AI checks chest X-rays globally, then inspects regions to explain its reports

Automated systems can produce plausible descriptions of chest X-rays, but a reader may not be able to tell what visual evidence led to a particular finding. Huang and colleagues propose a way to expose more of that path: have the model survey the whole image, inspect selected anatomical regions, interpret their visual evidence and then compose a report.

The framework, called ARIVE, is designed for chest radiographs. During generation, it first takes in the image as a whole, then requests selected anatomical regions. It encodes those local views and feeds them back into the model before it continues its interpretation. The resulting sequence records region requests, intermediate statements and the final report, so a reviewer can trace a report claim to earlier steps and associated image regions.

How the system was trained and tested

The researchers built 159,416 multimodal reasoning trajectories from the MIMIC-CXR dataset. They started with existing radiology reports and used a large language model to reorganize their findings into stepwise narratives. They then attached image-region boxes from Chest ImaGenome annotations, covering 29 predefined anatomical regions. These are report-derived training examples, not direct records of how a radiologist actually reasoned through each case.

ARIVE was first trained to produce the trajectories and reports. The team then applied an optimization method called Anatomical and Clinical-Content Optimization (ACCO), which rewards anatomical grounding and agreement with clinical content in both intermediate reasoning and the final report. Some rewards give extra weight to less common findings or regions.

The evaluation used MIMIC-CXR and MIMIC-ABN, an abnormality-focused subset of MIMIC-CXR. The authors compared ARIVE with six other systems using automated measures of clinical findings, relationships between findings and report errors. They also tested anatomical localization, asked an experienced radiologist to grade a sample of reports and reasoning traces, and compared review of reports with and without the traces.

What the study found

On both datasets, ARIVE scored higher than the compared models across the reported clinical-efficacy measures. On MIMIC-CXR, it led on all seven such measures; on MIMIC-ABN, it led on all six tabulated measures. The lead was not statistically significant in every comparison, and ARIVE did not lead on measures of wording overlap with reference reports. That distinction matters: matching clinical content and matching a reference’s phrasing are different goals.

The authors’ training comparisons suggest that regional image evidence and clinical-content optimization played complementary roles. Adding text-only reasoning did not consistently improve report quality. Interleaving regional inspection improved most clinical-efficacy measures compared with text-only reasoning, though one main CheXbert score dipped slightly on MIMIC-CXR. Adding ACCO to that setup improved all six tabulated clinical-efficacy measures on both datasets.

In blinded grading of 100 cases, ARIVE’s average report score was similar to Qwen3.5’s and higher than PriorRG’s, while the reference reports received the highest average. The radiologist rated ARIVE’s reasoning traces highly for diagnostic traceability. In a separate paired review, the radiologist identified errors in 11 reports when traces were available, compared with 8 without them. Median review time also increased, from 24.5 to 42.5 seconds per report. This is a preliminary result from one reviewer and a limited sample, not proof that traces improve clinical outcomes.

What remains uncertain

A visible reasoning trail is not the same as a correct one. The paper describes an unsupported finding that appeared in an intermediate step and carried through to the final report. The authors caution that generated reasoning should be checked against the image, rather than trusted because it is detailed or coherent.

The study also has important scope limits. Its evaluation used frontal chest X-rays from a single-institution dataset; it does not establish how the approach performs across hospitals, reporting styles, other imaging views or other body regions. The trajectories were reconstructed from reports and may inherit omissions, local conventions or rewriting errors. Anatomical supervision was sparse for some regions, and localization there remained weaker. Expert assessments involved small samples and a single radiologist.

The paper is an arXiv preprint, not evidence of established clinical effectiveness. Overall, ARIVE offers a way to make parts of AI-generated chest X-ray reports more inspectable. Whether that visibility leads to safer or more useful reporting in real clinical workflows remains to be tested.