Research On The Intelligent Reporting Technologies In Medical Imaging Information Systems

Image report as a permanent record of patients' image exam and diagnosis information in the medical imaging information system,has been widely used in the clinical routine and research.The image report is an important means of communication between the radiologist and the referring clinician.However,reading and writing image reports is a laborious and tedious work,especially when the amount goes large.Therefore,improving the efficiency of reading and writing image report is of great significance for reducing the workload of doctors and improving the quality of hospital services.The purpose of this study is to renew the report reading and writing functions in the medical imaging information system by automation derived from artificial intelligence,thereby improving the efficiency of doctors in processing image reports.Specifically,based on deep learning technology,three intelligent report technologies are studied in this paper,which are image report information extraction,image diagnosis conclusion generation,and structured report generation.These intelligent reporting technologies are modeled as sequence labeling,text summarization,and image segmentation task respectively,and solved by the newly proposed neural network model.The main innovative research work is as follows:For the reading of the image report,this paper proposed an information extraction method and a visualization method for medical imaging information systems.This method can extract,group,deliver and display the medical named entities in radiology information systems.The proposed neural network model-Multi-Embedding-BGRUCRF is based on a professional medical dictionary and multi-granularity embedding,which can encode the lexicon matching,radical decomposition,and context information.Due to the better represent of Chinese characters,our model achieved an F1 score of 95.88% in the key information extraction,which is 1.70% higher than the baseline model.The proposed rule-based entity grouping method got a 91.03% accuracy.This paper also studies the method of integrating information extraction into the radiology information system.The integration is realized by adding an information extraction server,a JSON object database,and a key information display interface.Through further experiments,we proved that the radiology information system integrated with the information extraction function can effectively reduce about 46% time consuming for report reading.For the writing of diagnosis conclusions in image reports,this paper proposed a method for automatically generating diagnosis conclusions.It can automatically generate diagnosis conclusions based on image findings and other data fields in the image report.The generated diagnosis conclusions can be used as a dynamically generated template for further editing by the doctor,thereby reducing the effort of edits and improving efficiency.The proposed neural network model-Info Fusion2 Seq based on information fusion encoding not only encodes the text of image findings but also encodes other short text fields and structured fields in the image report in their own ways.And all the encoded information are combined to decoding the outputs.The proposed methods can achieve a ROUGE-L value of 77.03% and an editing distance of 11.49,which is the best compared to the baseline models.Experimental results show that this method can reduce the editing efforts when doctors writing reports based on the generated template.For the generation of glaucoma screening reports,this paper proposes a refined image segmentation method,and based on this,further image measurements are made to generate glaucoma screening reports.The refined image segmentation method includes two main parts: contour transformation and SU-Net.Contour transformation is a newly proposed image coordinate transformation method,which can transform an image into a contour centered image according to a contour.SU-Net is a sequence labeling network for image segmentation inspired from natural language processing which directly predicts the segmentation boundary.The refined image segmentation method can be combined with any other segmentation method to form a two-step coarse-to-fine segmentation process,which can further suppress background noise and obtain a more balanced foreground-background ratio,thereby reducing segmentation error.For optic disc and cup segmentation,our method achieved the best results on two public datasets: MESSIDOR and Drishti-GS.We further tested the generation of glaucoma screening reports on Drishti-GS.The experimental results show that our method achieves the best Cup-to-Disc Ratio error(0.047)and AUC value(0.935).