Research On Evaluation Method Of Liver Fibrosis By Ultrasound Images Based On Radiomics And Deep Transfer Learning

China is a high-risk area for chronic hepatitis B(CHB).Liver fibrosis is a progressive disease of CHB.Without timely treatment,it can evolve into liver cirrhosis or even cancer.Relevant studies have shown that timely and proper treatment can reverse the evolution process of liver fibrosis and cirrhosis.Therefore,accurate assessment of liver fibrosis is of critical importance for clinical treatment and prognosis prediction of liver fibrosis.Though liver biopsy is the “gold standard” for diagnosis of liver fibrosis,it has several practical limitations as it is invasive,toxic and difficult to be accepted.Therefore,it is highly desirable to develop a non-invasive,safe and accurate method for evaluating liver fibrosis.Currently,the clinical non-invasive tools developed for liver fibrosis assessment are mainly based on serological testing and a variety of imaging studies.Among them,ultrasound imaging has the advantages of non-invasive,non-radiative,real-time,portable,etc.,and hence it is a major means for clinical examination of liver fibrosis.This study bases on radiomics and deep transfer learning to explore methods for non-invasive evaluation of liver fibrosis in ultrasound images.In cooperation with Shenzhen Third People's Hospital,we recruited 321 CHB patients where assessments with liver biopsy,serological examination and medical ultrasound examination were performed.Based on the radiomics method,121 radiomics features were extracted and screened from the B-mode ultrasound images of the right lobe of liver,and three machine learning methods,including support vector machine(SVM),logistic regression(LR)and deep transfer learning,were tested in the experiments.For discriminating liver fibrosis(?F2),the SVM performs the best,followed by LR method.The area under the receiver operator curve(AUC)averaged via ten cross-validation reaches up to 88% for SVM and LR.For the identification of liver cirrhosis(F4),the average AUC obtained by LR is 0.87,which is slightly higher than that of SVM,86%.For the deep transfer learning method,the average AUC,accuracy,sensitivity,and specificity are respectively 0.96,92%,92%,92% for discriminating ?F2 and 0.95,88%,88%,90% for discriminating F4.In order to overcome the limitations of manual labeling of region of interest(ROI)as done in other similar researches,the classical Unet segmentation network is further developed based on the depth residual structure to achieve accurate segmentation of liver parenchyma in B-mode ultrasound images.The segmentation works pave way for the subsequent prediction of liver fibrosis stages.Finally,fully automated and non-invasive assessment of liver fibrosis is achieved by the joint use of segmentation and evaluation models.The thesis studies the methods for non-invasive evaluation of liver fibrosis in ultrasound images based on radiomics and deep transfer learning.By the joint use of a deep learning network for liver parenchymal segmentation,automatic non-invasive prediction for liver fibrosis stages is achieved.Experimental results show that the evaluation model based on deep transfer learning has a better diagnostic performance for early liver fibrosis prediction.In the future,we will further improve the generalization capability of the model via multi-center cooperation and iterative model optimization,so that the method could be clinically recognized as soon as possible.