Lesion Regions Labeling And Detection For Idiopathic Pulmonary Fibrosis

Idiopathic pulmonary fibrosis is the commonest interstitial lung disease in adults,and patients have a poor prognosis with a median survival of 3-4 years.Early diagnosis of idiopathic pulmonary fibrosis can lead to antifibrotic therapy,which can prolong patient survival and reduce acute deterioration.The existing manual diagnostic approach requires a high level of expertise and extensive clinical experience,and reading large amounts of data is a significant challenge to the radiologist’s attention and energy.There are few studies related to automated detection of idiopathic pulmonary fibrosis in academia,and no computer-assisted detection system is available for clinical use.The lesions of idiopathic pulmonary fibrosis vary in size and shape,and there are many types of lesions.Automatic detection of them is a highly challenging and practical work.In this thesis,the automatic labeling and detection method of idiopathic pulmonary fibrosis lesion regions are designed for high-resolution CT images,and the main research elements are as follows:（1）Theories related to visual detection of idiopathic pulmonary fibrosis were studied,including the fundamentals of high-resolution CT imaging,image feature analysis of idiopathic pulmonary fibrosis,machine learning models and image classification networks,semantic segmentation networks and transfer learning theory in deep learning.（2）A lesion regions detection method based on corner point distribution is proposed.Firstly,the lesion candidate regions are extracted based on the corner point distribution in high-resolution CT images,and a modified U-Net segmentation of pathological lung parenchyma was used to mask the candidate regions,then combined with medical a priori knowledge to model the features of IPF lesions in order to further screen the candidate regions,realizing a simple and fast lesion regions detection method.（3）A lesion regions detection method based on multi-feature fusion is proposed.Firstly,lesion candidate regions are segmented by multiscale K-mean clustering and fractal area,then a classification network is designed to fuse classical visual features and depth features,which combines the advantages of both features and improves the sensitivity and precision of the detection method.（4）A lesion regions detection method based on semi-supervised semantic segmentation is proposed.The network includes a feature extraction backbone network,a bi-directional feature fusion module and an edge guidance module,which uses different CT value ranges to obtain different channels of the input image,and in training,the training set is expanded using label-free data to improve the generalization ability and segmentation performance of the network.