New AI Shrinks Gigantic Cancer Scans Without Losing the Details

Digital pathology works by scanning a slice of tissue into a whole-slide image — a single file that can run several gigabytes, roughly like stitching thousands of phone photos into one picture. Training AI to spot cancer in these files is slow, expensive, and power-hungry. So researchers try to shrink each slide into a much smaller stand-in that teaches the AI nearly as well. The problem: older shrinking methods guessed at what mattered and often threw away useful detail.

A team led by Duong M. Nguyen reframed the task as a matching problem. Instead of guessing, NICER compares the small stand-in to the original slide and adjusts until the two look statistically alike — keeping the patterns that matter for spotting disease. Crucially, each slide decides for itself how much detail to keep: messy, complex tissue gets more room, simple tissue gets less. The method learns from unlabeled slides, so hospitals don't need armies of humans tagging every image first.

The results, accepted at the NeurIPS 2026 conference, come from five histopathology datasets. NICER beat previous approaches by 7.44% in average accuracy, and a board-certified pathologist reviewed the outputs. It also used computing power more efficiently, improving the trade-off between speed and accuracy.

What it means for you: pathology labs are drowning in data and short on specialists. If slides can be compressed dramatically without losing diagnostic signal, AI tools become cheaper to train and easier to run in smaller hospitals — potentially shortening the wait for biopsy results. The honest catch: this is a research result, not a hospital product. The code is public, but real-world clinical testing, regulation, and validation across diverse patient groups are still ahead.