Microsoft Research has introduced CARE-X, a research model designed to address gaps in current radiology vision-language models by combining generative report writing with calibrated diagnostic predictions. The system uses reinforcement learning (DAPO) to reward clinical correctness and pairs the Qwen3-VL-4B-Instruct model with deterministic measurement tools to evaluate performance on conditions requiring precise calculations.
- CARE-X combines a SigLIP2-so400M vision encoder and a Phi-4-mini-instruct (3.8B) language model, augmented with task-specific auxiliary heads for classification and visual grounding.
- The model employs dual inference, producing both free-text responses and structured predictions with confidence scores in a single forward pass.
- Training involves a three-stage supervised fine-tuning pipeline followed by DAPO-based reinforcement learning to optimize rewards for clinical reporting and diagnostic accuracy.
- A separate experiment paired Qwen3-VL-4B-Instruct with deterministic measurement tools to calculate cardiac and thoracic widths, comparing direct computation against visual approximation.
- The model was validated on real-world Indian clinical data from Narayana Health, including rare ICU pathologies and CT-confirmed enlargement conditions.
CARE-X aims to provide a unified approach that supports diverse radiology workflows while maintaining clinical fidelity, addressing the need for both expressive report generation and precise, threshold-adjustable diagnostic outputs.