ARIVE: Anatomy-Guided Reasoning Makes AI-Generated Radiology Reports Traceable

ARIVE uses anatomy-guided interleaved visual reasoning to generate traceable radiology reports, improving clinical effectiveness and error review efficiency.
The ARIVE framework tackles black-box reasoning in radiology report generation with a multimodal chain-of-thought method that interleaves anatomical inspection and diagnostic reasoning. Using 159,416 reasoning trajectories built from MIMIC-CXR and optimized with anatomy and clinical content rewards, ARIVE outperforms strong baselines on MIMIC-CXR and MIMIC-ABN. Expert blind review shows competitive report quality, and paired review evaluation shows that access to the multimodal chain-of-thought helps reviewers identify errors, making interpretability directly serve quality control. The article notes that trajectory construction costs and generalization to real clinical settings remain open challenges.
Automated radiology report generation can already produce text that looks clinically plausible, but the reasoning path from visual evidence to final conclusion remains a black box—physicians find it hard to review, revise, or judge whether the report's basis is reliable. To address this pain point, researchers propose the ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence) framework, which interleaves anatomical region inspection with diagnostic reasoning, making the report generation process interpretable.

From "Black-Box Output" to "Interleaved Reasoning"
Traditional report generation models typically take images as input and output text directly, with no verifiable reasoning steps in between. ARIVE's core idea is to simulate the actual workflow of a radiologist: first locate and examine a specific anatomical region, form a preliminary judgment based on visual evidence, then move to the next region for inspection and reasoning, and finally aggregate everything into a complete report. This interleaved "anatomical inspection—diagnostic reasoning" Multimodal Chain-of-Thought (MCoT) explicitly records the visual basis and clinical interpretation at each step.
To train this process, the research team constructed 159,416 multimodal chain-of-thought trajectories from the MIMIC-CXR dataset, each containing the visual evidence of the examined anatomical region and its corresponding clinical explanation. The model is trained on these trajectories and further optimized with anatomy rewards and clinical content rewards, encouraging reasoning steps that both conform to anatomical structure and align with clinical meaning.
Multimodal Chain-of-Thought (MCoT) is an extension of the Chain-of-Thought approach to multimodal scenarios. Traditional chain-of-thought mainly targets pure text tasks, improving the model's ability to solve complex problems and its interpretability by having it explicitly generate intermediate reasoning steps. MCoT, in contrast, requires the model to simultaneously reference visual evidence (such as specific regions in an image or the appearance of lesions) and textual explanations during reasoning, forming an alternating "see—think—say" sequence. In medical imaging reports, each step of MCoT typically corresponds to the visual inspection of an anatomical region and its clinical interpretation, allowing the final conclusion to be traced back to specific image evidence. This structure not only facilitates human review but also provides finer-grained signals for process supervision during training.
Clinical Effectiveness Outperforms Strong Baselines
On both the MIMIC-CXR and the abnormality-centric MIMIC-ABN datasets, ARIVE's clinical effectiveness metrics outperform strong baseline models. This means it not only writes reports that "look more like" real ones but also performs better in the coverage and accuracy of key clinical information.
More noteworthy are the expert blind review results: reports generated by ARIVE are competitive with those from leading models. Meanwhile, its multimodal chain-of-thought trajectories received high scores for diagnostic traceability—reviewers were able to trace the source of report conclusions along the reasoning chain. A paired review evaluation further showed that when reviewers had access to the MCoT, they were more likely to identify errors in generated reports. This is particularly important for clinical deployment: interpretability is not just for "looking good" but directly serves report quality control.
MIMIC-CXR and MIMIC-ABN are widely used public datasets in the field of radiology report generation. MIMIC-CXR contains approximately 370,000 chest X-rays and corresponding free-text reports, sourced from Beth Israel Deaconess Medical Center, and is one of the largest open chest imaging-report paired resources currently available. MIMIC-ABN is a subset filtered from MIMIC-CXR that is abnormality-centric, focusing on cases with clear pathological findings, and is commonly used to evaluate a model's ability to capture abnormal signs. Since the report texts were written by radiologists, these two datasets both provide realistic clinical language distributions and retain the noise and incompleteness in the reports. Therefore, metric improvements on these datasets are usually regarded as an important reference for method effectiveness, but one should also note the generalization limitations brought by their single-center origin.
How Interpretability Assists Error Review
Errors in radiology reports are often hidden in the misreading or omission of a certain anatomical region. ARIVE associates each diagnostic conclusion with specific visual inspection steps, allowing reviewers to quickly locate which region's evidence is insufficient or where the reasoning jumps too far. The paired review evaluation shows that this traceability helps reviewers find errors, rather than judging reliability solely by the fluency of the final text.
From an application perspective, this framework offers new interaction possibilities for AI-assisted reporting systems: physicians no longer simply accept or reject a piece of text but can review the reasoning chain and correct specific steps, thereby increasing trust in the system. Of course, generating and annotating MCoT trajectories is costly, and how to extend to more modalities and a broader range of abnormality types remains a direction for future exploration.
Paired review evaluation is an experimental design used to measure the practical utility of interpretability: reviewers view model outputs together with their explanatory information (such as reasoning chains) and compare them in a paired manner against review results under no-explanation conditions. In radiology report scenarios, this design can quantify "whether access to MCoT truly helps reviewers find errors faster and more accurately," rather than merely subjectively evaluating whether explanations "look reasonable." The paired design controls for individual reviewer differences and case difficulty, making conclusions more statistically powerful. This evaluation paradigm originates from the need in the explainable AI field for empirical testing of "explanation usefulness," emphasizing that explanations should serve real downstream tasks (such as error detection and quality control) rather than being merely a byproduct of the model's internals.
Editorial View
ARIVE's value lies in turning "interpretability" from an add-on attribute into an intrinsic structure of the generation process. Interleaved reasoning makes the mapping between visual evidence and clinical conclusions explicit, which both benefits model optimization and provides a handle for clinical review. However, current evaluations are still concentrated on the MIMIC series datasets, and the diversity and noise levels in real clinical scenarios may bring new challenges. In addition, the quality of chain-of-thought trajectories depends heavily on how they are constructed, and how to ensure their clinical faithfulness rather than merely "looking reasonable" is a common issue faced by such methods.
One-line Summary
ARIVE uses anatomy-guided interleaved visual reasoning to generate traceable radiology reports, improving clinical effectiveness and error review efficiency.
Paragraph Summary
The ARIVE framework addresses the black-box reasoning problem in radiology report generation by proposing a multimodal chain-of-thought method that interleaves anatomical region inspection with diagnostic reasoning. Based on MIMIC-CXR, the study constructed 159,416 reasoning trajectories containing visual evidence and clinical explanations, and optimized the model with anatomy and clinical content rewards. On MIMIC-CXR and MIMIC-ABN, ARIVE's clinical effectiveness outperforms strong baselines, and expert blind review shows its report quality is competitive with leading models. Paired review evaluation further indicates that access to the multimodal chain-of-thought helps reviewers identify report errors, making interpretability directly serve quality control. The article argues that this method internalizes interpretability into the generation structure, but trajectory construction costs and generalization to real clinical scenarios remain unresolved challenges.
Related articles

GPT-6.2 and Astra Launch Together: OpenAI Has Finally Figured Out Its Release Cadence
OpenAI releases GPT-6.2 Sol and Astra together, with Plus users able to try them in ChatGPT Work and Codex. This article breaks down the positioning differences between the two models, Astra's practical use in long-document retrieval and rewriting, and how the simultaneous dual-model launch affects user workflows.

GPT-6.2 Dual-Model Hands-On: How Big Is the Gap Between Astra and Sol in Clock-Making and Game-Building?
GPT-6.2 Astra vs Sol hands-on: Blender clock, neon shooter, and Formula racing compared. Astra spends $419 with strong execution; Sol costs $197 but drops to 2 FPS. Both still trail Fable and Opus in games and aesthetics.

GPT-6.2 Astra and Sol Released: A Big Cost Drop, and How to Choose Between the Two Models
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol: Astra tackles professional tasks, Sol targets low-cost everyday use, with cost per task cut to under a third.