rate remains low, underscoring the need for improved prognostic tools. The current UICC staging system often lacks precision in patient stratification. Moreover, there is a paucity of scalable methods that explore and quantify tumor microenvironment (TME) features and their influence on patient outcome. Here, we present an artificial intelligence (AI) framework that operates on routine hematoxylin & eosin-stained (H&E) slides to enable systematic TME characterization and improve prognostic accuracy. Methods: In a bicentric cohort of over 700 resected BC patients, we developed and validated a deep learning approach for TME analysis. The model was trained using multiplex immunofluorescence-validated annotations but operates solely on H&E-stained images, maximizing clinical applicability. Key features included tissue compartment segmentation, cell classification, and spatially resolved cell patterns. We evaluated the model’s performance and integrated TME features with clinicopathological variables to improve prognostic stratification beyond UICC staging. Results: The model demonstrated robust tissue compartment segmentation (F1-score = 0.91) and accurately identified key immune cell populations in tissue regions. When integrating the spatially resolved cellular features with clinicopathological variables, we observed significant improvements in prognostic capabilities for overall survival. The integrated approach demonstrated a 22% relative improvement over the conventional UICC staging system alone (C-index increased from 0.59 to 0.61, p < 0.01), measured against the random baseline C-index of 0.5. Our integrated model displayed a hazard ratio of 1.859 (95% CI: 1.530-2.259, p = 4.390e-10), markedly stronger than traditional risk stratification which showed a hazard ratio of 1.477 for high versus low risk groups (95% CI: 1.219-1.791, p = 7.117e-05). These findings demonstrate that AI-driven analysis of the tumor microenvironment provides valuable prognostic information beyond current clinical staging methods, suggesting promising opportunities for enhancing patient risk stratification. Conclusions: We show that a combination of AI with UICC staging shows improved patient stratification compared to stratifying by UICC alone. This study demonstrates the feasibility of automated TME characterization from routine H&E slides in BC and suggests that incorporating TME features into prognostic models enhances accuracy and could support personalized patient management in individualized oncology. While further validation in larger, multicentric-datasets is required, our approach shows potential for facilitating systematic biomarker development and improving clinical decision-making in BC care.