Study On Staging Dagnosis Of Liver Fibrosis Based On Ultrasound Image

Liver fibrosis is a relatively serious liver function disease that can progress to cirrhosis or even cause death if left untreated.Studies have shown that early liver fibrosis can be cured,therefore,early detection and accurate staging of fibrosis and cirrhosis is essential for early diagnosis and timely initiation of appropriate treatment options.Liver biopsy is the "gold standard" for the diagnosis of liver fibrosis,but its clinical use is limited due to its invasive nature and the physical damage it can cause to the patient.Ultrasound images are widely used because they are inexpensive and noninvasive,and clinicians judge the grade of liver fibrosis by observing the roughness of the ultrasound image texture,which is subjective and prone to misdiagnosis.Therefore,in this thesis,ultrasound liver fibrosis images are used as the research object,and ultrasound liver fibrosis image classification methods are investigated by using the current popular deep learning-related technologies.In the image preprocessing stage,firstly,for the existence of non-ultrasound imaging regions in the original image,an algorithm is designed to automatically extract ultrasound imaging regions to exclude the influence of interfering regions on the classification of ultrasound this fibrosis image.Secondly,for the existence of a large amount of shadow noise in some images,Zero-DCE image enhancement technique is used to denoise the images;meanwhile,for the problem of difficult extraction of liver parenchyma region,manual extraction of liver parenchyma is performed;finally,for the problem of small ultrasound liver fibrosis data set,methods such as image geometric transformation are used to expand the data.In the ultrasound liver fibrosis image classification stage,three different classification models,including Res Net50,Mobile Net V2 and Efficient Net V2-S,were used to complete the early,middle and late ultrasound liver fibrosis image classification tasks,and four evaluation metrics were used to evaluate the models in a multidimensional and systematic analysis.The experimental results show that the Efficient Net V2-S-based classification model has good classification performance in the ultrasound liver fibrosis image classification task with an accuracy of 70% and F1-Score performance metrics of 80.65%,54.17%,and 72.11% for early,intermediate,and late liver fibrosis,respectively.To further improve the classification performance of the model,an ultrasound liver fibrosis image classification model with an improved model Efficient Net V2-BCH was proposed.Firstly,to address the problem of difficult extraction of liver fibrosis lesion detail information,feature pyramid branches are designed to fuse features at different scales in the images;secondly,to address the problem of insignificant differences between two adjacent liver fibrosis lesion stages in ultrasound images,Canny edge detection and Haar wavelet transform feature maps are extracted and fused with Bi FPN branches for features,respectively,and finally with the backbone network Efficient Net V2-S for multi-scale multi-feature fusion.The experimental results show that the improved model Efficient Net V2-BCH can achieve 89.41% classification accuracy,which is 19% more accurate than the original model Efficient Net V2-S;the recall rate of early,intermediate and advanced liver fibrosis is improved by 10%,36% and 13%,respectively;the F1-Score of early,intermediate and advanced liver fibrosis is improved by 11% and 33%,respectively.Score improved by 11%,33% and 13%,respectively.