Open AI programs were not reliable at spotting brain tumors on MRI
Original title: Benchmarking Open-Source Vision-Language Models for Brain Metastasis Assessment on Single-Slice Contrast-Enhanced MRI
How far along is this research?
This is a preprint. Other scientists have not checked it yet, so treat it as an early signal rather than an answer.
This only tested computer programs on saved scan images, not in real care.
The short version
Researchers tested free AI programs on brain MRI scans, and none read them well enough to trust.
What was studied. Researchers tested three general AI programs and three medical AI programs that can look at pictures. Each read brain MRI scans from 60 patients with cancer that had spread to the brain. Sixty scans showed a tumor and sixty did not.
What they found. One program, MiniCPM-V-4.5, did the best overall. It spotted a tumor in 78.3% of the scans that had one, and it correctly said no tumor in 85.0% of the scans without one. The programs built for medicine did not do better than the general ones.
What this means, and what it doesn't
What it could mean: These AI programs cannot be trusted to read a brain MRI on their own. A trained doctor still reads your scans. This study does not change your care.
What it doesn't mean: This is not a treatment, and it is not a cure. It is an early test of computer programs on saved scan pictures, not in real patient care. It does not mean AI will read your scans soon, and it is not a promise of anything for patients.
Source: medRxiv (preprint), August 26, 2026 · Read the original
This plain-language summary was written by AI and published automatically after passing our automatic safety checks. How we write.