TRACE: A Multimodal Multi-Agent System for Transparent Chest X-Ray Assistance

Abstract

Chest X-ray (CXR) AI has shown strong progress in abnormality detection, report generation, and multimodal medical reasoning, yet many existing systems remain limited as user-facing assistive tools. In particular, report-centered or text-only systems often provide limited support for evidence inspection, follow-up interaction, and cross-study continuity. This thesis presents TRACE (Transparent Radiology Assistance for Chest X-ray Explanation), a multimodal multi-agent system for transparent CXR assistance. TRACE integrates detector-supported image analysis, multimodal report drafting, report refinement, transparency-oriented interaction, consultation support, and archive-assisted comparison within a unified workflow.

TRACE was evaluated from four complementary perspectives: objective report quality, transparency-supported consultation preparation, archive-assisted cross-study comparison, and subjective usability and workload. Objective report quality was assessed on a 2,000-image frontal CXR benchmark constructed from MIMIC-CXR and the Indiana University Chest X-rays dataset, using external comparisons against LLaVA-Med and the MedRAX report-generation tool as well as an internal ablation of the TRACE reporting pipeline. TRACE achieved the best overall profile across most reported semantic and label-oriented metrics, indicating that detector-supported finding integration and report refinement can improve report quality within a broader assistive workflow. Two controlled user studies and a questionnaire-based human-factors assessment were also conducted with 48 participants. The transparency-oriented interface significantly improved users' ability to identify verification-worthy report conclusions, justify why they warranted clinician verification, and formulate more specific verification-oriented consultation questions. Archive support significantly improved report-level difference identification and reduced cross-study comparison time. In addition, the full TRACE interface was associated with higher perceived usability and lower subjective workload than the text-only interface, with the largest workload reductions observed in mental demand and frustration.

Overall, this thesis shows that, within the controlled benchmark and user-study settings examined here, CXR AI can provide measurable assistive value when it is designed and evaluated not only for report generation, but also for evidence inspection, verification-oriented interaction, and continuity across encounters within a unified assistive workflow. Clinical usefulness remains to be established through further validation with patients and clinicians.