Research On Big Data Deep Learning For Health Diagnosis And Treatment

Nowadays,there are a lot of visual data in the medical industry which will be of great significance to the diagnosis and evaluation of the curative effect.This thesis provides some theoretical basis and optimization suggestions for the detection of diabetic retinopathy through the application of deep learning algorithm based on retinal fundus images of diabetic patients.Convolutional neural network is one of the core algorithm of image processing in depth learning.So far,the study of convolution neural network has made great progress in many fields.But there are a few things to be solved completely.One of the points is that the number of different types of data sets(including the retinal fundus)is large and unbalanced,leading to the poor training effect of low proportion.Second,diabetic retinopathy detection is of high accuracy requirements,and the scale is relatively large which need more model parameters.In the traditional training time,the gradient difference from the back layer to the first layer is almost zero by the back-propagation algorithm.Third,for the huge amount of data sets,the traditional training method is difficult to adapt.In order to solve the kinds of problems above,following works have been done in this thesis:1.In order to improve the generalization ability of the model and the better training performance on imbalanced data sets,this thesis proposes an adaptive resampling method(Adaptive),and designs a hierarchical model to improve the efficiency of training model.The hierarchical model consists of an improved model of AlexNet and VGGNet with the number of convolution layers of 8,11 and 13,respectively.The experiments show that the classification model of evaluation index,classification accuracy?kappa coefficient and training speed have been greatly improved through adaptive over-resampling.The kappa coefficient(model consistency test)has reached 0.80.2.In order to improve the detection ability of diabetic retinopathy,a mean square feature fusion algorithm based on two-dimensional image is proposed.This thesis designs a full convolution network structure based on Residual Network,which can reduce a large number of parameters compared to the whole connection layer,and build a deeper level in the case of avoiding the gradient disappear.In this thesis,the features of the last layer of the model are extracted by the trained model,and then the extracted features are stacked and fused in two batches,and the mean value is taken as the input data of the two detection.Through the training and testing of diabetic retinal fundus dataset provided on the Kaggle big data platform,the kappa coefficient is increased by four percentage points on the previous basis.3.In order to solve the problem of deep learning under bigdata,this thesis achieves a distributed depth learning training system,based on the framework of big data computing.This system can be used in multi-CPU and GPU environment.Finally,based on the analysis of the parallel training and experiment,the results show that the training speed can be improved effectively and the classification accuracy can be guaranteed.