Report is a mixture of topics: Topic-guided radiology report generation

Area of Science:

• Artificial Intelligence
• Medical Imaging Informatics
• Natural Language Processing

Background:

• Radiologists face high workloads and diagnostic errors, necessitating automated radiology report generation (RRG).
• Existing RRG methods struggle to fully leverage the rich clinical information present in radiology reports.
• Radiology reports can be conceptualized as topic mixtures related to disease annotations.

Purpose of the Study:

• To develop a topic-guided framework to improve the quality of automatic radiology report generation.
• To incorporate expert knowledge, including topic graphs and n-grams, into the RRG process.
• To enhance RRG by conditioning word generation on probabilistically inferred radiographic topics.

Main Methods:

• Probabilistically inferring topics from radiographs.
• Integrating related topic graphs and n-grams as expert knowledge.
• Employing a bag-of-words planning mechanism as a novel encode-decode stream for guided report generation.

Main Results:

• The proposed topic-guided framework significantly outperforms previous state-of-the-art methods on the IU-Xray and MIMIC-CXR datasets.
• Experimental results validate the effectiveness of topic inference and expert knowledge integration in RRG.
• The method demonstrates superior performance in generating accurate and informative radiology reports.

Conclusions:

• The topic-guided RRG framework offers a promising approach to alleviate radiologist workload and reduce diagnostic errors.
• Incorporating topic knowledge and expert insights enhances the quality and reliability of automated radiology reports.
• This work introduces a novel probabilistic perspective on topic-based RRG.