ARIVE
An anatomy-guided radiology report generation framework that generates a traceable multimodal chain of thought by interleaving visual inspection of anatomical regions with diagnostic reasoning, improving report interpretability and clinical validity.
Timeline (last 90 days)
ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence) is a radiology report generation framework that interleaves anatomical region inspection with diagnostic reasoning.
The research team constructed 159,416 multimodal chain-of-thought trajectories from the MIMIC-CXR dataset, each trajectory containing visual evidence of the inspected anatomical region and its corresponding clinical interpretation.
The ARIVE model is trained on the constructed multimodal chain-of-thought trajectories and further optimized with an anatomy reward and a clinical content reward.
On both the MIMIC-CXR and MIMIC-ABN datasets, ARIVE's clinical efficacy metrics outperform strong baseline models.
Expert blind evaluation results show that the reports generated by ARIVE are competitive with those of leading models.
ARIVE's multimodal chain-of-thought trajectories received high scores for diagnostic traceability, with reviewers able to trace back along the reasoning chain to the sources of the report's conclusions.
A paired-review evaluation showed that when reviewers had access to MCoT, they more easily identified errors in generated reports.
Current evaluations of ARIVE remain concentrated on the MIMIC series of datasets, and the diversity and noise levels in real clinical settings may bring new challenges.
All Facts (8)
ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence) is a radiology report generation framework that interleaves anatomical region inspection with diagnostic reasoning.
50%UnverifiedThe research team constructed 159,416 multimodal chain-of-thought trajectories from the MIMIC-CXR dataset, each trajectory containing visual evidence of the inspected anatomical region and its corresponding clinical interpretation.
50%UnverifiedThe ARIVE model is trained on the constructed multimodal chain-of-thought trajectories and further optimized with an anatomy reward and a clinical content reward.
50%UnverifiedOn both the MIMIC-CXR and MIMIC-ABN datasets, ARIVE's clinical efficacy metrics outperform strong baseline models.
50%UnverifiedExpert blind evaluation results show that the reports generated by ARIVE are competitive with those of leading models.
50%UnverifiedARIVE's multimodal chain-of-thought trajectories received high scores for diagnostic traceability, with reviewers able to trace back along the reasoning chain to the sources of the report's conclusions.
50%UnverifiedA paired-review evaluation showed that when reviewers had access to MCoT, they more easily identified errors in generated reports.
50%UnverifiedCurrent evaluations of ARIVE remain concentrated on the MIMIC series of datasets, and the diversity and noise levels in real clinical settings may bring new challenges.
50%