Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
| Source: Microsoft Research Blog
Tags: Microsoft Research, CARE-X, radiology AI, chest X-ray, VLM, reinforcement learning, medical AI
Microsoft Research's CARE-X unifies free-text radiology report generation with calibrated structured diagnostics in a single chest X-ray VLM, using DAPO reinforcement learning for clinical alignment and validated on real Indian hospital data including rare ICU pathologies.
Details
Microsoft Research published a research note on CARE-X, a unified vision-language model for chest X-ray interpretation. Unlike existing radiology AI that handles either generative reporting or structured prediction, CARE-X combines both: it can generate detailed findings and impressions, answer structured diagnostic questions, identify medical device placement, and localize abnormalities—all from a single model.\n\nClinical alignment uses DAPO (a distributed actor-policy optimization approach), which rewards clinical correctness across multiple tasks simultaneously rather than optimizing for fluency alone. A separate research experiment paired Qwen3-VL-4B-Instruct with deterministic measurement tools, finding that direct computation outperforms visual approximation for measurement-dependent conditions like cardiac enlargement—a finding with practical implications for hybrid AI+tool pipeline design.\n\nValidation used real-world clinical data from Narayana Health in India, including rare ICU pathologies and CT-confirmed enlargement conditions—a meaningful departure from benchmark-only evaluations that strengthens generalizability claims across healthcare settings. Microsoft explicitly states CARE-X is a research model without regulatory clearance; it is not approved for clinical diagnosis, screening, or patient care in any market.