A new approach to radiology AI
Microsoft researchers introduced CARE-X, a research model designed to unify diverse radiology AI tasks such as report generation, abnormality localization, and device assessment while maintaining clinical accuracy.
Microsoft’s research team unveiled CARE-X, a radiology AI model developed to address the varied demands of chest X-ray analysis, including generating detailed reports, identifying medical devices, and pinpointing abnormalities. Unlike previous models, CARE-X integrates generative and discriminative capabilities, enabling it to produce both free-text outputs and structured predictions with calibrated confidence scores. The model is built on a SigLIP2 vision encoder and a Phi-4-mini-instruct language model, augmented with task-specific heads for classification and visual grounding.
The model employs a three-stage fine-tuning pipeline followed by reinforcement learning to optimize clinical reporting, diagnostic accuracy, and spatial grounding. Testing on datasets like MIMIC-CXR and IU-Xray showed CARE-X outperforming other models in most metrics, including a 94% accuracy rate on the ReXVQA benchmark. A key innovation is its ability to perform dual inference, producing both autoregressive responses and structured predictions in a single forward pass, which enhances flexibility and clinical relevance.
CARE-X’s auxiliary heads provide calibrated probability scores with adjustable thresholds, allowing clinicians to prioritize sensitivity or specificity based on clinical needs. This capability is particularly useful for screening versus confirmation scenarios. The model’s performance on anatomical grounding tasks improved by up to 28.2 percentage points in mean average precision, demonstrating its potential to enhance diagnostic precision in real-world settings.
In a separate experiment, researchers combined a vision-language model with deterministic measurement tools to enable quantitative assessments, such as aortic dilation screening, which are not typically quantified on chest X-rays. The approach improved accuracy in threshold-dependent diagnoses and could expand the range of assessments derived from routine radiographs, potentially enabling earlier detection of conditions like aortic dilation.