Research On Medical Image Registration And Segmentation Based On Deep Learning

Image registration and image segmentation,as two key technologies in medical image processing,play a very important role in actual medical scenarios such as disease diagnosis,surgical navigation,and therapy evaluation.Traditional medical image registration and manual segmentation are time-consuming and require higher professional level of doctors.With the rapid increase in the types and quantity of medical images,the workload of doctors has increased significantly,leading to missed and misdiagnosed phenomena.With the vigorous development of deep learning in recent years,more and more researchers have introduced deep learning into medical image registration and segmentation tasks,and use computers to assist doctors in diagnosis and improve their work efficiency.This paper analyzes and studies the deficiencies in the existing medical image registration and segmentation algorithms,and proposes corresponding improvement methods based on deep learning technology.In the task of medical image registration,due to the obvious differences in image information（such as grayscale and texture）of different modalities,the convolutional neural network in multi-modal image registration cannot capture the spatial position correspondence directly from the original image.In order to solve this problem,this paper proposes an unsupervised multi-modal image registration model based on edge features,and uses an improved Canny operator to extract edge features in the three directions of the image’s sagittal axis,vertical axis and coronal axis at the same time.Using edge information to guide the full convolutional network to learn multi-modal image registration.This paper uses this model to register rigid data sets and non-rigid data sets,and the Dice similarity coefficient（DSC）reach 97.3%and 97.5%,respectively,which proves the effectiveness of the method proposed in this paper.In the task of medical image segmentation,the classic U-Net model has been widely used,but the convolution operator that constitutes its network structure has certain limitations.To solve this problem,this paper proposes two ways to improve.Aiming at the problem that the convolution operator cannot effectively capture the spatial specific information,this paper integrates the Involution operator into the U-Net structure and proposes InvResUNet,which enhances the model’s ability to model spatial information.Regarding the lack of the ability of convolution operators to establish the associated information between long-distance pixels,this paper combines the self-attention mechanism with the U-Net structure and proposes ST-UNet,which strengthens the information exchange between different pixels in the image,so as to provide more accurate segmentation information for the model.This article evaluates the segmentation performance of InvResUNet and ST-UNet networks on the public datasets BraTS 2015 and BraTS 2020,as well as the glioma and meningioma datasets provided by Beijing 301 Hospital.Among them,on the BraTS 2020 data set,for the whole tumor area,the DSC of InvResUNet and ST-UNet are 91.8%and 91.1%,respectively,which are 19.1%and 18.2%higher than U-Net.Fully verify the reliability and superiority of the proposed models.