Joint Extraction of Entities and Relations of Breast Ultrasound Reports Based on Deep Learning

摘要

The breast disease has become a common and prevalent disease. Ultrasound is a common way to diagnose breast diseases, so a large number of ultrasound textual reports have been generated which are valuable for breast disease. Recognition of entities and semantic relations through the ultrasound report is the premise to meet the demand of modern clinical system. The paper proposes a joint extraction method of the entity and relation based on text data of breast ultrasound reports, using practical ultrasound reports of patients with breast diseases from a 3A grade hospital in Shanghai, and focuses on an innovative point that is depending on deep learning to extract entities and relations from text data of ultrasound reports. This method uses deep learning to realize the structuralization of breast ultrasound reports to extract partial features of the test on the basis of collecting features from training data. Then, the relation between entities is classified. Finally, the extracted relation classification and the partial features of text are used to improve the accuracy of structuralization. The experiments show the accuracy of text structure increases by nearly 10% via the joint extraction of entity and relation method based on deep learning, comparing with the common structured approach relying on pipeline model, which reaches up to 80.3% in practical text data of ultrasound reports.