Why Training AI on Whole Biopsies Beats Individual Tissue Slices for Prostate Cancer Grading

Artificial intelligence systems designed to grade prostate cancer aggressiveness perform better when trained on complete biopsy cores rather than individual tissue sections, according to new research from Radboud University Medical Center in the Netherlands. The finding addresses a fundamental challenge in deploying AI for cancer diagnosis: ensuring that algorithms trained on tissue samples can accurately assess disease severity across an entire patient.

The study, published in Scientific Reports, analyzed 2,660 histopathology sections from 803 core biopsies taken from 106 patients. Researchers compared deep learning models trained at different scales—individual sections versus complete biopsy cores—to determine which approach produced more reliable patient-level Gleason scores, the standard measure of prostate cancer aggressiveness.

Biopsy-Level Models Outperform Section Combinations

The biopsy-level model achieved a Cohen's quadratic-weighted kappa of 0.87 when evaluating individual biopsies and maintained performance at 0.79 when predictions were aggregated to the patient level. This consistency proved superior to approaches that combined predictions from section-level models, which showed greater sensitivity to variations in the number of biopsy cores and heterogeneity between tissue samples.

Lead author N. Uysal and colleagues found that voting-based combination methods—where individual section predictions are pooled to generate patient-level assessments—struggled with the inherent variability in prostate cancer presentation. Prostate tumors are often multifocal, with different areas showing varying degrees of aggressiveness, making it difficult to extrapolate from isolated tissue fragments to overall disease severity.

Implications for Clinical AI Deployment

The research carries implications for pathology AI companies and healthcare providers implementing automated diagnostic systems. For developers of digital pathology platforms, the findings suggest that training strategies should prioritize biopsy-core-level data rather than maximizing the number of individual section images. This approach appears to better capture the biological context needed for clinically meaningful predictions.

Healthcare institutions may benefit from more reliable AI-assisted Gleason grading that accounts for intrabiopsy heterogeneity—the variation in cancer characteristics within a single biopsy sample. The study suggests that biopsy-level models naturally incorporate this variability during training, producing predictions that align more closely with how pathologists assess overall cancer aggressiveness.

Research Context and Funding

The study was conducted at Radboud University Medical Center and received ethics approval from the institution's Ethics Committee. Lead author N. Uysal was supported by a Turkish Ministry of National Education YLSY scholarship, while co-author G.L. received funding from the Dutch Research Council, European Research Council, and Oncode Institute. The research utilized a slide packing algorithm developed by Witali Aswolinskiy. Co-author G.L. disclosed an equity relationship with Aiosyn BV.

The findings underscore the importance of matching AI training methodologies to clinical workflows, particularly as pathology departments increasingly adopt computational tools for cancer diagnosis. By demonstrating that biopsy-level training produces more accurate patient-level assessments, the research provides guidance for developing AI systems that translate effectively from laboratory validation to real-world clinical use.

Sources

1. Towards patient-level Gleason grading from AI-based analysis of individual sections — Nature Scientific Reports