Non-rigid Registration Of Medical Images And Its Application To Lung Ventilation Modeling

Modeling the lung ventilation is of great significance in image-guided radiotherapy of lung cancer through quantitatively displaying the spatial distribution of air content in lungs.The lung ventilation is the process of content change in lungs,and also the process of deforming movement of the lung tissue.Image registration has ability to estimate the lung deformation by establishing the correspondence between images,which is thus the key technology for lung ventilation modeling.However,lungs are typical moving organs,with complex deformation occurring during breathing cycle,which results in lung image registration rather difficult.At present,even many medical image registration methods have been proposed,they still cannot achieve the ideal registration results when applied to lung images.Therefore,registration of lung images is still a challenging problem which is urgent to solve.In this paper,according to the lung anatomic and respiratory characteristics,we focus on development of non-rigid registration algorithms for three-dimensional Computed Tomography(CT)images of lungs.The main work is summarized as follows:(1)A adaptive non-uniform B-spline method is proposed for lung CT image registration.First,a non-uniform B-spline deformation model is constructed,in which the control point gird is initialized with spatial adaptive sparsity based on the curvature according to the characteristics of the physiological structure of the lung.Then,a smooth regularization term and a total variation regularization term are combined to form a spatially weighted deformation field constraint,which is connected with the data fidelity constraint for establishing the similarity metric of registration.Finally,the non-uniform B-spline registration is implemented from coarse to fine under a multi-scale registration framework.On this basis,the lung ventilation distribution is modeled,and the superiority of the proposed method is verified through experiments.(2)An unsupervised registration method based on Generative Adversarial Network(GAN)is proposed.The proposed GAN registration model is constructed in which a registration network is built as the generator based on the encoder-decoder architecture and a discriminant network is created as the discriminator based on the U-net convolutional neural network architecture.The degree of matching between image pairs is checked according to the output probability by the discriminant network.The proposed GAN model is a general registration framework because it uses a specific similarity measure as the loss function for different registration problems,which effectively avoids registration errors caused by improper selection of similarity measures.Experiment shows that the proposed GAN model obtain more accurate registration results than some other commonly used deep learning models.Moreover,the spatial distribution map of lung ventilation generated by the proposed method is consistent with the physiological characteristics of lung tissue.(3)A lung image registration and ventilation modeling system is designed and implemented.The system mainly includes four functional modules: multi-view image display,image registration,registration result evaluation and ventilation modeling.The system supports the selection of different non-rigid medical image registration methods,including the proposed non-uniform B-spline registration method,the proposed GAN model and some other classical medical image registration methods.In addition,the function of the system also allows for multi-view display of the three-dimensional images and the spatial distribution of lung ventilation,as well as various evaluation of the registration results.