Graph Enhanced Contrastive Learning for Radiology Findings Summarization

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4677 - 4688 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Graph Enhanced Contrastive Learning for Radiology Findings Summarization Jinpeng Hu♠♡∗, Zhuo Li♠♡∗, Zhihong Chen♠♡, Zhen Li♠♡ Xiang Wan♡♢†, Tsung-Hui Chang♠♡† ♠The Chinese University of Hong Kong (Shenzhen) ♡Shenzhen Research Institute of Big Data ♢Pazhou Lab, Guangzhou, 510330, China ♠{jinpenghu, zhuoli3, zhihongchen}@link.cuhk.edu.cn ♡wanxiang@sribd.cn ♠{lizhen, changtsunghui}@cuhk.edu.cn Abstract The impression section of a radiology report summarizes the most prominent observation from the findings section and is the most im- portant section for radiologists to communicate to physicians. Summarizing findings is time- consuming and can be prone to error for inex- perienced radiologists, and thus automatic im- pression generation has attracted substantial at- tention. With the encoder-decoder framework, most previous studies explore incorporating ex- tra knowledge (e.g., static pre-defined clinical ontologies or extra background information). Yet, they encode such knowledge by a sepa- rate encoder to treat it as an extra input to their models, which is limited in leveraging their re- lations with the original findings. To address the limitation, we propose a unified framework for exploiting both extra knowledge and the original findings in an integrated way so that the critical information (i.e., key words and their relations) can be extracted in an appro- priate way to facilitate impression generation. In detail, for each input findings, it is encoded by a text encoder, and a graph is constructed through its entities and dependency tree. Then, a graph encoder (e.g., graph neural networks (GNNs)) is adopted to model relation informa- tion in the constructed graph. Finally, to empha- size the key words in the findings, contrastive learning is introduced to map positive samples (constructed by masking non-key words) closer and push apart negative ones (constructed by masking key words). The experimental results on OpenI and MIMIC-CXR confirm the effec- tiveness of our proposed method.1 1 Introduction Radiology reports document critical observation in a radiology study and play a vital role in commu- nication between radiologists and physicians. A *Equal Contribution. †Corresponding author. 1Our code is released at https://github.com/ jinpeng01/AIG_CL. Figure 1: An example of the findings and corresponding impression, where the relation information, as well as positive and negative examples, are also shown in the figure. Note that △ represents the removed word. radiology report usually consists of a findings sec- tion describing the details of medical observation and an impression section summarizing the most prominent observation. The impression is the most critical part of a radiology report, but the process of summarizing findings is normally time-consuming and could be prone to errors for inexperienced radi- ologists. Therefore, automatic impression genera- tion (AIG) has drawn substantial attention in recent years, and there are many methods proposed in this area (Zhang et al., 2018; Gharebagh et al., 2020; MacAvaney et al., 2019; Shieh et al., 2019). Most existing studies focus on incorporating extra knowledge on the general encoder-decoder framework. For example, Zhang et al. (2018) uti- …

Content truncated to 500 words