# ORGAN: Observation-Guided Radiology Report Generation via Tree Reasoning

**Authors:** Wenjun Hou, Kaishuai Xu, Yi Cheng, Wenjie Li, Jiang Liu  
**arXiv:** [2306.06466](https://arxiv.org/abs/2306.06466)  
**Date:** 2023-06-10  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.06466v1/

## Key points

- Radiology report generation is a challenging task that requires consistency between images and detailed reports.
- Previous research focused on planning-based methods but lacked essential information for accurate report generation.
- The Observation-guided Radiology Report Generation framework (ORGAN) generates an observation plan and utilizes it with radiographs for report generation, incorporating an observation graph and tree reasoning mechanism.
- ORGAN outperforms previous methods in text quality and clinical efficacy according to experimental results.
- Automated radiology report generation can help alleviate the burden on healthcare professionals by providing textual descriptions for radiographs.
- Maintaining consistency between images and textual reports is crucial in automated radiology report generation.
- ORGAN leverages image information with textual plans and advanced reasoning mechanisms to produce accurate and informative reports for medical practitioners.

## Summary

, , , , 

This paper delves into the task of radiology report generation, focusing on generating descriptive free-text reports for a set of radiographs. The primary challenge lies in maintaining consistency between the images and the detailed reports. Previous research has attempted to address this issue through planning-based methods that generate reports based on high-level plans. However, these plans often lack essential information such as observation characteristics and preliminary clinical diagnoses, leading to a need for stronger reasoning during the report generation process. To tackle this problem, the proposed Observation-guided Radiology Report Generation framework (ORGAN) first generates an observation plan and then utilizes both the plan and radiographs for report generation. This framework incorporates an observation graph and a tree reasoning mechanism to enhance the plan information by capturing various formats of each observation. Experimental results demonstrate that ORGAN outperforms previous state-of-the-art methods in terms of text quality and clinical efficacy. play a crucial role in  by providing textual descriptions for radiographs. However, interpreting radiographs can be time-consuming even for experienced radiologists, with an average of 5-10 minutes per image. Automated  has emerged as a potential solution to alleviate the burden on healthcare professionals, garnering increasing research attention. The paper highlights the importance of maintaining consistency between images and textual reports in automated . While previous approaches have made progress by identifying major observations from radiographs before generating word-level realizations, they still face challenges in ensuring comprehensive coverage across different modalities. In conclusion, this study introduces ORGAN as a novel framework that leverages  to enhance the quality and clinical relevance of generated radiology reports. By integrating image information with textual plans and employing advanced reasoning mechanisms, ORGAN demonstrates superior performance compared to existing methods in producing accurate and informative reports for medical practitioners.