Microsoft's CARE-X radiology VLM uses RL and tool-augmented measurement for clinical accuracy

Microsoft Research unveiled CARE-X, a unified chest X-ray vision-language model designed to handle diverse clinical interpretation tasks. Unlike typical VLMs that only generate free-text reports, CARE-X combines generative and discriminative capabilities, delivering both narrative findings and calibrated diagnostic scores. This dual-output approach allows clinicians to tune sensitivity-specificity trade-offs—a key requirement for real-world deployment. The model is trained using DAPO reinforcement learning, which rewards clinical correctness rather than token-level accuracy, meaning a missed life-threatening finding incurs a higher penalty than a harmless wording change.

CARE-X also explores tool-augmented reasoning for measurement-dependent conditions. In a separate research experiment, the team paired Qwen3-VL-4B-Instruct with deterministic measurement tools to assess whether direct computation improves performance on signs like cardiomegaly and mediastinal widening, compared with visual approximation alone. Validated on real-world Indian clinical data from Narayana Health—including rare ICU pathologies and CT-confirmed enlargement—CARE-X directly addresses three critical gaps in current radiology AI: missing calibrated confidence, cross-entropy loss that ignores clinical consequences, and no capability for measurement-based findings. Microsoft emphasizes CARE-X is a research model, not a cleared medical device.