Research On Fatty Liver B-ultrasonic Image Recognition Based On Deep Learning

Fatty liver is the main cause of liver dysfunction,and ultrasound has become the most commonly used method for detecting fatty liver.With the rapid development of computer technology and medical imaging technology,doctors are also facing more and more medical image diagnosis tasks.In order to reduce the pressure of doctors' work,computer-aided diagnosis systems are especially important.In recent years,deep learning has achieved rapid development and has made tremendous breakthroughs in various fields.In the medical aided diagnosis system,deep learning has greatly improved the diagnosis efficiency,but there has not yet been relevant mature research results for the fatty liver ultrasound image recognition.Therefore,the study of liver ultrasound image classification based on deep learning has important practical application value.For deep learning neural networks,this article mainly discusses the establishment of learning network models,parameter selection and optimization methods in fatty liver ultrasound image classification.The main tasks are:1.The feasibility of convolutional neural network to classify fatty liver ultrasound images is verified,and a CNN-based auxiliary diagnosis algorithm is proposed for classifying fatty liver ultrasound images.This method first preprocesses the real ultrasound fatty liver image,then analyzes its texture features through the gray level co-occurrence matrix and visualizing the intermediate process of convolution,and finally designs a shallow CNN network with jump connections according to the texture features.Then compare the classification accuracy of different CNN parameters,and finally determine an optimal CNN model structure.The experimental results show that,compared with the traditional method,this method has greatly improved the classification accuracy of fatty liver ultrasound images.2.A CNN model optimization method based on grid search is given.This method first determines a baseline CNN model,and then restricts the parameters space range of the CNN model through the constraints of Flops.On the basis of the baseline CNN model,a constraint is added to search for a scalable CNN model.Finally,by scaling in on the searched CNN model,an optimized CNN model is obtained.Compared with the design of conventional CNN models,this method does not require manual adjustment of parameters based on experience,saving time and effort;and the classification accuracy of fatty liver ultrasound images exceeds 92%.