Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
Microsoft says CARE-X improved chest X-ray reporting and structured prediction in retrospective tests, but it is still research-only and not cleared for clinical use.
The model combines free-text radiology responses with auxiliary heads for calibrated classification and grounding. Microsoft reports stronger results across several chest X-ray benchmarks, including 94% accuracy on ReXVQA and top CRIMSON scores in its comparison set. On Narayana Health data, CARE-X showed more balanced performance on rare ICU conditions, while a separate tool-augmented measurement experiment improved F1 scores for measurement-dependent findings. The note repeatedly limits the claims to retrospective research and says CARE-X is not a product, medical device, or diagnostic system. Microsoft Research AI's note
The model combines free-text radiology responses with auxiliary heads for calibrated classification and grounding. Microsoft reports stronger results across several chest X-ray benchmarks, including 94% accuracy on ReXVQA and top CRIMSON scores in its comparison set. On Narayana Health data, CARE-X showed more balanced performance on rare ICU conditions, while a separate tool-augmented measurement experiment improved F1 scores for measurement-dependent findings. The note repeatedly limits the claims to retrospective research and says CARE-X is not a product, medical device, or diagnostic system. Microsoft Research AI's note
score 5