Mayo Clinic's New AI Reads Biopsy Slides and Drafts the Report

When you get a biopsy, a pathologist studies dozens of thin-sliced tissue slides, zooming in and out at different magnifications, then writes up what they see. It's slow, highly skilled work. A team at Mayo Clinic built WILSON, an AI that treats a patient's whole case as one single picture rather than thousands of separate image tiles. It learned by studying about 189,000 real slides, spanning 42 organs and 829 types of diagnosis, along with the written reports that went with them.

The efficiency results are striking. WILSON's accuracy score came in at 0.52 versus 0.38 for a model built specifically for case-level work. It also matched rival AI models up to 9.4 times larger while using up to 2,155 times less computing power. In plain terms: similar or better answers from a much smaller, cheaper engine. That matters because most hospitals and labs don't have the budget or the electricity for the biggest AI systems.

WILSON also writes. Asked to find the text matching a slide, it landed on the right description first try 75.6% of the time, beating an earlier system that managed 58.1%. Its generated captions read closer to real pathology reports than competing models. For patients, that hints at faster turnaround and less clerical work for the people reading the slides.

The catch: this is a research paper, not an approved medical device. An accuracy score of 0.52 means it is often wrong, so a pathologist still has to check every finding. It was trained mostly on Mayo Clinic data, so its performance at other hospitals isn't guaranteed. Think of it as a promising assistant that speeds up the paperwork — not a replacement for your doctor's judgment.