TL;DR
ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence) is a new framework that makes AI-generated radiology reports explainable by mimicking a radiologist's workflow: iteratively inspecting specific anatomical regions and interpreting them before writing the final report. By grounding diagnostic reasoning in localized visual evidence, ARIVE achieves state-of-the-art clinical accuracy while providing a "traceable" path that helps clinicians identify and correct errors.
What Problem Does This Paper Address?
Automated radiology report generation has made huge strides, but it suffers from a "black box" problem. Current models often produce "clinically plausible" reports that sound professional but may contain hallucinations—findings not actually supported by the image.
Even when models use Chain-of-Thought (CoT) reasoning to explain their steps, these explanations are often purely textual. As the reasoning sequence grows longer, the model can suffer from "visual forgetting," where it relies more on its own generated text than the original image. This lack of transparency makes it difficult for radiologists to trust, review, or revise AI-generated drafts.
Methods
The authors introduce ARIVE, which treats explainability as a core requirement rather than an afterthought. The framework's workflow is illustrated in Figure 1:
Key technical components include:
- Interleaved Multimodal Chain-of-Thought (MCoT): Unlike standard CoT, ARIVE's reasoning process is interleaved with visual inspection. The model first surveys the image globally, then iteratively requests and "looks" at specific anatomical crops (e.g., the right hilar structures or cardiac silhouette) to inform its next reasoning step.
- Anatomical and Clinical-Content Optimization (ACCO): To move beyond simple token-level training, the authors use Group Relative Policy Optimization (GRPO). This reinforcement learning stage optimizes the model using four distinct rewards:
- Anatomical Grounding ($R_{BBOX}$): Accuracy of region selection and localization.
- Reasoning Accuracy ($R^{rea}_{CXB}$): Clinical correctness of the intermediate reasoning steps.
- Report Accuracy ($R^{ans}_{CXB}$): Clinical correctness of the final report.
- Entity & Relation Preservation ($R_{RG}$): Agreement with the reference report's clinical entities and their relationships.
- Architecture: ARIVE uses a 4.85 billion parameter backbone (Qwen3-VL-Instruct-4B) augmented with a specialized anatomical visual module to process local image crops.
Datasets
- MIMIC-CXR: The primary benchmark for chest X-ray report generation.
- MIMIC-ABN: A specialized subset containing only abnormal studies, providing a focused assessment of the model's ability to report significant findings.
- Chest ImaGenome: Used to provide the anatomical bounding box annotations for 29 regions.
- MCoT Dataset: A newly constructed dataset of 159,416 multimodal trajectories derived from MIMIC-CXR reports using radiologist-style templates.
Key Insights and Ideas
- "Looking" Prevents "Forgetting": The most critical insight is that interleaving regional inspection with text generation keeps the model anchored to the visual evidence. This significantly reduces the propagation of unsupported assumptions.
- Traceability is a Safety Feature: Because every claim in the final report can be traced back to a specific anatomical inspection step, clinicians can verify the AI's work. As shown in Figure 6, an unsupported claim of "mild cardiomegaly" can be traced back to the exact moment the model misinterpreted the cardiac silhouette.
- Grounding $\neq$ Accuracy: The authors found that better anatomical grounding alone doesn't guarantee a better report. Clinical efficacy and spatial grounding must be optimized jointly to ensure the model isn't just looking at the right place, but also seeing the right thing.
Key Results
ARIVE demonstrates that you don't need a massive 400B parameter model to achieve expert-level results if your reasoning process is sound.
- Clinical Efficacy: ARIVE achieved the highest scores across seven clinical efficacy measures on MIMIC-CXR, outperforming much larger models like Qwen3.5-397B and MedGemma-1.5.
- Expert Evaluation: In blinded grading by an experienced radiologist, ARIVE's reports were rated as competitive with leading models and superior to task-specific baselines like PriorRG.
- Diagnostic Traceability: 95.8% of report claims were fully supported by the generated MCoT trajectories.
- Human-AI Collaboration: In a paired review study, radiologists identified 37.5% more errors in generated reports when they had access to the MCoT trajectories compared to reviewing the report and image alone (Figure 4d).
Limitations and Future Work
While promising, the authors acknowledge several areas for improvement:
- Sparse Supervision: Some anatomical regions (like the lung apices) had very few training examples, leading to weaker grounding in those areas.
- Modality Constraints: The current work is limited to frontal chest X-rays. Expanding to lateral views, CT, or MRI will require new trajectory designs and spatial supervision.
- Fixed Regions: The model currently relies on 29 predefined anatomical boxes. Future iterations could benefit from learning to identify diagnostically relevant regions dynamically without relying on fixed bounding-box annotations.
Summary updated