Diagnosis Of Liver Diseases Based On Deep Learning

Liver cirrhosis and other related liver diseases are increasingly threatening human health,so the diagnosis of liver diseases is of great significance.However,the medical liver images collected in clinic have some problems,such as uneven echoes and blurred lesions.It makes it easy for doctors to have certain errors in subjective diagnosis.According to these issues,computer aided diagnosis technology is used.This paper firstly uses MB-LBP and Gabor as feature fusion in the field of traditional machine learning.Then I combines learning vector neural network LVQ.Finally it achieves a better classification of liver diseases.But there are many disadvantages in machine learning methods.It is time-consuming and labor-consuming to extract features manually.In view of the above problems,this paper adopts the method of deep learning,which can automatically classify learning features,greatly improving the efficiency of operation and generalization ability.In this paper,two recognition methods are proposed in the field of deep learning:(1)Firstly,a lightweight convolution neural network called SqueezeNet is used in this paper.Traditional convolutional neural networks have many layers and large structural parameters.Compared with it,SqueezeNet has not only fewer parameters but also more convenient and efficient training.This model is more suitable for in-depth learning applications of mobile or embedded devices.In this paper,SqueezeNet and support vector machine are used to recognize and classify liver cirrhosis.(2)A new model structure GoogleNet-PNN is proposed for the first time.It combines the advantages of GoogleNet convolution neural network in learning image features efficiently and accurately,and PNN probabilistic neural network in training easily and convergence quickly.Then,Particle swarm optimization is used to optimize the structure.Finally it achieves good results in the experiment of liver disease classification.Because the in-depth learning network model needs a certain sample set to train parameters to optimize,and the number of sample sets of medical images is far less than that of natural image sets.If the training is not enough,it is easy to produce over-fitting phenomena and affect the performance of the final model.To address these problems,migration learning and data enhancement are used for pre-training and data set expansion in this paper.