Medical Image Registration Based On Multi-sourece Feature Learning

The goal of medical image registration is to obtain the spatial transformation relations,so that the corresponding regions of fixed image and moving image can be aligned by rule and line.Medical image registration is one of the key issues in the field of medical image processing as well as the research foundation for medical applications such as computer-aided diagnosis,surgical assistance and medical image segmentation.Current mainstream registration methods consist of methods based on point set registration and methods based on deep learning.Point features can restore the structure and shape of tissues and organs and hardly affected by image degradation.After extracting the feature points of the medical image,the point set registration algorithm can be employed to estimate the space transformation.At present,the mainstream point set registration algorithms improve the registration accuracy by introducing different types of prior knowledge.However,due to the lack of active evaluation of the prior information,the invalid priors that are affected by the degradation of the point set will have a negative impact on the registration results.The medical image registration methods based on deep learning capture image features through neural network.Most of the existing registration networks employ a single network to train medical images,and lack the cumulative learning of image features.What's more,existing registration networks introduce image pyramids or additional segmentation information to improve the registration accuracy,but lack differentiation learning of image features,resulting in insufficient estimation of local tissue deformation.In order to address the above problems,this thesis focus on the following two aspects:1)a point set registration algorithm based on multi-source prior learning and 2)an iterative medical image registration network based on differentiated learning.The main research contents of this thesis are as follows:1.This thesis proposes a multi-level point set registration framework based on prior evaluation.The algorithm can dynamically evaluate the current registration state based on the posterior information and prior knowledge obtained from the current registration.This thesis also proposes a feature description based on the local topological structure of the point set.The descriptor extracts features from the multi-scale neighbour area,and proposes related geometric feature descriptors according to the topological characteristics of the point set.Experiments proves that the algorithm can effectively reduce the impact of prior failure and has achieved good results in medical image registration tasks.2.This thesis proposes an iterative registration network including a differentiated attention mechanism.Through iterative training,the network can fully learn the structural features among the image.By separately calculating the regular terms of different deformation fields,the network can reduce information loss and better constrain network parameters.This thesis also proposes a differentiated learning attention mechanism based on feature dimensionality reduction.This mechanism enables the network to focus on important features among different dimensions.Experimental results show that the registration network has a good registration result on the MR image of human brain.