Lessons from deploying AI-assisted assessment of donor kidney

Introduction / background: Accurate assessment of global glomerulosclerosis (GGS) in donor kidneys is critical to transplant decision-making, yet manual estimation is subject to significant inter-pathologist variability. AI-assisted quantification is promising, but real-world deployment introduces challenges beyond model accuracy-including workflow integration, system robustness, and pathologist adoption.

Methods / design: From September 2024 through April 2026, we deployed an AI-assisted GGS workflow across two institutions: 1,123 cases, 1,331 donor organs, 1,758 slides, and 2,316 reviews by 14 pathologists. Most slides were reviewed twice-by the on-call pathologist and a second reviewer the following day. Key UI features included a guided tour ensuring complete slide coverage, one-click rejection of AI-predicted glomeruli, and one-click addition of missed glomeruli. Rejected and added glomeruli and review times were captured prospectively.

Results: Workflow design proved as consequential as model performance. Streamlined correction tools drove adoption across all 14 pathologists even when the AI failed. AI inference averaged ~30 seconds per slide-compatible with intraoperative use-though achieving this required significant engineering effort. Prospective data enabled structured quality assessment, revealing disagreement patterns between pathologists and between pathologists and the algorithm. Silent errors from undocumented changes to slide acquisition protocols emerged as a critical and underappreciated risk.

Conclusion / discussion: Successful AI deployment requires complete workflows-not just algorithms-that reduce cognitive load and remain robust to failure. UI features guiding coverage and simplifying correction are essential to adoption. Our 19-month, two-institution experience offers a real-world foundation for understanding how AI tools perform and fail in time-sensitive clinical settings.

Learning Objectives:

1. Describe how guided slide navigation and one-click correction tools drive pathologist adoption of AI in clinical workflows.
2. Identify operational risks—silent upstream errors and inference latency—that affect AI reliability in real-world deployment.
3. Apply lessons from this deployment to design AI workflows that reduce cognitive load and remain robust to model failure.