Research On Transformer-Based Medical Image Registration

The task of image registration is to search the spatial mapping relationship between images.Image registration is an important step in the processing and analysis of medical images.Researchers can conduct subsequent studies more accurately and conveniently by aligning the structures in medical images with image registration.Therefore,it is essential to propose more accurate and faster medical image registration methods for medical image processing,clinical diagnosis,surgical navigation,and so on.Recently,with the rapid development of computer science,medical image registration has been widely studied and improved.However,the current medical image registration methods have some shortcomings,such as complex calculation process and insufficient registration accuracy,which limit the application of image registration in medical image analysis.To solve the above problems,this thesis proposes two registration methods based on Transformer for electron microscopy images and magnetic resonance images among various medical image modalities with the help of deep learning algorithms and achieves competitive registration accuracy and speed.The main work of this thesis is as follows:(1)For the current situation that the existing electron microscopy image registration methods cannot effectively utilize the information of multiple reference images,this thesis proposes an electron microscopy image registration network based on Transformer.In this method,Transformer is used to fully model the correlation between multiple reference images.In the Transformer encoder,we use the attention mechanism to correlate related regions in multiple reference images;in the Transformer decoder,the prediction query is introduced,which is able to interact with the enhanced features from the Transformer encoder.Then the ground truth feature of the image to be registered is predicted and used as a reference to calculate more accurate registration results.The method is evaluated and analyzed on Cremi and FIB25 datasets.Experimental results show that the electron microscopy image registration network proposed in this thesis makes full use of Transformer to model long-range dependence of multiple images,which extracts the changing tendency of images and achieves more accurate and faster electron microscopy image registration results.(2)For the current situation that existing magnetic resonance image registration methods cannot fully model the correlation of multi-scale image features,this thesis proposes a magnetic resonance image registration network based on Transformer.In this method,the image pyramid of different resolutions is constructed,and Transformer is used in each layer to globally correlate image features and extract attention-enhanced features.The predicted registration field of each layer is taken as initial deformation of the next layer,and the residual of registration field is predicted for further registration.In this way,the registration field is gradually refined through multi-scale residual prediction,which is able to reduce the calculation cost of multiple registration steps with higher accuracy.This method is evaluated and analyzed on OASIS and LPBA datasets.Experimental results show that the magnetic resonance image registration network proposed in this thesis can make full use of Transformer to model the correlation of multi-scale image features and correlate information of different levels,which refines registration fields in the image pyramid and improves the accuracy of registration with competitive speed.