Research On The Model Classification And Features Analysis Of Ultrasonic Images Of Hepatic Fibrosis

Liver fibrosis is an intermediate stage in the pathological process of chronic liver disease.Early clinical intervention in liver fibrosis can slow the progression of cirrhosis and reduce the risk of liver cancer.Liver biopsy is the gold standard for viral liver disease management,but it has some limitations,such as invasiveness and high sampling error rate.The accuracy and specificity of the diagnostic model based on serological laboratory indicators were not ideal,and the combined serological and medical imaging models failed to meet clinical needs.Therefore,we continued to explore a non-invasive,accurate and economical method for grading liver fibrosis.The changes on the ultrasonographic caused by the hepatic fibrosis and those caused by the noise is similar to each other.The existing Liver Fibrosis Index(LFI)of Fibrosis grading is low efficiency;therefore,in this dissertation,by reducing the ultrasonic image speckle,machine learning algorithms,and build a new liver diagnosis index,it explored a method to grade liver fibrosis based on the ultrasonic image of hepatitis B liver fibrosis,hoping to provide new method for clinical diagnosis and evaluation.(1)Aiming at the interference of a common speckle in ultrasonic images with lesion information,a new method was proposed in this chapter.It is based on super-pixel segmentation and detail compensation,with the goal of reducing speckle in ultrasonic images.In this method,certain parts of the structure are better protected by the bilateral filter of the super-pixel segmentation version.At the same time,a strategy inspired by human visual system is introduced,with spatial compensation,to reduce the high frequency noise while recovering the complex edge as much as possible.Through the experiments on synthetic images and real ultrasound images of different organs,this method has already been proved to a more efficient way to reduce ultrasonic speckle.(2)Real Time Elastography(RTE)has the advantages of non-invasive and accurate assessment of lesions,but it is a challenging to determine the stages of liver fibrosis from RTE liver fibrosis images.Aiming at this problem,this study adopted four classical classification methods,including Support Vector Machine,Na?ve Bayes,Random Forest and K-Nearest Neighbor to establish four better decision supporting systems for grading diagnosis of hepatitis B.In this multicentre collaborative study,11 RTE image features were obtained from 513 subjects who underwent liver biopsies.The experimental results show that comparing with the previous diagnostic prediction LFI depending on RTE image and multiple regression analysis,the classifier adopted in this method is superior.Among the four machine algorithm,the random forest classifier has the highest average accuracy.The results show that the machine learning method can be a powerful tool for evaluation of liver fibrosis stage and has a good clinical application prospect.(3)There were numerous methods for diagnosis liver fibrosis emerging in clinic and their parameters were looming.To solve this problem,this dissertation mainly discuss the diagnostic efficiency of all auxiliary examination method for grading fibrosis of hepatitis B.This experiment has collected the subjects’ information,including demographic characteristics,clinical symptoms,serum indexes,two-dimensional ultrasound images,real-time ultrasound elastography data.Then radiomics was applied to extract features from two-dimensional ultrasound images,and machine learning of random forest was used to combine all types of features into multiply fusion modes and classify,hoping to find the optimal checking method.After that,with the Lasso regression model(Least Absolute Shrinkage and Selection Operator,Lasso),a new diagnostic index for liver DCSRT(Demographic Characteristics,Clinical Symptom,Serum Index,Real-time Ultrasound Elastography,Two-dimensional Ultrasound,DCSRT)is constructed.The Area Under Curve(AUC)of the wavelet 1-direction2-texture4 and wavelet1-direction4-texture1 in the test group and the training group was greater than 0.8 when differentiating whether there was liver fibrosis,indicating that normal ultrasound played a crucial role in disease screening.When identifying fribosis of middle stage,the AUC of wavelet 1-direction2-texture4 and GGT in the test group was higher but slightly lower than that in the training group,indicating that with the progression of the disease,serological examination was indispensable.When identifying fribosis of severe stage,both the AUC of the gammna-glutanyl transpeptidase(GGT)and age in the test group were higher and larger than those of the training group,indicating that age was an important factor in aggravating the disease.When distinguishing whether there is cirrhosis or not,the AUC of Platelets(PLT)and age is relatively large,the age of the test group has reached 0.91.Therefore,blood test for patients with cirrhosis of terminal stage is very important.These rules can guide doctors in clinical.From the perspective of medical image processing,this dissertation proposed a noise reduction model based on visual mechanism,which aims at the interference factors of grading of liver fibrosis,and quantify the image information.These high-dimensional data were input into machine learning model to solve clinical problems puzzled the doctor.This is a new solution to the problem of the diagnosis of liver fibrosis grading and will play an important role in the clinical research and application.