Microsoft researchers introduced CARE-X, a unified chest X-ray vision-language model combining generative report generation, dual-inference auxiliary heads, and reinforcement learning.

Key Details

• CARE-X is built on SigLIP2-so400M vision encoder and Phi-4-mini-instruct language backbone.
• Evaluated on real-world Indian clinical data from Narayana Health across 1,047 de-identified chest radiographs.
• Achieved 94% accuracy on the ReXVQA benchmark and was selected as a finalist at the IHF Innovation Hub, World Hospital Congress 2026.

In This Story