Disease-Assisted Diagnosis Research And System Design Based On Deep Learning

Intelligent assisted diagnosis of disease not only relieves the shortage of professional physicians,and improves the efficiency of medical treatment,but also promotes the overall medical level of society.The development of deep learning techniques in medically assisted diagnosis has attracted widespread attention.Diabetic retinopathy(DR)is one of the early diseases that use actively deep learning techniques for intelligent diagnosis.At present,major companies are exploring more accurate DR intelligent diagnosis algorithms that can be applied to the clinic to improve the recognition accuracy,which is essential for curbing retinopathy and preventing blindness.Compared with the DR dataset,the autism spectrum disorder(ASD)brain imaging data is small due to factors such as difficulty in acquisition.It has great challenge to use deep learning techniques for predicting ASD and finding objective biomarkers on small sample datasets.For remote areas with many patients and insufficient medical resources,the establishment of a disease assisted diagnosis system can make full use of accurate auxiliary diagnostic models to provide effective diagnosis.Firstly,the DR classification algorithm and ASD prediction algorithm based on deep learning are studied in this paper.Then a cross-platform DR auxiliary detection system is built based on the DR diagnosis model.The specific research contents are as follows:(1)In view of the problems of training memory and time cost caused by the deepening of convolutional neural networks(CNN),this paper proposes the use of Inception-ResNet-v2 network model based on the migration learning method for DR classification.The Inception-ResNet-v2 network combines the Inception module with the ResNet residual network for stronger feature representation.The Inception-ResNet-v2 pre-training model is used to hierarchically fine-tune the network parameters on the retinal images for DR classification.The accuracy of the binary-classification and five-classification is 91.3%and 70.5%accuracy,respectively,which is higher than the accuracy of the Inception-v3 network.(2)For the small sample dataset of structure magnetic resonance imaging,this paper makes a preliminary exploration of ASD prediction from the aspects of morphological features and texture features by using deep learning method.For the brain volume and thickness morphological features,the five-layer multilayer perceptron(MLP)network and the one-dimensional CNN are used to predict ASD.For texture features,hierarchical extraction and two-dimensional tiling are used to reduce the dimensionality of the three-dimensional image,and the neural network based on CNN and MLP fusion is proposed for ASD prediction.The recognition rate of two-dimensional tiling method is 9.8%higher than that of stratified extraction,which indicates the effectiveness of neural network using CNN and MLP fusion based on two-dimensional tiling for ASD prediction.(3)In order to solve the shortage of medical resources in remote areas,the cross-platform DR assisted detection system is designed using C/S architecture.The system is designed and implemented in the LAN.The architecture of client function development uses the MVVM mode.The server uses the HTTPServer port in the Python class library for monitoring.The disease assisted diagnostie model i5 invoked through the common gateway interface.This system implements the basic functions of cross-platform DR assisted diagnosis and can be extended to ASD prediction.