Classification Of Liver Fibrosis Based On Machine Learning And Multiphoton Imaging

Non-alcoholic Fatty Liver Disease(NAFLD)is the most common liver disease worldwide and the leading cause of liver-related morbidity and mortality.Only the grade of liver fibrosis can independently predict the long-term prognosis of NAFLD patients based on liver histological indices.At present,the imaging resolution of traditional imaging methods is insufficient,and liver histopathologic biopsy is time-consuming and laborious and affected by human error.Therefore,the development of a new technique,which can be used to rapidly,objectively and accurately stage liver fibrosis at a higher resolution,will significantly improve the classification and management of this disease.Multiphoton microscopy is a fast microscopic imaging technique with high spatial resolution.It can obtain strong optical signal in biological tissue without special staining and tissue processing.It causes minimal damage to biological tissues,has less phototoxicity and photobleaching,and can image deep into tissues.Multiphoton imaging can effectively overcome the shortcomings of traditional liver biopsy,and is expected to provide a new method for quantifying the grade of liver fibrosis.The recognition and analysis of a large number of images and medical data need to rely on the increasingly developed machine learning means.In particular,as a subset of machine learning,deep learning has achieved high accuracy in tasks such as image classification,and semantic segmentation.Therefore,the application of machine learning for quantifying liver fibrosis has great potential in reducing the human error and improving the objectivity and accuracy of clinical diagnosis.In this thesis,we conducted a study on the liver fibrosis grading diagnosis of NAFLD,utilizing machine learning and multiphoton imaging.The main findings are summarized as follows: Firstly,our study utilized multiphoton imaging to capture images of hepatocellular carcinoma tissue.We then applied machine learning to extract features of collagen fibers and conducted qualitative and quantitative analysis.This approach provided a solid foundation for further investigations.Subsequently,high-resolution images of fatty liver tissue were acquired via multiphoton imaging.Based on three major deep learning models,namely VGG16,Res Net34,and Mobile Net V3,a transfer learning-based automatic classification model for liver fibrosis was established.The results indicate that the VGG16 model has the best performance,demonstrating the great potential of combining deep learning and multiphoton imaging in clinical liver fibrosis grade.Finally,in order to optimize the performance of the deep learning model,the multi-layer perceptron was used to perform data fusion on three image features(clinical information features,machine learning-based manual features and deep learning features).A multi-dimensional and multi-mode Auto Fibro Net(Automated Liver Fibrosis Grading Network)model was developed.The results show that the Area Under the Receiver Operating Characteristic Curve(AUROC)of the model for different fibrosis grades is 1.00,0.99,0.98 and 0.98,respectively The model’s outstanding performance underscores its immense potential for clinical application.In conclusion,the research work in this thesis demonstrate that multiphoton imaging combined with machine learning can achieve rapid and label-free classification of liver fibrosis grades,thus enhancing the efficiency of clinical diagnosis.In particular,the introduction of deep learning has made it possible to achieve automatic liver fibrosis grade.