A Stricter Test Shows MRI Scan Predictions Are Weaker Than They Looked
Original title: Nested Cross-Validation Reveals Performance Inflation in MRI Radiomics for Early Mortality Prediction in IDH-Wildtype Glioblastoma.
How far along is this research?
- Lab cells
- Animals
- Review
- Tested in people
This was tested in people. That is the most reliable kind of research we share.
This was a check of computer methods on stored scans, not a new treatment for patients.
The short version
Computer tools that read brain scans to predict survival may work less well than earlier studies suggested.
What was studied. Researchers used brain MRI scans from 60 people with IDH-wildtype glioblastoma: The fastest-growing type of glioma (grade 4). Treatment usually starts soon after diagnosis: surgery first, then radiation and chemotherapy. See the glossary in a public dataset. They tested whether computer measurements from those scans could predict who would die within one year, or within six months.
What they found. The scan tools looked good under a loose testing method. Under a stricter testing method, they did much worse at predicting death within one year. The looser method made the scans look about 27% better than they really were. For death within six months, the gap was smaller, about 8%. Simple measures of tumor size stayed the most steady across tests.
What this means, and what it doesn't
What it could mean: Be careful with news that brain scans can predict how long someone will live. Some of those results may come from weak testing, not real skill. In this group, plain facts still mattered. People who had more of the tumor removed lived longer, and older age went with shorter survival.
What it doesn't mean: This does not mean a new treatment, and it is not a cure. It is not a test you can ask your doctor for. Nothing here changes your care today. The study only looked back at scans from 60 people at one center. The authors say their numbers apply to this dataset alone and should not be treated as a general rule.
Source: PubMed, August 12, 2026 · Read the original
This plain-language summary was written by AI and published automatically after passing our automatic safety checks. How we write.