Making hierarchically aware decisions on short findings for automatic summarisation

Abstract

An impression in a typical radiology report emphasises critical information by providing a conclusion and reasoning based on the findings. However, the findings and impression sections of these reports generally contain brief texts, as they highlight crucial observations derived from the clinical radiograph. In this scenario, abstractive summarisation models often experience a degradation in performance when generating short impressions. To address this challenge in the summarisation task, our work proposes a method that combines well-known fine-tuned text classification and abstractive summarisation language models. Since fine-tuning a language model requires an extensive, well-defined training dataset and is a time-consuming task dependent on high GPU resources, we employ prompt engineering, which uses prompt templates to programme language models and improve their performance. Our method first predicts whether the given findings text is normal or abnormal by leveraging a fine-tuned language model. Then, we apply a radiology-specific BART model to generate the summary for abnormal findings. In the zero-shot setting, our method achieves remarkable results compared to existing approaches on a real-world dataset. In particular, our method achieves scores of 37.43 for ROUGE-1, 21.72 for ROUGE-2, and 35.52 for ROUGE-L.