BI-RADS BERT & Using Section Segmentation to Understand Radiology Reports

TL;DR

This work pre-trained a contextual embedding BERT model using breast radiology reports and developed a classifier that incorporated the embedding with auxiliary global textual features in order to perform section segmentation, which resulted in an overall accuracy increase of 17% in the field extraction tasks.

Abstract

Radiology reports are one of the main forms of communication between radiologists and other clinicians and contain important information for patient care. In order to use this information for research and automated patient care programs, it is necessary to convert the raw text into structured data suitable for analysis. State-of-the-art natural language processing (NLP) domain-specific contextual word embeddings have been shown to achieve impressive accuracy for these tasks in medicine, but have yet to be utilized for section structure segmentation. In this work, we pre-trained a contextual embedding BERT model using breast radiology reports and developed a classifier that incorporated the embedding with auxiliary global textual features in order to perform section segmentation. This model achieved a 98% accuracy at segregating free text reports sentence by sentence into sections of information outlined in the Breast Imaging Reporting and Data System (BI-RADS) lexicon, a significant improvement over the Classic BERT model without auxiliary information. We then evaluated whether using section segmentation improved the downstream extraction of clinically relevant information such as modality/procedure, previous cancer, menopausal status, the purpose of the exam, breast density, and breast MRI background parenchymal enhancement. Using the BERT model pre-trained on breast radiology reports combined with section segmentation resulted in an overall accuracy of 95.9% in the field extraction tasks. This is a 17% improvement compared to an overall accuracy of 78.9% for field extraction with models using Classic BERT embeddings and not using section segmentation. Our work shows the strength of using BERT in radiology report analysis and the advantages of section segmentation in identifying key features of patient factors recorded in breast radiology reports.

Figures (3)

• Figure 1: This project aims to improve health indicator field extraction tasks by using section segmentation to narrow down the free text report length. When a Radiologist reads a report, they can divide the report into sections useful for finding specific information. With a classic BERT framework the report is fed into the model without narrowing the report into sections, resulting in some confusion as to where the information is located. With a BI-RADS BERT model used to segment sections before field extraction, we achieve a higher performance.
• Figure 2: Visual representation of the model architectures used for classification. A) Text sequence classifier: This model takes a contextual embedding of the input text using a BERT architecture and then feeds the embedding into a fully connected linear layer to output a classification. B) Text sequence classifier with auxiliary data: This model uses a auxiliary feature encoder to build an encoded auxiliary data vector that is concatenated with the contextual embedding to use for classification. C) Auxiliary data encoder architecture: This encoder architecture include 3 fully connected layers followed by a Tanh activation function.
• Figure 3: Histograms of the labels each field being extracted from breast radiology reports fine tuning dataset. We can see that each task suffers from a dominating label that make G.F1 better at quantifying performance over accuracy.