Research On Medical Image Registration And Segmentation Technology Based On Deep Learning

As two key technologies of medical image processing,medical image registration and segmentation are widely used in many fields such as clinical diagnosis,precision medicine,and postoperative evaluation.However,in practical applications,due to the complex characteristics of the target individual structure and the variety of medical imaging,there are still many problems in medical image registration and segmentation technology,which cannot meet the needs of clinical applications in speed and accuracy,so medical image registration and segmentation are still the focus of research in the field of medical image processing.This paper combines deep learning methods to analyze and study some main problems in medical image registration and segmentation technology.The main work and innovations of this paper include:Aiming at the problems of slow convergence and easy to fall into local maximum in the traditional rigid registration method for multi-modal image,a hybrid registration method based on Fully Convolutional Networks(FCN)and mutual information is proposed.This method uses FCN to extract the deep features of the images and perform coarse registration.Then,the transformation parameters obtained from FCN are used as the initial search point of the mutual information algorithm,so as to narrow the search range of the optimal solution.Finally,the mutual information algorithm is used to further fine tune and optimize the parameters to obtain the final registration results.Experimental results show that the proposed method can not only effectively avoid falling into local extremum and obtain higher registration accuracy,but also greatly shorten the registration time and have faster registration speed.In order to solve the problem that the traditional non-rigid registration method has slow registration speed and the deep learning method has difficulty in obtaining the label,an unsupervised non-rigid registration method based on U-Net and Spatial Transformer Network(STN)is proposed.This method first builds a model that can achieve fast non-rigid registration based on the U-Net structure;then combines STN to achieve unsupervisedadaptive optimization of model parameters.This method can learn the non-linear correspondence between image features without training labels,can quickly complete non-rigid registration tasks,and has better registration performance than traditional methods.Accurately segmenting gliomas from multi-modal magnetic resonance images is a very challenging task.The existing segmentation algorithms have more or less deficiencies in segmentation accuracy,in order to further improve the segmentation accuracy of glioma,a segmentation method based on multi-scale Densenet and ensemble learning is proposed.Firstly,the multi-scale mechanism is introduced into the Densenet model to construct the base classifiers with different cross-sections.Then,the information obtained by the base classifiers with different cross-sections is integrated and analyzed with the ensemble learning to achieve the segmentation of brain tumours.Finally,the Conditional Random Field(CRF)is used to further optimize the segmentation boundary to get the final segmentation results.In this paper,we verify the proposed method based on open datasets and compare it with other mainstream segmentation methods.The experimental results show that the proposed method has higher segmentation accuracy.