AI heatmaps on chest X‑rays often miss radiologists' markings, study finds

UPDATE - follow-up to IIIT-H study caution doctors against relying entirely on AI for medical diagnosis: A research team from IIIT‑H examined four vision‑language models that produce heatmaps on chest X‑rays. They found the highlighted regions often did not line up with the areas a radiologist marked as diseased. The lead author Parameswari Krishnamurthy said the goal was to check whether model heatmaps truly indicate disease location. Principal investigator Syed Faizan explained that a model could first decide on a diagnosis and then generate a heatmap to match that conclusion, rather than pinpointing the abnormality from the image itself. When the team stripped the diagnostic label from the input, the models’ ability to locate abnormalities fell, suggesting they rely on disease expectations to guide heatmaps. The audit also revealed performance differences among the four models, with some focusing tightly on a suspected spot but missing the spread of disease that radiologists assess. The work, accepted for presentation at MICCAI 2026’s iMIMIC satellite event in Strasbourg, also probed how phrasing of clinical questions affects model responses, showing sensitivity to wording. The researchers stressed that AI tools should be checked against clinical judgment and that future AI evaluation must include how well visual explanations match expert markings, not just overall accuracy. ## Context Vision‑language models are being explored to help doctors read images faster, but their explanatory outputs can mislead if they do not truly reflect disease locations.