Research On Nonalcoholic Fatty Liver Disease Method Based On Ultrasound Images

NAFLD is a common global disease and an important cause of liver cancer and cirrhosis,and its serious impact on human health makes timely and accurate screening and diagnosis of the severity of NAFLD crucial.However,due to the extreme scarcity of ultrasonographers,the use of computer technology to assist in NAFLD screening and diagnosis is imperative.The task of NAFLD grading is of great significance in the medical and scientific fields as a key task of the ancillary diagnostic system that can help physicians to quickly and accurately confirm the diagnosis of NAFLD and its severity.In this thesis,we investigate the grading method of NAFLD using deep learning method with ultrasound images of NAFLD.For the image preprocessing of NAFLD,wavelet-bilateral filtering algorithm is used to remove the interference of strong speckle noise in ultrasound images;for the problem of unbalanced image brightness,two-dimensional gamma function is used for adaptive correction of image brightness;for the problem of small and unbalanced kinds of data in medical image dataset,a combination of deep convolutional adversarial network with fused residual structure and traditional data enhancement methods is used for data enhancement of the dataset.The classification models for NAFLD are selected according to the three series models:Dense Net,Mobile Net and Efficient Net.Secondly,the accuracy,accuracy,specificity,recall rate and F1-score other evaluation indexes are combined,the 86.4% Efficient Net-B2 model is selected as the basic network model.Finally,a classification method based on channel attention mechanism and Ghost Net thought was proposed to solve the problem of poor grading effect of severe nonalcoholic fatty liver disease.The introduction of ECA attention mechanism can avoid the side effect that the channel and weight do not correspond directly due to the reduction of the dimension of SE module,so as to improve the image feature extraction ability of the hierarchical model.The introduction of Ghost module can reduce the storage space and computing power of the original model,thereby speeding up the training speed of the hierarchical model;The introduction of h-swish activation function can accelerate the convergence rate of the model and further optimize the performance of the hierarchical model.The experimental results showed that the ultrasound image grading method of NAFLD proposed in this thesis achieved accurate grading of NAFLD with an accuracy of 92.5%.It improved 6.1% compared with the original model Efficientnet-B2.The accuracy,recall,specificity and F1-score also respectively increased by 6.1%,6.1%,4.3% and 6.1%.