Microsoft Research has introduced CARE-X, an experimental vision-language model designed to improve how AI systems interpret chest X-rays across reporting, classification and image-grounding tasks.
The research combines generative AI with structured prediction capabilities, reinforcement learning and a separate experiment involving deterministic measurement tools.
CARE-X achieved 94% overall accuracy on the ReXVQA benchmark, which includes 41,007 question-and-answer pairs across clinically relevant categories. Microsoft said that result was six percentage points higher than the next-best publicly reported model as of August 2026.
One Model Handles Multiple Radiology Tasks
CARE-X is designed to perform several types of chest X-ray interpretation.
The model can generate detailed findings and diagnostic impressions, identify the presence or absence of abnormalities, determine disease locations, identify medical devices and detect incorrectly positioned tubes and lines.
It combines generative responses with auxiliary classification and grounding components that can provide structured predictions and confidence scores.
CARE-X uses a SigLIP2 vision encoder connected to the 3.8-billion-parameter Phi-4-mini-instruct language model.
Microsoft trained the system using supervised fine-tuning followed by DAPO reinforcement learning with rewards designed around reporting accuracy, diagnostic performance and spatial grounding.
Measurement Tools Improve Performance
Microsoft separately explored whether vision-language models could improve measurement-dependent diagnoses by calling deterministic tools instead of estimating measurements visually.
Researchers connected Qwen3-VL-4B-Instruct to tools capable of identifying anatomical landmarks and calculating measurements.
Across five evaluated conditions, tool-assisted inference improved F1 performance by an average of 43.6 percentage points compared with perception-only analysis.
The approach allows the AI model to handle image interpretation while specialized software performs calculations requiring greater numerical precision.
Tested on Real-World Clinical Data
CARE-X was evaluated retrospectively using 1,047 de-identified chest X-rays from Narayana Health covering five relatively rare high-acuity conditions.
A separate evaluation involving 122 CT-confirmed positive cases found the tool-augmented approach achieved 94.26% recall, 10.65 percentage points above the strongest perception-only baseline in that study.
Microsoft stressed that CARE-X remains a research project rather than a clinical product. It has not been cleared or approved by a regulatory authority and is not intended for diagnosis, screening, patient care or clinical decision-making.
The research nevertheless demonstrates a broader approach to specialized AI automation in which models can orchestrate purpose-built tools rather than relying exclusively on their own generated predictions.