Development and validation of artificial intelligence-based model for bladder cancer immunophenotyping using whole...

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate immunophenotype classification in muscle-invasive bladder cancer (MIBC) is crucial for predicting immunotherapy response.
• Current methods for MIBC immunophenotype assessment lack standardization and scalability.
• Developing reproducible and scalable diagnostic tools is essential for precision immuno-oncology.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI)-based system for reproducible immunophenotype classification in MIBC using routine hematoxylin and eosin-stained whole-slide images.
• To assess the generalizability and diagnostic performance of the AI system across external validation cohorts.
• To evaluate the system's utility in human-AI collaboration and its correlation with immunotherapy response.

Main Methods:

• Development of an AI system integrating Hover-Net-based nuclear classification and cell structure graph networks for spatial cellular interaction modeling.
• Retrospective analysis of consecutive MIBC patients from multicenter cohorts (two Chinese hospitals, The Cancer Genome Atlas) and an independent cohort treated with immune checkpoint inhibitors.
• Validation of the AI model using macro-area under the curve (AUC) and macro-accuracy metrics, and assessment in a human-AI collaboration study.

Main Results:

• The AI system demonstrated robust generalizability with high performance across external validation cohorts (macro-AUC 0.922-0.956, macro-accuracy 0.922-0.950).
• In human-AI collaboration, the system outperformed junior and senior pathologists, enhancing diagnostic accuracy and reducing review time for junior pathologists.
• Predicted 'Inflamed' tumors showed increased CD8+ T-cell infiltration, elevated checkpoint gene expression, and a stronger correlation with immunotherapy response.

Conclusions:

• The AI-based MIBC Immunophenotype Diagnostic System offers a reproducible and scalable solution for classifying immunophenotypes from routine pathology slides.
• The system shows significant potential for clinical translation, aiding in precision immuno-oncology for bladder cancer.
• AI-driven immunophenotype analysis can improve diagnostic workflows and predict patient response to immunotherapy, optimizing treatment strategies.