Microsoft Research introduces CARE-X, a unified vision-language model for chest X-ray interpretation

Microsoft Research describes CARE-X as a unified chest X-ray vision-language model (VLM) designed to support diverse clinical interpretation tasks, combining both generative and structured prediction outputs. The model uses reinforcement learning (DAPO) to reward clinical correctness across multiple tasks.

The research team notes that radiology AI must handle task diversity—such as generating detailed findings or concise impressions, answering presence/negation questions, identifying medical devices, and pinpointing abnormalities—while maintaining clinical accuracy. CARE-X integrates these capabilities into a single system, switching between generative and dual inference modes depending on the task.

CARE-X was validated on real-world clinical data from Narayana Health in India, including rare ICU pathologies and CT-confirmed enlargement conditions. The model’s outputs include free-text reasoning and deterministic predictions, with reinforcement learning used to align optimization with clinical correctness.

The announcement highlights gaps in current radiology VLMs, including a lack of calibrated confidence for diagnostic decisions, training objectives that do not prioritize clinically consequential errors, and an inability to perform measurement-dependent tasks like assessing cardiomegaly via cardiothoracic ratio calculations.

To address measurement-dependent findings, the team paired a separate research experiment with Qwen3-VL-4B-Instruct and deterministic measurement tools, evaluating whether direct computation could outperform visual approximation for conditions like cardiomegaly and mediastinal widening.