Repeat MRI images can make AI tumor software look better than it is
Original title: Near-duplicate contamination inflates brain MRI benchmark performance: a size-matched, multi-seed MedNeXt-B study.
How far along is this research?
- Lab cells
- Animals
- Review
- Tested in people
This was done on cells in a lab, not in people. It is a very early step.
This was a computer study about how AI software is tested, not a treatment for people.
The short version
This study checked how AI scan software is tested, not a new treatment.
What was studied. Researchers trained one AI program to read single brain MRI slices. It sorted each slice into one of four groups and outlined the tumor. They ran it two ways, with and without near-copy images shared between the training and test images.
What they found. The same test set of 995 slices was used each time. The AI scored higher when near-copy images were left on both sides. Its scores went up for naming the tumor type and for outlining the tumor, and the biggest jump was in spotting meningioma.
What this means, and what it doesn't
What it could mean: This does not change your care or your scans. It is a warning for people who build and test AI scan tools. Some reported AI scores may look better than the tool really is.
What it doesn't mean: This is not a treatment and not a cure. It was not tested in patients. No one was treated, and no one got a diagnosis from it in this study. It does not say AI reading of scans works or does not work. It is early research about testing methods, far from everyday care. The authors also say their own copy-finding tool found only 5.8% of known same-patient image pairs in another set, so some repeats may have been missed.
Source: PubMed, October 8, 2026 · Read the original
This plain-language summary was written by AI and published automatically after passing our automatic safety checks. How we write.