Visual Information Modeling And Lesion Detection For Idiopathic Pulmonary Fibrosis

Idiopathic pulmonary fibrosis(IPF),as a typical manifestation of interstitial lung diseases(ILDs),has the characteristics of invisibility at the initial stage and high fatality rate in the middle/late stages.Unfortunately,there is no efficient treatment for IPF disease currently,so medical check-up early,diagnosis early,and treatment early can effectively improve the survival rate of IPF patients.Artificial diagnosis of IPF,however,needs a lot of complex steps,which not only requires experienced doctors but consumes a great deal of energy and time of doctors.For one thing,there are few related researches of automatic detection of IPF in CADs,and for another,IPF tissues ordinarily have significantly different visual features from the other lung tissues in CT images.Consider that,this thesis designs a new algorithm based on high-resolution CT for IPF automatic detection.The main research content of this thesis as follows:(1)The basic target detection of lung CT images is studied,including the principle of CT imaging,image texture analysis,machine learning model,deep learning based semantics segmentation algorithm and migration learning method;(2)A visual feature modeling method of IPF tissue based on image texture analysis is proposed.A multiscale rotation invariant visual feature model of IPF tissue is constructed by combining the Gabor transformation,LBP feature and main orientation HOG feature of IPF tissue.On the ILD-database dataset,the precision of this model for lung tissue classification in common interstitial diseases can reach more than 88%;(3)A detection method of IPF tissue in CT image based on K-means clustering is proposed.Firstly,the candidate areas of IPF tissue in the original lung CT images are roughly extracted by K-means clustering algorithm.Then,the final candidate areas of IPF are obtained by super-pixel segmentation and image morphological processing on the roughly candidate areas of IPF.Finally,the real IPF tissue are detected based on the IPF visual feature model in(2);(4)An IPF tissue detection method based on UNet is proposed.Firstly,this thesis replaces the original UNet encoder with VGG11 and adds the attention module to the UNet decoder.Then,IPF-texture set,a collection of natural texture image data similar to IPF tissue image,is constructed and pre-trained.Finally,the pre-training weight is transferred to VGG11 and the network is fine-tuned,which greatly improves the precision of the network for IPF tissue detection.