Introducing CARE-X: Implementing a Clinically Useful Radiology VLM via Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement

CARE-X is a research vision-language model (VLM) developed to handle various requirements of chest X-ray interpretation within a single model. It supports both generative tasks—such as detailed findings and impression generation, diagnostic yes/no and localization question answering, and medical device identification and placement assessment—and structured predictions.

Existing radiology VLMs present diagnoses in free-form text, which limits accuracy and confidence calibration, and makes it difficult to balance sensitivity and specificity according to clinical contexts. Conversely, discriminative models can provide calibrated scores but lack flexibility for open-ended report generation. CARE-X takes an approach that combines generative and discriminative capabilities to bridge this gap.

The core training methodology includes:

• Auxiliary supervision for diverse clinical tasks
• DAPO-based reinforcement learning that rewards clinical correctness
• A multi-task design that provides both free-text reasoning and deterministic structured outputs

In separate research experiments, deterministic measurement tools were connected to Qwen3-VL-4B-Instruct to investigate whether measurement-dependent disease assessment improved compared to methods relying solely on visual estimation. The tools calculate cardiac width and thoracic width to derive the cardiothoracic ratio (CTR) and display the results on the image.

CARE-X was evaluated using real Indian clinical data from Narayana Health, including rare ICU pathologies and cases of organ enlargement confirmed by CT. However, CARE-X is a research model, not a Microsoft product or medical device, and has not received regulatory approval or authorization for clinical diagnosis, screening, or patient care.