Predicting Doctor's Impression For Radiology Reports with Abstractive Text Summarization

Predicting doctor's impression (summarization) for radiology reports saves doctors and patients tremendous time from manually digging through the reports. But there are few pre-trained language models for summarization, especially for radiology datasets. We solve abstractive summarization for the free-text radiology reports in the MIMIC-CXR dataset by building ClinicalBioBERTSum, which incorporates domain-specific BERT-based models into the state-of-the-art BERTSum architecture. We give a well-rounded evaluation of our model performance utilizing both word-matching based metrics and semantic based metrics. Our best-performing model obtains a ROUGE-L F1 score of 57.37/100 and a ClinicalBioBERTScore of 0.55/1.00. With comprehensive experiments, we showcase that domain-specific pre-trained and fine-tuned encoders and sentence-aware embeddings could significantly boost the performance of abstractive summarization for radiology reports. Our work also provides a set of pre-trained transformer weights that could further facilitate practitioner's future research with radiology reports.