Experience-guided multi-agent interpretable framework for radiology report summarization.

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Medical Informatics

Background:

• Radiology report summarization is crucial for medical decision-making.
• Automating this task with large language models (LLMs) shows promise.
• Existing methods often neglect historical data and interpretability.

Purpose of the Study:

• To develop an advanced framework for radiology report summarization.
• To leverage historical reports for enhanced model performance.
• To improve the interpretability of automated summarization.

Main Methods:

• Proposed the Experience-guided Multi-Agent Interpretable (EMAI) framework.
• Introduced a self-evolutive nearest-neighbor explicit experience induction algorithm.
• Developed an interpretable findings analysis module and a collaborative multi-agent system.

Main Results:

• EMAI demonstrated improved accuracy and interpretability on MIMIC-CXR and Open-I datasets.
• Experience induction enhanced LLM generalization from diverse data.
• The interpretability mechanism increased the trustworthiness of generated summaries.

Conclusions:

• The EMAI framework effectively enhances radiology report summarization.
• Utilizing historical data and interpretability improves impression generation.
• This approach has the potential to significantly support medical decision-making.