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

Background and Objective: Radiology report summarization, which involves generating concise impressions from detailed findings, is a critical task in medical decision-making. Recent advancements in large language models have shown promise in automating this task, yet existing methods often overlook the importance of leveraging historical radiology reports for experience induction and the interpretability of the prediction process. Methods: To address these challenges, we propose an Experience-guided Multi-Agent Interpretable framework (EMAI) for radiology report summarization. Our framework introduces a novel self-evolutive nearest-neighbor explicit experience induction algorithm to automatically extract and generalize knowledge from historical reports, enhancing the robustness and generalizability of the model. Additionally, we design an interpretable findings analysis module that deconstructs and explains the relationships between findings and impressions, providing human-comprehensible insights into the impression generation process. To dynamically integrate these components, we develop a collaborative multi-agent framework that adaptively determines when and how to leverage experiences or relevant reports for impression generation. Results: Experimental results on two public datasets, MIMIC-CXR and Open-I, demonstrate that our framework improves the accuracy and interpretability of radiology report summarization. The induction learning process enables LLMs to generalize better from diverse case data, while the interpretability mechanism provides valuable explanations that enhance the trustworthiness of the generated results. Conclusions: The EMAI framework effectively improves radiology report summarization by utilizing historical data and enhancing interpretability. It has the potential to better support medical decision-making through more accurate and trustworthy impression generation.