CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
CARE-X is a chest X-ray vision-language model that unifies auxiliary discriminative supervision with reward-aligned generation. It augments a generative backbone with focal-loss classification and composite-loss grounding heads, co-trained with the language-modeling objective. This produces discriminative diagnostic predictions with tunable decision thresholds and precise spatial localization. The system also uses Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) with task-specific reward signals for report generation, visual question answering, and spatial grounding.
CARE-X is a chest X-ray VLM that integrates auxiliary discriminative supervision and reward-aligned generation to improve clinical usefulness. It adds classification and grounding heads to a generative model, enabling tunable diagnostic thresholds and spatial localization while enhancing report quality. DAPO leverages task-specific rewards for report generation, VQA, and grounding.
The model demonstrates that co-training generative language modeling with discriminative heads (focal loss for classification, composite loss for grounding) can mutually reinforce performance, yielding both better structured predictions and improved report generation. The use of DAPO for reward-aligned fine-tuning across multiple tasks suggests a scalable approach to aligning VLMs with clinical utility metrics.
This work addresses a critical gap in radiology AI by moving beyond report generation to include actionable outputs like tunable classification and spatial localization. It signals a trend toward more clinically integrated VLMs that can support diagnostic workflows with measurable, threshold-adjustable outputs.
By providing tunable decision thresholds and precise localization, CARE-X could reduce radiologist workload, improve diagnostic consistency, and enable new AI-assisted measurement tools. This has potential for commercial deployment in radiology departments and teleradiology services.
Next signals to watch include validation on external datasets, integration with PACS systems, regulatory submissions, and extensions to other imaging modalities. The approach may also influence the design of generalist medical VLMs that combine generation with structured prediction.