Research And Application Of Medical Image Registration Method Based On Deep Learning

As a key link in the medical image processing process,medical image registration aims to align the misplaced and different positions of the same organ structure in the image,and provide medical personnel with accurate diagnosis and treatment images.There are important applications in research and other fields.At present,the registration process of traditional medical image registration methods is complex and slow,and the quality of registration is poor,which is difficult to meet the accuracy and real-time requirements of clinical treatment.At the same time,due to inconsistencies in image shooting conditions,such as inconsistencies in shooting angles,positions,and degree of organ compression,there are often large displacement deviations and complex deformation relationships in the images.If these relationships are not corrected and optimized,it is difficult to accurately align the inconsistencies in the images The region that affects the final registration quality of the method.To this end,this paper use deep learning technology and multi-scale strategy to improve the above problems,and use MRI data sets of human brain and knee cartilage to analyze and verify.Its main research focuses and contributions are as follows:(1)This article proposes a multi-scale medical image affine registration method based on Transformer.In response to the problems of large linear misalignment relationships in images and low registration efficiency of traditional methods,this method combines Swin Transformer and image pyramid technology in deep learning.By inputting multi-level feature maps and image block mapping concatenation,it provides multi-scale image features for the registration network.By learning the details of high-resolution feature maps,it guides the improvement of the prediction parameter results in the previous stage,Realized optimization of affine registration results from coarse to fine.At the same time,for 3D medical images,this method proposes a stage vision module,introduces a self attention mechanism with a shift window,and implements its 3D version to achieve global feature relationship modeling over long distances.The position embedding matrix is removed and continuous small convolutional kernel image block embedding is used instead of linear image block embedding to better encode the pixel level spatial information of the image;By constructing an affine alignment head,the image block features are converted into geometric transformation parameters,and then the affine transformation matrix is constructed.Finally,the accuracy and real-time performance of this method were demonstrated through comparative experiments.(2)This article proposes a Transformer based cascaded multi-level medical image registration method.In response to the complex deformation and large displacement deformation situations in medical images,combined with multi resolution strategy and cascading strategy,the most difficult deformation areas in the image are optimized,breaking through the registration upper limit of single-layer networks.This method defines the deformable registration process as a hierarchical optimization process,cascading three hierarchical subnetworks with gradually decreasing input scales.By continuously screening local difficult deformation regions and inputting them into a single scale registration network for optimization,the overall registration quality is gradually improved;In a single hierarchical sub network,a single scale U-shaped registration network was constructed by combining CNN and Transfomer,maximizing the global and local modeling capabilities of the model and learning the complex displacement relationships of local regions;At the level connection,a difficult deformation perceptron based on MSE coefficient is equipped to filter difficult deformation regions and input subsequent level optimization.The experimental results show that the proposed method has better registration quality compared to the currently popular unsupervised registration methods.(3)Aiming at the difficulty and poor practicability of traditional registration systems,this paper develops a medical image registration system based on the two proposed multiscale registration methods,which gives the specific application value of the method.The system integrates the algorithm model into the operable application interface,and realizes the affine registration and deformable registration functions of click operation.The system provides medical researchers with high-quality registration quality and efficient registration efficiency,thereby helping doctors better understand the condition,provide patients with more accurate diagnosis and treatment,and also promote the development of medical research.