Research On Key Technologies Of Medical Assisted Diagnosis Based On Deep Learning

With the rapid development of deep learning technology and its extensive application in the medical field,smart medical care is playing an increasingly important role in our lives,and a new technological revolution has gradually formed.However,at present,China has an average of 57 million misdiagnosed medical records each year,with a misdiagnosis rate as high as 27.8%.Secondly,the shortage of doctor resources in our country is also very serious.Each person has an average of only 0.0021 doctors.Especially in remote areas,medical resources are evenly distributed.Not only are medical devices and daily medicines scarce,doctors and experts with clinical experience are even It is one of the few.If computers can be used to automatically assist in the diagnosis of disease symptoms in patients,patients in remote areas can enjoy expert-level diagnosis and treatment,which will greatly alleviate the shortage of medical resources and other problems.Therefore,this thesis studies the key technologies of medical-assisted diagnosis based on deep learning to improve the accuracy of computer-aided diagnosis.The medical assisted diagnosis technology studied in this article is completed through a medical assisted diagnosis model.Patients can obtain the probability of their own disease according to the main complaint,which can effectively assist doctors to judge the patient’s main complaint information.This article uses deep learning technology to build a medical assistant diagnostic model and analyze patient complaints,that is,medical text classification.However,traditional deep learning algorithms have the following shortcomings: 1)Although the bidirectional threshold recurrent network model has good training effects,it is good for sentences.The interpretability is not enough;2)TextCNN algorithm can effectively classify text,but there are problems of overfitting and loss of feature information.In response to the above problems,this thesis proposes the following three models:First,a medical assistant diagnosis model based on a combination of a two-way threshold recurrent network and a self-attention mechanism is proposed.The bidirectional threshold recurrent network model(BiGRU)has better performance and fewer parameters to converge faster,but it lacks interpretability of sentences,and the self-attention mechanism can obtain more information about disease symptoms that need attention In order to suppress other information not related to symptoms and enhance the interpretability of sentences,the self-attention mechanism is combined with a two-way threshold recurrent network,and a BiGRU-SA model is proposed to classify medical text data to improve the accuracy of diagnosis.Second,a medical assistant diagnosis model based on F-TextCNN is proposed.The traditional TextCNN model is implemented using convolutional neural networks for text classification.Its biggest feature is the simple network mechanism,which has the advantages of fewer parameters,less calculation,and fast training.However,the pooling layer of the model uses the largest pool.The pooling operation is performed by the chemical method,which will cause the loss of characteristic information of disease symptoms.This thesis proposes to improve the maximum pooling method in the pooling layer to a form in which all features are connected,which can effectively improve the training speed and accuracy of the model,and significantly improve the classification effect of medical text data.Third,a hybrid model of medical assistant diagnosis based on the combination of BiGRU-SA and F-TextCNN is proposed.In this model,the output results of the two models are fused,and the weighted average method is used to obtain the final result.It can effectively balance the negative effects brought by the two models and keep the accuracy,recall and F1-evaluation values at a relatively balanced and reasonable level.In the end,this article uses the three indicators of accuracy,recall,and F1-evaluation as the evaluation criteria,and designs corresponding experiments for the theoretical part.The experimental analysis shows that the improvement of the model in this thesis is reasonable and effective.