IIIT Hyderabad study found AI chest X‑ray heatmaps can diverge from radiologist markings

Researchers at IIIT Hyderabad’s Language Technologies Research Centre evaluated four vision-language models (MAIRA-2, MedGemma-4B, LLaVA-Med-1.5 and LLaVA-1.5) on thousands of publicly available chest X-rays and compared model attention heatmaps with regions identified by two radiologists. The study reported discrepancies between the regions highlighted by the AI models and those marked by radiologists, and it was accepted at MICCAI 2026 for presentation at the iMIMIC Satellite Event in Strasbourg.