Archive · 9 Oct 2026
Study Finds AI May Mislocate Disease on Chest X-rays
Researchers tested four computer programs that help read chest X-rays.
They wanted to see whether the areas marked by the programs matched the areas doctors thought showed disease.
Sometimes a program named the disease correctly but marked the wrong place.
The programs also had more trouble finding abnormalities when they were not given the diagnosis.
A computer comparison favored MAIRA-2, but two radiologists gave MedGemma a higher rating.
Doctors may prefer to examine a wider area around a suspected problem.
The researchers say these tools should support doctors, not replace their decisions.
The study was accepted for a medical imaging event in Strasbourg in 2026.
Researchers at IIIT-H tested four AI models that interpret chest X-rays.
The models could identify a disease yet highlight an area that did not match radiologists’ judgments.
When diagnosis information was removed, the models became less effective at locating abnormalities.
MAIRA-2 scored better against doctors’ markings in computer-based comparisons, while radiologists rated MedGemma higher.
The study was accepted for the iMIMIC satellite event at MICCAI 2026 in Strasbourg.
- Who
- Researchers at the Language Technologies Research Centre at the International Institute of Information Technology Hyderabad, led by Prof. Parameswari Krishnamurthy, with two radiologists assessing highlighted areas.
- What
- A study evaluated whether four AI models’ marked areas on chest X-rays matched radiologists’ assessments.
- When
- The study was accepted for the iMIMIC satellite event at MICCAI 2026; no study date is specified.
- Where
- The research was conducted at IIIT-H; the event is in Strasbourg.
- Why
- To examine whether AI heatmaps correspond to where radiologists judge disease to be on chest X-rays.
- Research institution
- International Institute of Information Technology Hyderabad (IIIT-H)
- Research group
- Language Technologies Research Centre (LTRC)
- Models tested
- MAIRA-2, MedGemma-4B, LLaVA-Med-1.5 and LLaVA-1.5
- Image data
- Thousands of publicly available chest X-rays
- Clinical evaluators
- Two radiologists assessed the areas highlighted by the models
- Acceptance
- Accepted for the iMIMIC satellite event at MICCAI 2026
- Event location
- Strasbourg