Introducing CARE-X: Towards Clinically Useful Radiology VLMs

Summary

Microsoft Research introduces CARE-X, a unified chest X-ray vision-language model (VLM) designed for diverse clinical interpretation tasks. It combines free-text generation with structured prediction to provide both narrative reports and deterministic outputs. The model uses reinforcement learning (DAPO) to reward clinical correctness in a multi-task setting. In a separate research experiment, the team paired Qwen3-VL-4B-Instruct with deterministic measurement tools to evaluate whether direct computation could improve performance on measurement-dependent conditions compared with visual approximation alone. CARE-X was validated on real-world Indian clinical data from Narayana Health, including rare ICU pathologies and CT-confirmed enlargement conditions. It is a research model, not approved by regulatory authorities, and not intended for clinical use.