Research On Non-rigid Medical Brain Image Registration Technology

With the rapid development of computer technology,medical imaging has been widely used in modern clinical disease diagnosis and auxiliary treatment.Non-rigid medical image registration technology can accurately describe the spatial correspondence between images.It plays an important role in many application areas such as monitoring and tracking of pathological development changes,the formulation of radiation treatment plans,and feedback evaluation of treatment effects.This paper first studies the non-rigid medical image registration based on the traditional B-spline algorithm with the aim of improving the accuracy and robustness of multi-modal registration,and verifying and evaluating it with actual multimodal brain images.Then studies the non-rigid medical image registration tasks based on deep learning algorithms to achieve end-to-end high-efficiency registration of images.The main contributions of the dissertation are as follows:(1)In the conventional B-spline registration algorithm,the normalized mutual information used as the similarity measure is unable to register multi-modal images due to the lack of its own spatial structure information,a similarity measure ALST-NMI(Adaptive Local Structure tensor-Normalized Mutual Information)combining normalized mutual information and spatial information is proposed.The measure extracts the weight information of the spatial structure of the image according to the local structure tensor parameters based on regional variance.Experimental results show that ALST-NMI can improve the registration accuracy of single-mode and multi-modal images,enhances the robustness of registration,and reduces the risk of local extremes during the registration process.(2)The deep neural network model is prone to degradation during the training process as well as the limitation of convolutional kernels' detection range,an improved multi-scale residual full convolutional network model(MS_ResFCN)is proposed.The model introduces a residual structure into the FCN model to ensure that the model is effectively and smoothly trained.At the same time,a hierarchical multi-scale convolution kernel is constructed in the residual layer,which simultaneously extracts and fuses the local texture information andcontext information of the image,can generate multiple types of features and improve the nonlinear mapping ability of the registration network to some extent.Experiments show that the MS_ResFCN model effectively eliminates the degradation of the deep network during the training process and improves the registration accuracy.