Microsoft Research introduces CARE-X radiology VLM for chest X-ray interpretation
Microsoft Research unveiled CARE-X, a research vision-language model for chest X-ray interpretation that combines generative and discriminative capabilities. The model aims to move beyond simple report generation by adding calibrated predictions and measurement-based tools. It is built on a SigLIP2 vision encoder and Phi-4-mini language model, with auxiliary heads for classification and grounding. The team reports strongest performance on most metrics across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient. CARE-X is explicitly a research model, not cleared for clinical use. The approach highlights a shift toward clinically useful AI that provides auditable, quantitative evidence alongside textual reports, which could aid regulatory approval and reduce diagnostic ambiguity.
Microsoft Research introduces CARE-X radiology VLM for chest X-ray interpretation
Microsoft Research introduces CARE-X, a radiology VLM for chest X-ray interpretation that combines generative and discriminative capabilities. It claims the strongest performance on most metrics across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient. CARE-X is a research model and not cleared for clinical use.
Key takeaway
Radiology AI's next frontier is calibrated, measurement-based reasoning that provides auditable evidence, rather than fluent report generation alone.
What happened
Microsoft Research introduced CARE-X, a vision-language model designed for chest X-ray interpretation. Unlike typical models focused on report generation, CARE-X integrates generative and discriminative capabilities, combining free-text flexibility with calibrated structured predictions and measurement-based tools. According to the research blog, the model is built on a SigLIP2-so400M vision encoder and a Phi-4-mini-instruct (3.8B) language model, connected via a lightweight adapter, with auxiliary heads for classification and visual grounding.
The team reports that CARE-X achieves the strongest performance on most reported metrics across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient. The model uses a three-stage supervised fine-tuning pipeline followed by DAPO-based reinforcement learning to optimize rewards for clinical reporting, diagnostic accuracy, and spatial grounding. Importantly, the authors note that CARE-X is a research model and not cleared for clinical use, with results being retrospective research findings.
Evidence
CARE-X combines generative and discriminative capabilities to support a broader range of radiology workflows.
Microsoft Research AI · attributed
CARE-X explores a unified approach that combines flexible reasoning, calibrated predictions, and measurement-based tools for chest X-ray interpretation.
CARE-X achieves the strongest performance on most metrics across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient.
Microsoft Research AI · attributed
It claims the strongest performance on most metrics across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient.
CARE-X is a research model and not cleared for clinical use.
Microsoft Research AI · attributed
CARE-X is a research model and not cleared for clinical use.
Why it matters
For healthcare operators, this shift towards measurement-based tools reduces diagnostic ambiguity and supports regulatory approval by providing auditable, quantitative evidence alongside textual reports.
Limits and uncertainties
CARE-X is a research model and has not been cleared or approved by any regulatory authority; it is not intended for clinical diagnosis, screening, or patient care.
Results described are retrospective research findings and do not establish the safety, effectiveness, or suitability of CARE-X for any clinical use.
Practical implications
Builders should prioritize calibrated predictions and measurement tools to enhance clinical trust and regulatory compliance. Integrating auxiliary heads for classification and grounding can provide auditable evidence alongside text output.
Operators must await regulatory clearance and validated prospective studies before any clinical deployment.
What to watch
Watch for Microsoft's follow-up on potential regulatory submissions or partnerships with clinical sites, and for updates on the model's generalization to other modalities like CT or MRI.
Original reporting: Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
