Research On Brain MRI Image Registration Method Based On Convolutional Variational Autoencoding Network

Medical image registration is a key technology for processing medical image information in clinical medical research.Traditional methods mainly rely on special markers manually identified by doctors as key points to guide image registration.However,the process is complicated and lengthy and relies more on doctors’experience.It’s easy to make mistakes.With the continuous development of digital images,traditional methods can no longer cope with more and more high-precision medical images.Therefore,the use of deep learning(DL)networks for medical image registration has gradually become a research hotspot.The research on registration methods for high-precision digital images mainly uses end-to-end registration methods,although this method can achieve rapid One-step registration,but compared with traditional methods,important key part feature information,such as the location of the lesion,will be lost during registration,which is not conducive to clinical pathological analysis and judgmentAiming at the shortcomings of existing registration methods,combined with the advantages of artificial intelligence(AI)in feature selection,this paper proposes an image registration method based on deep convolutional variational autoencoding network(DCVAE),It can not only extract the key point feature information of medical images intelligently,but also improve the speed and accuracy of medical image registration.The main content and research results of the thesis include:A key point feature extraction method based on deep convolutional variational autoencoding network is proposed.A feature-based medical image registration method is proposed.First,the HAMMER algorithm and the Diffeomorphic demons algorithm are improved to adapt the feature information of the key points extracted by the convolutional variational self-encoding network.Then the brain MRI image is preprocessed to make it suitable as the input of the convolutional variational self-encoding network,and then the key point feature information output by the network is used to match the key points of the image to be registered one by one.Finally,two improved traditional algorithms are used to realize the registration of medical images.Based on the above research results,medical image registration experiments were carried out on the ADNI(Alzheimer’s Disease Neuroimaging Initiative)and LONI datasets,and the registration effect was evaluated.The experiments confirmed the feasibility and effectiveness of the proposed method.The experimental results show that compared with the key point pairs obtained by the traditional methods,the key point pairs guided by the deep convolution variational self-coding network constructed in this thesis can obtain higher accuracy in guiding medical image registration.At the same time,because convolution variational self-coding network does not depend on artificial feature points or image labels,it belongs to unsupervised network,and has more advantages in the practicability and efficiency of clinical medicine.In addition,the consistent registration accuracy is obtained on both datasets,which indicates that the constructed network has certain stability and versatility.In summary,the medical image registration method based on feature point extraction proposed in this thesis can carry out clinical medical image registration more quickly and effectively,and has a good auxiliary effect on doctors’ work,and has certain practical value.