Automatically Generate Impression From Findings In Radiology Report Using BERT

The vital work of summarizing medical reports so that the general public can easily access them can be considerably aided by the recent developments in deep learning and large language models (LLM). One medical specialty that can make use of it is radiology, where it's essential to summarize the most crucial information from the “findings” part into the “impression” part of a report to help radiologists and referring physicians communicate more easily. To automate the development of the “impression” part in radiology reports, we propose a unique architecture for summarization. The workload for radiologists is greatly reduced by this method. Because of the specialized vocabulary employed in radiology, existing models and approaches for abstractive summarization-such as those that utilize recent advances in the field of natural language processing (NLP), such as BERT and its derivatives, are not immediately relevant. As such, our system is specially designed to tackle the particular difficulties presented by radiology reports. We create a pre-trained language model in radiology to address the issue of automatically summarizing reports on radiography. More specifically, In this research BERT Encoder is pre-trained on the MIMIC-CXR dataset. Next, an abstractive summarization model is built by combining a Transformer decoder and a pretrained BERT. In conclusion, for the abstractive summarization task, the model will be improved utilizing reports from chest X-rays. Furthermore, the publically available OPEN-I datasets is used for evaluation.