Now Reading: IIIT-Hyderabad study warns against blind trust in medical AI models
-
01
IIIT-Hyderabad study warns against blind trust in medical AI models
A study by IIIT-Hyderabad researchers has warned that medical AI tools used to analyse chest X-rays may identify a disease correctly but highlight the wrong area. The team audited four vision-language models, MAIRA-2, MedGemma-4B, LLaVA-Med-1.5 and LLaVA-1.5, on thousands of chest X-rays and found that the models’ performance dropped when diagnostic information was removed, suggesting some image-based reasoning may be influenced by an anatomical expectation learned from the diagnosis.
The researchers also conducted a reader study with two radiologists. While computer-based assessments ranked MAIRA-2 higher for overlapping heatmaps, the radiologists rated MedGemma higher because it provided a broader area around the abnormality. Dr Syed Faizan, principal investigator, said a radiologist may want to see not only the disease spot but also the surrounding area to assess the extent of spread. The study has been accepted at the iMIMIC satellite event of MICCAI 2026 in Strasbourg.
Indian Opinion Analysis
Both telanganatoday and deccanchronicle reported the IIIT-H study as straight, neutral wire coverage of a research finding, with no pro-government or critical framing. Telanganatoday led with the caution against blind trust, while deccanchronicle led with the finding that AI may mark the wrong area. Both outlets attributed the same quotes to Dr Faizan and reported the same details about the models, the two-radiologist study, and the conference acceptance. The uniform coverage means there is no framing divide to reconcile. The key takeaway from the study itself is the practical gap between what an AI model’s heatmap shows and what a radiologist needs: the model’s narrow focus may be less useful than a broader view, and the same model that scores best in overlap may not be preferred by human readers.
Coverage: 2 sources, 2 neutral
Sources (2): telanganatoday.com (neutral report), deccanchronicle.com (neutral report)
This brief was synthesised by AI from the 2 sources linked above, so one read covers every framing they carry. Methodology and corrections.
