Rad-Former: Structuring Radiology Reports using Transformers*

Abstract

Several professional societies have advocated for structured reporting in radiology, citing gains in quality, but some studies have shown that rigid templates and strict adherence may be too distracting and lead to incomplete reports. To gain the advantages of structured reporting while requiring minimal change to a radiologist's work-flow, the present work proposes a two-stage abstractive summarization approach that first finds the key findings in an unstructured report and then generates and organizes descriptions of each finding into a given template. The method uses a large manually annotated dataset and a taxonomy and other domain knowledge that were prepared in consultation with several practising radiologists. It can be used to structure reports dictated by radiologists and as post- and pre-processing steps for machine-learning pipelines. On the subtask of label extraction, the method achieves significantly better performance than previous rule-based approaches and learning-based approaches that were trained on automatically extracted labels. On the task of summarization, the method achieves more than 0.5 BLEU-4 score across 8 of the 10 most common labels and serves as a strong baseline for future experiments.