Chest X-ray reports that show their work: ARIVE inspects anatomy before diagnosing, matching Qwen3.5 in blinded grading

Chest X-ray report generation can now inspect the image region by region, state an interpretation, and only then write the report, so each conclusion traces back to an anatomical region and an intermediate judgment. Earlier attention visualizations, retrieval of similar cases or graph-based methods could point to regions tied to a finding, but did not show how evidence from a region updates the diagnostic interpretation; text-only chain-of-thought made reasoning readable without guaranteeing each step was grounded in the image. On MIMIC-CXR, compared with Qwen3.5, MedGemma-1.5, MAIRA-2, Libra, MLRG and PriorRG, ARIVE scored highest on all seven clinical-efficacy measures, with 40 of 42 comparisons significant; the two exceptions were 14-observation macro against Qwen3.5 and against PriorRG.

During reasoning the model requests an anatomical region, has that local crop encoded and inserted back into its context, and continues interpreting on that evidence. The pipeline is trained with multimodal chain-of-thought supervision plus Anatomical and Clinical-Content Optimization, which jointly rewards anatomical grounding, reasoning-stage observation accuracy, final-report observation accuracy and preservation of report entities and relations. Training trajectories come from MIMIC-CXR and are anchored to 29 Chest ImaGenome regions; removing the anatomical-grounding reward lowered every reported measure, and removing the report-stage clinical reward produced the largest drop in final-report micro, from 0.505 to 0.463 on the Gold subset.

The findings rest on single-institution MIMIC-CXR data and its abnormality subset MIMIC-ABN, use frontal radiographs only, and draw anatomical supervision from 29 predefined Chest ImaGenome regions rather than lesions or volumetric structures; the authors propose future work in which models learn to identify diagnostically relevant regions without bounding-box annotations.

Upper and mid lung zones plus apical regions together account for only about 2.8% of bounding-box annotations, and grounding for them stayed weaker; the authors suggest this may relate to low CheXbert scores for Consolidation and Pneumothorax but call that explanation hypothetical. The chain-of-thought trajectories were rewritten from reports by a large language model and may inherit omissions and institutional reporting habits, so they should not be read as faithful reconstructions of radiologist cognition. With trajectories available, the radiologist flagged errors in 11 rather than 8 reports, while median review time rose from 24.5 s to 42.5 s.