TRACER: A Multimodal Multi-agent Workflow for Transparent Chest X-ray Assistance

Artificial intelligence (AI) systems for chest X-ray (CXR) interpretation can detect abnormalities, generate reports, provide visual grounding, and support dialogue but often offer limited mechanisms for inspection and follow-up. We present Transparent Radiology Assistance for Chest X-ray Explanation and Review (TRACER), a multimodal multi-agent workflow integrating CXR analysis, consultation support, and prior-report comparison with structured findings, gradient-weighted class activation mapping (Grad-CAM) cues, report-generation steps, and execution trajectories. Evaluation included report-quality benchmarking on 2000 frontal CXRs and user studies with 60 non-radiologist participants. Experiment II used 12 cases, with one fixed report per case in both text-only and transparency-supported conditions; the latter additionally displayed structured findings, finding-specific Grad-CAM cues, and execution trajectories. Open-ended responses were scored by an independent automated judge with clinician validation on a stratified subset. Experiment III evaluated eight prior-current pairs with or without an archive-assisted comparison paragraph. The full report pipeline achieved the best values on six of eight automatic metrics. Transparency support increased Consultation Preparation Quality from 2.133 ± 1.302 to 3.142 ± 1.154 (odds ratio 6.76, 95% confidence interval 3.92-11.67; $P<0.001$ ). Archive assistance improved difference identification from 0.642±0.338 to 0.833±0.285 (odds ratio 4.16, 95% confidence interval 2.33-7.42; $P<0.001$ ) and reduced mean completion time from 108.825 to 74.225 s. The transparency-supported condition also yielded higher usability ratings and lower perceived workload. These findings support TRACER as an inspectable workflow for CXR report review, verification-oriented consultation preparation, and prior-current comparison; clinical usefulness requires evaluation with patients and clinicians.