A Transformer based approach for Abstractive Text Summarization of Radiology Reports

Abstract

Abstractive text summarization has emerged as a promising approach for generating concise and informative summaries from radiology reports. The topic of research focuses on developing a mechanism for abstractive text summarization specifically tailored for radiology reports, to generate informative summaries from the voluminous and complex information contained within reports. Manual summarization is time-consuming and prone to errors, while automated techniques can save time, reduce human bias, and improve the overall quality of the generated summaries. The mechanism involves preprocessing the text with NLP techniques, utilizing deep learning-based architectures and transformer models, and generating summaries that capture the essence of the original reports in a more concise form. Challenges such as handling complex medical terminologies, incorporating contextual information, and evaluating the quality of generated summaries are important considerations in this mechanism. The potential applications of radiology report summarization include improving report readability, facilitating decision-making, and enabling large-scale data analytics. In this research, Biobart-V2 model is used for summarization and Rouge-L value of 69.42% is being achieved whereas the dataset used is MIMIC III. Further research is needed to address the remaining challenges in this domain and integrate summarization into clinical practice for more effective and efficient radiology report interpretation and patient care.