Research On Structuring Method Of Medical Imaging Report

As part of the electronic health record,the medical imaging report contains a wealth of patient health records.However,most of the imaging reports are free text,and the quality of the content depends on the doctor’s subjective judgment and experience.The structuring of medical texts requires a combination of medical knowledge,artificial intelligence,and natural language technologies,which poses a huge challenge to data mining.We investigated the trends in structured text study,and compared the language characteristics of Chinese and English.We proposed a text structuring processes:named entity recognition,entity relationship extraction and results visualization.This method is effective in MRI reports of nasopharyngeal carcinoma.The main work of this thesis is as follows:Firstly,we discussed with clinicians who have professional medical knowledge,summarized the description characteristics of MRI report of nasopharyngeal carcinoma,and formulated a set of structured templates suitable for image reports.According to the template,we can extract medical entities and relationships from the image reports,and then annotate the MRI reports to form a training set and test set.Secondly,we proposed a network architecture that incorporates word structure features,which has excellent performance in words semantic coding.Our experiments show that the architecture surpasses benchmark models in the task of named entity recognition.And the accuracy has reached more than 97%.Thirdly,we extracted five features for relation classification tasks.We compared multiple machine learning classifier through experiments and chosen the optimal model as the relation extraction model.What’s more,we input the entity vector obtained by the semantic coding architecture into multilayer perceptron.Experimental results show that entity vectors are effective in relation classification,and are better than manually extracted features.Finally,we applied the model to the structuring of the MRI report and derived a knowledge tree of nasopharyngeal cancer.A satisfactory conclusion is that the image reports structuring method is effective.The extraction rate and accuracy are 84.74%,89.3 9%respectively.In summary,based on the existing structured technology,we proposed a semantic coding network with a self-attention mechanism that incorporates the structural features of words.This architecture can capture the association between any words in a sentence.It provides a feasible method for clinical decision support,data management and database construction.