# RECAP: Towards Precise Radiology Report Generation via Dynamic Disease Progression Reasoning

**Authors:** Wenjun Hou, Yi Cheng, Kaishuai Xu, Wenjie Li, Jiang Liu  
**arXiv:** [2310.13864](https://arxiv.org/abs/2310.13864)  
**Date:** 2023-10-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2310.13864v1/

## Key points

- Growing interest in automating radiology report generation to alleviate radiologists' workloads
- Previous research focused on concise observations but lacked precise attributes for disease severity determination
- Historical records with temporal information are crucial for accurate diagnosis and treatment planning
- Introduction of novel approach called RECAP for generating precise and accurate reports through dynamic disease progression reasoning
- RECAP incorporates historical records and spatiotemporal information, utilizing a disease progression graph and dynamic reasoning mechanism
- Contributions include capturing spatial and temporal information, using fine-grained attributes for improved accuracy
- Limitations may include reliance on predefined observations and progressions for training, as well as potential challenges in certain scenarios due to the sequential nature of the framework
- Despite limitations, RECAP shows promise in improving clinical factuality and diagnostic accuracy in radiology practice

## Summary

In the field of radiology report generation, there has been a growing interest in automating the process to alleviate radiologists' workloads. Previous research has focused on producing concise observations but neglects precise attributes for determining disease severity. This can lead to imprecise reports and highlights the need for stronger attribute modeling. Historical records containing temporal information are crucial for accurate diagnosis and treatment planning but are often overlooked in evaluation. To address these issues, a novel approach called RECAP has been proposed. It aims to generate precise and accurate reports through dynamic disease progression reasoning by incorporating historical records and spatiotemporal information. A disease progression graph captures prior and current observations, progressions, and precise attributes while a dynamic reasoning mechanism effectively selects relevant attributes. The contributions include capturing spatial and temporal information and utilizing fine-grained attributes for improved accuracy. However, limitations such as reliance on predefined observations and progressions for training may exist, along with potential challenges in certain scenarios due to the sequential nature of the framework. Despite this, RECAP shows promise in improving clinical factuality and diagnostic accuracy in radiology practice.