Source
Care-X makes radiology AI more helpful in the clinic
CARE-X is a research vision-language model that combines generative report writing and structured diagnostic predictions for chest X-ray interpretation. It utilizes auxiliary supervision and reinforcement learning to improve clinical accuracy across various radiology tasks.
Why it matters
Combining text generation with precise measurement tools could help clinicians identify borderline conditions like aortic dilation often missed during initial screenings, potentially improving cardiovascular care outcomes.
The details
- CARE-X achieved 94% accuracy on the ReXVQA benchmark as of August 2026.
- Tool-augmented measurement improved diagnostic accuracy by an average of 43.6 percentage points.
- The model was validated using retrospective clinical data from Narayana Health in India.
Show entities and relationshipsHide entities and relationships
In this article
Topics
Technologies
Countries
Key connections
Microsoft developed CARE-X as a research model for chest X-ray interpretation.
Microsoft owns Phi-4-mini-instruct
Microsoft developed Phi-4-mini-instruct, which serves as the language backbone for CARE-X.
CARE-X is built with Phi-4-mini-instruct
CARE-X is built on the Phi-4-mini-instruct language model backbone.
CARE-X is built with SigLIP2-so400M
CARE-X incorporates the SigLIP2-so400M vision encoder.
CARE-X is built with Vision Language Models
CARE-X is a unified vision-language model for radiology tasks.
CARE-X uses Reinforcement Learning
CARE-X uses reinforcement learning (DAPO) to reward clinical correctness.
Show 11 more connectionsShow fewer connections
Microsoft is a partner of Narayana Health
Microsoft researchers collaborated with Narayana Health to evaluate CARE-X on real-world Indian clinical data.
Microsoft is a partner of Medha AI
Microsoft collaborated with Medha AI on clinical AI research.
CARE-X was evaluated against CheXOne on chest X-ray interpretation benchmarks and ICU pathology classification.
CARE-X was evaluated against MedGemma across radiology report generation and diagnostic classification.
CARE-X competes with CheXOne-R1
CARE-X was benchmarked against CheXOne-R1 on the ReXVQA dataset.
CARE-X was validated on real-world clinical data from Narayana Health in India.
Narayana Health is located in India
Narayana Health is a major healthcare provider network based in India.
Qwen3-VL-4B-Instruct is related to CARE-X
Qwen3-VL-4B-Instruct was evaluated in a separate research experiment alongside CARE-X to test tool-augmented quantitative measurement.
CARE-X is related to IHF Innovation Hub
The AI-based aortic dilatation screening application based on CARE-X research was selected as a finalist for showcase at the IHF Innovation Hub.
CARE-X is related to World Hospital Congress
CARE-X research findings are showcased at the World Hospital Congress 2026.
A study on CARE-X measurement-driven reasoning for aortic dilation was accepted at the EACTS conference 2026.
Related events
Microsoft Researchers Present CARE-X Radiology VLM
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.