# Can Prompt Learning Benefit Radiology Report Generation?

**Authors:** Jun Wang, Lixing Zhu, Abhir Bhalerao, Yulan He  
**arXiv:** [2308.16269](https://arxiv.org/abs/2308.16269)  
**Date:** 2023-08-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.16269v1/

## Key points

- Radiology report generation is a challenging task that requires prior medical knowledge
- The authors propose a method called PromptRRG that leverages prompt learning to activate a pretrained model and incorporate prior medical knowledge
- Different prompt designs are investigated, including common prompts, domain-specific prompts, and disease-enriched prompts
- An automatic prompt learning mechanism is proposed to alleviate the burden of manual prompt engineering
- The proposed method is evaluated on MIMIC-CXR dataset and achieves state-of-the-art performance in generating high quality radiology reports
- The approach utilizes prompt learning and incorporates prior medical knowledge to improve the effectiveness and efficiency of radiology report generation
- The method has the potential to enhance automated radiology reporting systems and improve patient care in clinical settings.

## Summary

In the field of radiology, generating accurate and clinically meaningful descriptions of radiology images such as MRI and X-ray is a challenging task that requires prior medical knowledge. While there have been significant advancements in natural scene image captioning tasks, radiology report generation remains a complex problem. To address this issue, the authors propose a method called PromptRRG that leverages prompt learning to activate a pretrained model and incorporate prior medical knowledge. Prompt learning for radiology report generation has not been extensively explored before, so the authors begin by investigating different prompt designs. They categorize these prompts based on varying levels of knowledge: common prompts that capture general information, domain-specific prompts that focus on specialized medical knowledge, and disease-enriched prompts that provide detailed insights into specific conditions. To alleviate the burden of manual prompt engineering, the authors also propose an automatic prompt learning mechanism. This approach enables the model to learn prompts automatically without requiring explicit input from human experts. By incorporating prompt learning into their method, the authors aim to improve the effectiveness and efficiency of radiology report generation. The proposed method is evaluated on MIMIC-CXR which is the largest benchmark dataset for radiology report generation. The experimental results demonstrate that PromptRRG achieves state-of-the-art performance in generating high quality radiology reports. Overall this paper presents a novel approach to radiology report generation by utilizing prompt learning and incorporating prior medical knowledge. It systematically examines the effectiveness of prompt learning for this task and provides empirical evidence of its benefits. The proposed method has the potential to enhance automated radiology reporting systems and improve patient care in clinical settings.