Microsoft Research Unveils CARE-X Radiology Model
Microsoft Research has developed CARE-X, a chest X-ray vision-language model that combines text generation with structured predictions to improve diagnostic accuracy and clinical utility.
Microsoft Research has introduced CARE-X, an experimental chest X-ray vision-language model designed to bridge the gap between generative AI and clinical precision. Built on a SigLIP2-so400M vision encoder and a 3.8-billion-parameter Phi-4-mini-instruct language model, the system uses a lightweight adapter and task-specific auxiliary heads. This allows the model to perform dual inference, generating free-text reports while outputting calibrated diagnostic scores. Researchers trained it using three-stage supervised fine-tuning followed by Direct Alignment Parameter Optimization (DAPO) reinforcement learning.
On the Chest ImaGenome anatomical grounding benchmark, the auxiliary head increased mean average precision (mAP) by 28.2 percentage points and mean intersection over union (mIoU) by 6.2 percentage points. On PadChest phrase grounding, it gained 24.6 percentage points mAP and 14.1 percentage points mIoU. Through DAPO, the model's generative output reached 0.868 mAP on anatomical grounding, surpassing the supervised detection head's 0.865 mAP. Additionally, CARE-X ranked first on the ReXVQA leaderboard as of August 2026, achieving 94% overall accuracy across 41,007 question-answer pairs.
The researchers validated CARE-X on 1,047 de-identified chest radiographs from Narayana Health, targeting rare intensive care unit pathologies with prevalence rates between 2.6% and 5.2%. In a separate experiment, they paired Qwen3-VL-4B-Instruct with deterministic measurement tools. This tool-augmented approach improved the average F1-score by 43.6 percentage points across five conditions, including cardiomegaly (rising from 74.56 to 96.00) and descending aorta enlargement (from 28.57 to 100.00). On 122 CT-confirmed cases, the tool-augmented pipeline achieved 94.26% recall.
For medical practitioners, these developments address the limitations of purely generative models. By offering adjustable sensitivity thresholds, CARE-X allows clinicians to tune the model for high-sensitivity screening or high-specificity confirmation. The integration of measurement tools also helps identify subtle conditions like mild aortic dilation, which was detected in 93% of cases (40 of 43) compared to just 12% (5 of 43) in initial human radiology reads.
This is our own summary of reporting by Microsoft Research Blog
