Research On 3D Head MR Image Registration Method Based On Deep Learning

Magnetic Resonance(MR)imaging of the head is a non-invasive imaging technique that uses magnetic fields and electromagnetic waves to generate high-resolution,threedimensional images of the head to reveal detailed information about head structures and brain tissue.In the medical field,MR images of the head are widely used in medical diagnosis,treatment and research,and are an important tool for physicians’ decision making and surgical planning.With the wide application of computer-aided diagnostic techniques in various medical image analysis tasks,image registration techniques have become an important medical image pre-processing step.To achieve accurate registration and fusion of head MR images,we propose the following three medical image registration methods based on deep learning:(1)In order to solve the problem that general registration methods cannot handle the complex spatial and positional information of head MR images,this paper proposes a multiscale feature fusion registration network(MFF-net)based on deep learning.The network consists of three sequentially trained modules,the first module is an affine registration module that implements affine transformation.The second module is a deformable registration module that implements a non-rigid transformation and consists of top-down and bottom-up feature fusion subnetworks in parallel.The third module is a deformable registration module that also implements the non-rigid transformation and consists of two feature fusion subnetworks connected in series.This network decomposes the deformation field of large displacement into multiple deformation fields of small displacement by multiscale progressive registration,which reduces the difficulty of registration.At the same time,the multi-scale information in the MR images of the head is learned in a targeted manner through the connection of two ways of feature fusion subnetworks,which improves the registration effect of complex space and position.Four index values were calculated for this network after registration: DSC(Dice Score)was 0.747±0.020,HD(Hausdorff Distance)was 3.350±0.882 mm,ASD(Average Surface Distance)was 0.695±0.087 mm,Std.Jacobian(Standard deviation of the Jacobian Matrix)was 0.379±0.032.(2)In order to solve the problem that the cascade network is too complex in performing cascade registration of head MR images,the model is too complicated resulting in too large amount of parameters,a deep learning-based channel attention cascade registration network(CAC-net)is proposed in this paper.The network consists of several subnetwork cascades,including two different subnetworks: one is an affine registration subnetwork that predicts affine transformations,and the other is a deformable registration subnetwork that predicts deformation fields.To further enhance the feature representation,we add an attention mechanism to the deformable registration subnetwork to assign feature weights.Then,the affine collocation network and a few channel attention networks are cascaded in order,and the output of each subnetwork is used as the input of the next subnetwork.During training,the parameters of all subnetworks are updated simultaneously for faster convergence.The channel attention mechanism allows the network to selectively focus on the more important channels in the feature map and ignore the less important ones,thus reducing the number of network parameters without sacrificing the registration accuracy and reducing the training burden of the cascaded network.Four metric values were calculated for this network after registration: DSC was 0.739±0.024,HD was 3.559±0.689 mm,ASD was 0.763±0.103 mm,Std.Jacobian was 0.457±0.058.(3)In order to solve the problem of predicting large displacement deformation difficulties and local distortion of deformation field due to small perceptual field when the registration model based on Convolutional Neural Networks(CNN)is aligned,a deep learning-based Transformer-CNN hybrid cascade registration network(TCHC-net)is proposed in this paper.The network consists of several subnetwork cascades,including an affine registration subnetwork for predicting affine transformations and a deformable registration subnetwork for predicting deformation fields.In the encoder stage of the deformable registration subnetwork,we employ three Swin-Transformer blocks for downsampling;in the decoder stage,we use conventional convolutional layers and fuse feature maps of the same resolution in the encoder and decoder stages using a crossconnected structure.In the Swin-Transformer block,the self-attention mechanism helps the network to learn and capture the interdependencies between long-distance pixel points in the image more easily,and thus learn the large displacement deformations between pixels more easily.Four metric values are calculated for this network after registration: DSC was0.762±0.015,HD was 3.179±1.267 mm,ASD was 0.625±0.109 mm,Std.Jacobian was0.524±0.070.Comparing the registration results of the above three methods,it is found that MFF-net utilizes different registration modules to extract multi-scale features and handles the best results in terms of registration details.the CAC-net network has the smallest number of parameters,the simplest structure,the easiest implementation,and the most computational resources saving under the condition of achieving almost the same registration results.The TCHC-net network has a large receptive field,which facilitates the extraction of macroscopic and abstract features.Furthermore,it predicts a more continuous deformation field,leading to the best registration performance on 3D head MR images.