Reason like a radiologist: Chain-of-thought and reinforcement learning for verifiable report generation

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Natural Language Processing

Background:

• Current radiology report generation models lack expert-like reasoning and anatomical grounding, limiting clinical trust and explainability.
• Generating structured, verifiable, and explainable radiology reports is crucial for clinical efficiency and decision-making.

Purpose of the Study:

• To introduce BoxMed-RL, a unified training framework for generating spatially verifiable and explainable chest X-ray reports.
• To improve the clinical trust and explainability of automated radiology reports.

Main Methods:

• A two-phase training framework: Pretraining Phase (medical concept learning, reinforcement learning for spatial grounding) and Downstream Adapter Phase (lightweight adapter for fluency and credibility).
• Utilized reinforcement learning to enforce spatial grounding of findings in anatomical evidence.
• Employed a lightweight adapter to ensure clinical credibility and fluency of generated reports.

Main Results:

• Achieved an average 7% improvement in METEOR and ROUGE-L metrics on MIMIC-CXR and IU X-Ray benchmarks compared to state-of-the-art methods.
• Demonstrated an average 5% improvement in large language model-based metrics, indicating robust generation of high-quality reports.
• Generated reports are spatially verifiable and enhance explainability.

Conclusions:

• BoxMed-RL offers a novel approach to chest X-ray report generation, enhancing both accuracy and clinical utility.
• The framework successfully integrates reasoning, spatial grounding, and clinical credibility for trustworthy AI in radiology.
• Publicly available code and templates facilitate further research and application of BoxMed-RL.