Background Radiology reports are a permanent record of patient's health information often used in clinical practice and research. Reading radiology reports is common for clinicians and radiologists. However, it is laborious and time-consuming when the amount of reports to be read is large. Assisting clinicians to locate and assimilate the key information of reports is of great significance for improving the efficiency of reading reports. There are few studies on information extraction from Chinese medical texts and its application in radiology information systems (RIS) for efficiency improvement. Objectives The purpose of this study was to explore methods for extracting, grouping, ranking, delivering, and displaying medical-named entities in radiology reports which can yield efficiency improvement in RISs. Methods A total of 5,000 reports were obtained from two medical institutions for this study. We proposed a neural network model called Multi-Embedding-BGRU-CRF (bidirectional gated recurrent unit-conditional random field) for medical-named entity recognition and rule-based methods for entity grouping and ranking. Furthermore, a methodology for delivering and displaying entities in RISs was presented. Results The proposed neural named entity recognition model has achieved a good F1 score of 95.88%. Entity ranking achieved a very high accuracy of 99.23%. The weakness of the system is the entity grouping approach which yield accuracy of 91.03%. The effectiveness of the overall solution was proved by an evaluation task performed by two clinicians based on the setup of actual clinical practice. Conclusions The neural model shows great potential in extracting medical-named entities from radiology reports, especially for languages, that lack lexicons and natural language processing tools. The pipeline of extracting, grouping, ranking, delivering, and displaying medical-named entities could be a feasible solution to enhance RIS functionality by information extraction. The integration of information extraction and RIS has been demonstrated to be effective in improving the efficiency of reading radiology reports.

Background: the automated identification of mentions of ontological concepts in natural language texts is a central task in biomedical information extraction. Despite more than a decade of effort, performance in this task remains below the level necessary for many applications.Results: recently, applications of deep learning in natural language processing have demonstrated striking improvements over previously state-of-the-art performance in many related natural language processing tasks. Here we demonstrate similarly striking performance improvements in recognizing biomedical ontology concepts in full text journal articles using deep learning techniques originally developed for machine translation. For example, our best performing system improves the performance of the previous state-of-the-art in recognizing terms in the Gene Ontology Biological Process hierarchy, from a previous best F1 score of 0.40 to an F1 of 0.70, nearly halving the error rate. Nearly all other ontologies show similar performance improvements.Conclusions:A two-stage concept recognition system, which is a conditional random field model for span detection followed by a deep neural sequence model for normalization, improves the state-of-the-art performance for biomedical concept recognition. Treating the biomedical concept normalization task as a sequence-tosequence mapping task similar to neural machine translation improves performance.

We compare the use of LSTM-based and CNN-based character-level word embeddings in BiLSTM-CRF models to approach chemical and disease named entity recognition (NER) tasks. Empirical results over the BioCreative V CDR corpus show that the use of either type of character-level word embeddings in conjunction with the BiLSTM-CRF models leads to comparable state-of-theart performance. However, the models using CNN-based character-level word embeddings have a computational performance advantage, increasing training time over word-based models by 25% while the LSTM-based characterlevel word embeddings more than double the required training time.