A systematic analysis of the impact of data variation on AI-based histopathological grading of prostate cancer.

The histopathological evaluation of biopsies by human experts is a gold standard in clinical disease diagnosis. While recent artificial intelligence-based (AI) approaches have reached human expert-level performance, they often display shortcomings caused by variations in sample preparation, limiting clinical applicability. This study investigates the impact of data variation on AI-based histopathological grading and explores algorithmic approaches that confer prediction robustness. To evaluate the impact of data variation in histopathology, we collected a multicentric, retrospective, observational prostate cancer (PCa) trial consisting of six cohorts in 3 countries with 25,591 patients, 83,864 images. This includes a high-variance dataset of 8,157 patients and 28,236 images with variations in section thickness, staining protocol, and scanner. This unique training dataset enabled the development of an AI-based PCa grading framework by training on patient outcome, not subjective grading. It was made robust through several algorithmic adaptations, including domain adversarial training and credibility-guided color adaptation. We named the final grading framework PCAI. We compare PCAI to a BASE model and human experts on three external test cohorts, comprising 2,255 patients and 9,437 images. Variations in sample processing, particularly section thickness and staining time, significantly reduced the performance of AI-based PCa grading by up to 8.6 percentage points in the event-ordered concordance index (EOC-Index) thus highlighting serious risks for AI-based histopathological grading. Algorithmic improvements for model robustness, credibility, and training on high-variance data as well as outcome-based severity prediction give rise to robust models with grading performance surpassing experienced pathologists. We demonstrate how our algorithmic enhancements for greater robustness lead to significantly better performance, surpassing expert grading on EOC-Index and 5-year AUROC by up to 21.2 percentage points.