Medical image registration method based on multi-scale deformation decomposition

Abstract

The invention belongs to the technical field of medical image processing, and relates to a medical image registration method based on multi-scale deformation decomposition. Firstly, a multi-scale feature extraction network is constructed, the method focuses on extracting features of a mobile medical image, and the features are transmitted to a hierarchical feature fusion module to be processed; secondly, the hierarchical feature fusion network is used for establishing a mapping relation between the mobile medical image and the fixed medical image; in a feature decoding stage, a mobile decoding feature network and a fusion decoding feature network are respectively constructed, multi-scale decoding features are utilized to output a multi-scale deformation field, and a coarse-to-fine registration strategy is gradually realized by fusing the deformation fields of adjacent stages layer by layer. By designing the double-branch network structure and connecting the gating updating modules in the adjacent stages, the complex deformation processing capability of the registration network is enhanced, so that the final registration effect is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of medical image processing and relates to a medical image registration method of multi-scale deformation decomposition. Background Art

[0002] Deformable medical image registration aims to align the anatomical structures of mobile medical images with those of fixed medical images so that multiple medical images can be transformed into the same coordinate system. Medical registration is crucial in the field of medical image analysis and is widely used in early disease diagnosis, preoperative planning, and intraoperative navigation. However, registration faces complex deformation challenges due to individual differences, posture changes, and differences in imaging devices. Despite the existence of a variety of advanced models, it remains difficult to achieve accurate and efficient large-volume deformation registration. Traditional methods regard registration as an optimization problem, aiming to improve the grayscale consistency between mobile and fixed medical images, but these methods are usually computationally expensive and time-consuming, and have difficulty coping with complex deformations, which is not conducive to real-time clinical applications.

[0003] In recent years, deep neural network-based registration technology has become the mainstream method for medical image registration due to its fast inference time and high accuracy. Compared with traditional iterative registration methods, deep learning models adjust parameters through global optimization and are applicable to the entire training set rather than a single image pair, thus enabling rapid alignment of new images.

[0004] VoxelMorph is an unsupervised deep registration network based on the U-Net architecture. It generates a deformation field by splicing pairs of images to be registered as input. During its training, the network parameters are optimized by evaluating the similarity between fixed medical images and deformed moving medical images, without the need for real labels. Although some methods try to directly predict the deformation field, single-stage networks often under-model large displacements. For this reason, cascade-based methods use multiple U-Nets to gradually predict the deformation field, but the computational burden is heavy. The pyramid method processes moving and fixed medical images separately through the encoder, and then gradually predicts the deformation field in the decoding stage, capturing global displacements at low resolution and focusing on local deformations at high resolution. This coarse-to-fine registration strategy improves the accuracy and robustness of the registration.

[0005] However, the iterative framework based on the coarse-to-fine strategy, whose network is relatively independent in each iterative step, is not the best solution for enhancing image features. Mobile medical images and fixed medical images are often connected or fed to the same network layer, which may cause the iterative learning and distortion of mobile medical images to be entangled with fixed medical images, thus affecting medical image registration. In addition, the existing iterative registration framework based on the coarse-to-fine strategy usually only upsamples the deformation field of the previous stage and then deforms the mobile medical image features with the deformation field to achieve the coarse-to-fine strategy, that is, the registration of adjacent stages is only connected by the deformation field. However, the connection between the registrations of adjacent stages cannot be fully established by the above method. Summary of the invention

[0006] The technical problem to be solved by the present invention is: to solve the problem of insufficient accuracy of current medical image registration, and propose a registration method based on multi-scale deformation decomposition, which significantly improves the registration performance through hierarchical feature fusion and gated update module.

[0007] The present invention is a new medical image registration framework with a dual-branch parallel network structure, which makes full use of the prior information provided by mobile medical images. Specifically, the network is a dual-branch structure, one of which only uses mobile medical images as the input of the network branch, and uses a convolutional neural network (CNN) as its backbone network to give full play to the feature extraction ability of the CNN network. The other branch first uses a fixed medical image as input, and at the same time fuses the features of the mobile medical images at the same stage, that is, to achieve hierarchical feature fusion. This branch focuses on the mapping relationship between mobile medical images and fixed medical images. In the feature decoding stage in the dual-branch network, intermediate features are taken from each branch to achieve continuous deformation from coarse to fine.

[0008] Because the original intention of proposing the iterative registration framework from coarse to fine strategy is to hope that the initial deformation field (global large displacement deformation field) is more accurate and reasonable, and it is hoped that the registration network can dynamically pay attention to the multi-scale deformation field and adaptively adjust the importance between adjacent scales. Therefore, the present invention proposes a gated update module (GUM), improves the gated recurrent unit (GRU) to make it more suitable for our registration task, connects adjacent scale registrations through GUM, establishes long-range dependencies between adjacent scales, and allows the network to adaptively adjust the importance of multi-scale deformation fields.

[0009] The present invention first constructs a multi-scale feature extraction network, focusing on extracting the features of mobile medical images, which are passed to the hierarchical feature fusion module for processing; secondly, the hierarchical feature fusion network is used to establish the mapping relationship between mobile medical images and fixed medical images; in the feature decoding stage, a mobile decoding feature network and a fusion decoding feature network are constructed respectively, and the multi-scale decoding features are used to output a multi-scale deformation field, and the deformation fields of adjacent stages are fused layer by layer to gradually realize the coarse-to-fine registration strategy. By designing a dual-branch network structure and a gated update module connecting adjacent scales, the ability of the registration network to handle complex deformations is enhanced, thereby improving the final registration effect.

[0010] The technical solution of the present invention:

[0011] A multi-scale deformation decomposition medical image registration method, the steps are as follows:

[0012] Step (1) Construct a multi-scale feature extraction network

[0013] The mobile medical image is fed into a pyramid feature extraction network to obtain mobile feature maps of different scales. The core component of the pyramid feature extraction network is the 3D convolution layer, which uses 3D convolution, instance normalization and nonlinear activation layers as a basic block structure for stacking, and generates multi-scale mobile feature maps by adjusting the stride of the 3D convolution layer. That is, L+1 sets of feature maps are generated. Where l represents the resolution of the original resolution times, i represents the total number of multi-scale mobile feature maps. Except for the L+1 group which contains one feature map, the other groups contain two feature maps. The process of generating a feature map of each scale is called a stage. In the next step (feature decoding stage), starting from the feature map of the smallest size, it is gradually upsampled to generate a larger and more detailed feature map.

[0014] Step (2) Constructing a hierarchical feature fusion network

[0015] The input of this step consists of two parts: one is the multi-scale feature map obtained in step (1), and the other is the fixed medical image. First, the fixed medical image is passed through a common 3D convolutional layer to generate high-dimensional features. Secondly, and The splicing is performed in the channel dimension, and then the spliced ​​feature map is input into the 3D residual module to obtain the fusion feature After a 3D convolution module with a stride of 2, the fusion feature is obtained by downsampling by 2 times Then and Perform concatenation operation on the channel dimension and input it into the 3D residual module to obtain fusion features Similarly, except for the last feature map, repeat the above steps to obtain a multi-scale hierarchical fusion feature map. That is, S+1 sets of feature maps are generated. Where s represents the resolution of the original resolution. times, i represents the total number of multi-scale fusion feature maps. Except for the S+1 group which contains 3 feature maps, other groups contain two feature maps. The last fusion feature map Yes The input is obtained by the residual linear self-attention module. In the next step (feature decoding stage), starting from the smallest size feature map, it is gradually upsampled to generate a larger size and more detailed fusion feature map.

[0016] Step (3) Construct feature decoder

[0017] This step consists of two parts: one is to build a feature decoder for mobile medical images, and the other is to build a fusion feature decoder for the network branch where the fixed medical image is located.

[0018] The multi-scale feature extraction network of mobile medical images is symmetrical with its feature decoding network. The specific steps are as follows: First, the minimum size mobile feature map Upsampled by 2x by trilinear interpolation, followed by a 3D convolution, and then combined with the mobile feature map The concatenation is performed on the channel dimension, and finally a 3D convolution module is used to obtain the decoding features on this scale. Similarly, repeating the above steps can obtain multi-scale mobile decoding feature maps Where l means the resolution is the original resolution The final obtained motion decoding features are used in the subsequent estimation of the deformation field.

[0019] The hierarchical feature fusion network of the network branch where the medical image is fixed is symmetrical with its fusion feature decoder. The specific steps are as follows: First, the fusion feature map of the smallest size and concatenated in the channel dimension, then passed through a 3D residual convolution module, and then Concatenate on the channel dimension and obtain the decoding features on this scale through a 3D residual convolution module Will The input is sent to the upsampling module (including trilinear interpolation and a 3D convolution) for 2x upsampling, and then the above steps are repeated to obtain multi-scale hierarchical fusion decoding features. Where s means the resolution is the original resolution The final fused decoding features are used in the subsequent estimation of the deformation field.

[0020] Step (4) Prediction of deformation field

[0021] The input of this step consists of three parts. The first part is the mobile decoding feature of the current stage obtained by step (3). The second part is the fused decoding feature of the current stage obtained by step (3). The third part is the feature obtained by the upsampling module through the fused decoding feature of the previous stage obtained by step (3), that is, the intermediate feature of the fused decoding feature process of the current stage (except the hidden feature of the smallest scale), which is called the hidden feature. These three features are input into the gated update module to predict the deformation field of the current stage. The gated update module is improved from the gated recurrent unit and is a dual-input and single-output module. Specifically, the forget gate in the gated recurrent unit is removed, and the addition operation is replaced by the concatenation operation. The specific steps are as follows: first, and The concatenation is performed in the channel dimension as the current input of the gated update module, except for the smallest size feature map, denoted by H s for A feature map processed by an upsampling module including trilinear interpolation and 3D convolution is the current hidden state. Then the current input and the current hidden state are input to the gated update module to obtain the deformation field of the current scale. After obtaining the deformation field of the current scale, it is first upsampled by 2 times and then the deformation field is used to decode the motion features of the next stage. Distortion and deformation,The distortion and deformation operation is performed by the spatial transformer network STN, and then the above steps are repeated until the final deformation field is obtained.

[0022] Step (5) Constructing an unsupervised loss function

[0023] When training a neural network model, a loss function needs to be constructed. The present invention adopts an unsupervised loss function, and the registration loss L reg It consists of two parts: similarity and deformation field smoothing. Its purpose is to align the moving medical image with the fixed medical image while ensuring the smoothness of the deformation field. The similarity part uses the local normalized cross-correlation as the similarity metric, while the smoothness term is achieved through L2 regularization of the deformation field gradient. The specific calculation formula is:

[0024]

[0025] where LNCC stands for Local Normalized Cross Correlation, ω is its window size, and F and Represent the fixed medical image and the deformed moving medical image respectively. represents the deformation field gradient, and λ1 is a hyperparameter that balances similarity and smoothness.

[0026] Step (6) Network training

[0027] The entire network is implemented based on the Pytorch framework, using the Adam optimization algorithm, the batch size of the network training is set to 1, the initial learning rate is 0.0001, and the training is performed for a total of 80 epochs. The experiment was conducted on two public brain magnetic resonance imaging datasets (LPBA and Mindboggle), which are divided into training sets and test sets to evaluate the performance of the model.

[0028] After the training is completed, the mobile medical image and the fixed medical image are input into the overall network constructed in step (1) to step (4), the deformation field is output, and the deformation field is used to generate the deformed image.

[0029] Beneficial effects of the present invention: The present invention fully utilizes the advantages of the dual-branch network architecture. The multi-scale feature extraction network focuses on extracting mobile medical image features, and the hierarchical feature fusion network focuses on hierarchical feature fusion to better learn the mapping relationship. In the decoding stage, the prior information provided by the mobile decoding features is fully utilized. At the same time, the long-range dependency relationship between the registrations of adjacent stages is established through the gated update module, allowing the network to adaptively adjust the importance of the multi-scale deformation field, thereby improving the final registration performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The entire network structure diagram of the present invention is a dual-branch parallel network structure, the upper network branch is a multi-scale feature extraction network and a mobile feature decoding network for mobile medical images, and the lower network branch is a multi-scale feature fusion network and a fusion feature decoding network;

[0031] Figure 2 It is a hierarchical feature fusion module.

[0032] Figure 3 The detailed structure of the gated update module. is element-wise multiplication, Indicates concatenation in the channel dimension. Represents a deformation operation.

[0033] Figure 4 Example of a mobile medical image as input.

[0034] Figure 5 An example of a fixed medical image as input.

[0035] Figure 6 An example of the registration result.

[0036] Figure 7 is an example of a deformation field. DETAILED DESCRIPTION

[0037] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0038] A multi-scale deformation decomposition medical image registration method, the overall framework of the network is composed of a dual-branch parallel network. Figure 1 The upper network branch is a multi-scale feature extraction and feature decoding network for mobile medical images, which is an encoder-decoder structure; the multi-scale features extracted in the encoding stage will be used in the hierarchical feature fusion module, and the multi-scale features output in the decoding stage will be input into the gated update module to predict the deformation field. The lower network branch is a hierarchical feature fusion and fusion feature decoding network. The hierarchical feature fusion network focuses on the mapping relationship between mobile medical images and fixed medical images, and the features output in the fusion feature decoding stage are used to predict the deformation field. The details are as follows:

[0039] (1) Constructing a multi-scale feature extraction network

[0040] The input of this step is a mobile medical image (such as Figure 4 As shown in Figure 2, the mobile medical image is fed into a pyramid feature extraction network, which is composed of 9 stacked 3D convolution blocks. Specifically, the 3D convolution block includes a 3D convolution, instance normalization, and a LeakyReLU activation function with a negative slope of 0.2. By adjusting the stride of the convolution layer, 5 groups (9 in total) of mobile medical image features with sizes of 1 / 1, 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image are generated. in and The number of channels of these nine mobile medical image feature maps are 8, 8, 8, 16, 16, 32, 32, 32, and 32, respectively.

[0041] (2) Constructing a hierarchical feature fusion network

[0042] The input of this step consists of two parts: one is the multi-scale moving medical image features obtained in step (1), and the other is the fixed medical image features. Figure 5 As shown) 3D convolution generates high-dimensional features Secondly, and Feed to Figure 2 The hierarchical feature fusion module shown above obtains fused features (FusedFeatures) Among them, the hierarchical feature fusion module consists of a residual block including group normalization, Swish activation function and 3D convolution. After a 3D convolution module with a stride of 2, the fusion feature is obtained by downsampling by 2 times Then and Input to the hierarchical feature fusion module to obtain fused features Similarly, repeat the above steps to obtain and In order to further achieve deep integration, Input to the residual linear self-attention module to obtain Finally, 4 groups (9 in total) of fused feature maps with sizes of 1 / 1, 1 / 2, 1 / 4, and 1 / 8 of the input image are obtained. in and The number of channels of these nine fusion feature maps is the same as the number of channels of the nine mobile medical image feature maps extracted by the multi-scale feature extraction network.

[0043] (3) Constructing feature decoder

[0044] This step consists of two parts: one is to build a feature decoder for mobile medical images, and the other is to build a fusion feature decoder for the network branch where the fixed medical image is located.

[0045] The multi-scale feature extraction network of mobile medical images is symmetrical with its feature decoding network. The specific steps are as follows: first, the minimum size mobile feature map Upsampled by 2x by trilinear interpolation, followed by a 3D convolution, and then combined with the mobile feature map The concatenation is performed on the channel dimension, and finally a 3D convolution module is used to obtain the decoding features on this scale. Similarly, repeat the above steps to obtain and The final four sets of motion decoding features are used for subsequent deformation field estimation.

[0046] The hierarchical feature fusion network of the network branch where the medical image is fixed and its fusion feature decoder are combined to form an encoder-decoder structure. The specific steps are as follows: First, the minimum size fusion feature map and are input into the hierarchical feature fusion module and then combined with Input together into the hierarchical feature fusion module to obtain the decoding features at this scale Will Input to the upsampling module (including trilinear interpolation and a 3D convolution) for 2 times upsampling, and then repeat to obtain The steps can be obtained Similarly, repeat the above steps to obtain and The final four sets of fused decoding features are used in the subsequent deformation field estimation.

[0047] (4) Deformation field prediction

[0048] The input of this step consists of three parts. The first part is the mobile decoding feature of the current stage, the second part is the fusion decoding feature of the current stage, and the third part is the feature obtained by the upsampling module of the fusion decoding feature of the previous stage obtained in step (3), that is, the intermediate feature of the fusion decoding feature process of the current stage (except the hidden feature of the smallest scale), which is called the hidden feature. These three features are input into Figure 3 The gated update module shown in the figure is used to predict the deformation field of the current stage. The gated update module is improved from the gated recurrent unit and is a dual-input and single-output module. Specifically, the forget gate in the gated recurrent unit is removed, and the gating mechanism is implemented using convolution and Sigmoid functions. At the same time, the addition operation is replaced by the concatenation operation, because the concatenation operation retains the original semantic information, effectively avoiding the semantic confusion of the current input and hidden state, and is more conducive to outputting the deformation field of the current scale through convolution. The specific steps are as follows: First, and Concatenate in the channel dimension as the current input of the gated update module, record H 3 , H 3 is the current hidden state, and then the current input and the current hidden state are input into the gated update module to obtain the deformation field of the current scale By trilinear interpolation Perform 2x upsampling and use the deformation field to decode the motion features of the next stage Warp operation is performed by STN (spatial transformer network), denoted by H 2 for A feature map after processing by an upsampling module including trilinear interpolation and 3D convolution, that is, the current hidden state, and then repeat the steps of the previous stage to obtain the deformation field of this stage Except for the lowest size, the deformation fields predicted in other stages must be fused with the deformation fields predicted in the previous stage. Specifically, the deformation fields are first Perform 2x upsampling, and then use the deformation field to predict the deformation field of the current stage Perform a warp operation, and finally add the two deformation fields to obtain the fused deformation field. Similarly, repeat the above steps until the final deformation field is obtained.

[0049] (5) Constructing an unsupervised loss function

[0050] When training a neural network model, a loss function needs to be constructed. The present invention adopts an unsupervised loss function, and the registration loss L reg It consists of two parts: similarity and deformation field smoothing. Its purpose is to align the moving medical image with the fixed medical image while ensuring the smoothness of the deformation field. The similarity part uses the local normalized cross-correlation as the similarity metric, while the smoothness term is achieved through L2 regularization of the deformation field gradient. The specific calculation formula is:

[0051]

[0052] where LNCC stands for Local Normalized Cross Correlation, ω is its window size, and F and Represent the fixed medical image and the deformed moving medical image respectively. represents the deformation field gradient, and λ1 is a hyperparameter that balances similarity and smoothness.

[0053] (6) Network training

[0054] The entire network is implemented based on the PyTorch framework, using the Adam optimization algorithm. The batch size of network training is set to 1, the initial learning rate is 0.0001, and the training is performed for a total of 80 epochs. The experiments were conducted on two public brain magnetic resonance imaging datasets, LPBA and Mindboggle. The LPBA40 dataset contains a total of 40 MRI volumes, each MRI volume contains 54 manually labeled regions of interest, 30 volumes (30×29 pairs) are used for training, and 10 volumes (10×9 pairs) are used for testing. Each MRI volume in the Mindboggle dataset contains 62 manually labeled regions of interest, of which 42 volumes (42×41 pairs) are used for training, and 20 volumes (20×19 pairs) are used for testing. After training is completed, the deformation field can be predicted, and the model will Figure 4 and Figure 5 They are input into the network as mobile medical images and fixed medical images respectively, and the output Figure 7 The deformation field shown, then the deformation field acts on Figure 4 Get the deformed image, such as Figure 6 As shown, medical image registration is completed.

Claims

1. A multi-scale deformation decomposition medical image registration method, characterized in that: Here are the steps: Step (1) Construct a multi-scale feature extraction network The mobile medical image is fed into a pyramid feature extraction network to obtain mobile feature maps of different scales; the pyramid feature extraction network uses 3D convolution, instance normalization and nonlinear activation layers as a basic block structure to stack, and generates multi-scale mobile feature maps by adjusting the stride of the 3D convolution layer That is, L+1 groups of feature maps are generated; Where l means the resolution is the original resolution times, i represents the total number of multi-scale mobile feature maps. Except for the L+1th group which contains one feature map, other groups contain two feature maps. The process of generating a feature map of each scale is called a stage. Step (2) Constructing a hierarchical feature fusion network The input of this step consists of two parts: one is the multi-scale feature map obtained in step (1), and the other is the fixed medical image. First, the fixed medical image is processed through a common 3D convolutional layer to generate high-dimensional features. Secondly, and The splicing is performed in the channel dimension, and then the spliced ​​feature map is input into the 3D residual module to obtain the fusion feature After a 3D convolution module with a stride of 2, the fusion feature is obtained by downsampling by 2 times Then and Perform concatenation operation on the channel dimension and input it into the 3D residual module to obtain fusion features Similarly, except for the last feature map, repeat the above steps to obtain a multi-scale hierarchical fusion feature map. That is, S+1 groups of feature maps are generated; Where s means the resolution is the original resolution times, i represents the total number of multi-scale fusion feature maps. Except for the S+1 group which contains 3 feature maps, other groups contain two feature maps. The last fusion feature map Yes Input to the residual linear self-attention module; Step (3) Construct feature decoder This step consists of two parts: one is to build a feature decoder for mobile medical images, and the other is to build a fusion feature decoder for the network branch where the fixed medical image is located; The multi-scale feature extraction network of mobile medical images is symmetrical with its feature decoding network. The specific steps are as follows: First, the minimum size mobile feature map Upsampled by 2x by trilinear interpolation, followed by a 3D convolution, and then combined with the mobile feature map The concatenation is performed on the channel dimension, and finally a 3D convolution module is used to obtain the decoding features on this scale. Similarly, repeat the above steps to obtain multi-scale mobile decoding feature maps Where l means the resolution is the original resolution times; the finally obtained motion decoding features participate in the subsequent estimation of the deformation field; The hierarchical feature fusion network of the network branch where the medical image is fixed is symmetrical with its fusion feature decoder. The specific steps are as follows: First, the fusion feature map of the smallest size and concatenated in the channel dimension, then passed through a 3D residual convolution module, and then Concatenate on the channel dimension and obtain the decoding features on this scale through a 3D residual convolution module Will The input is sent to the upsampling module for 2x upsampling. The upsampling module includes trilinear interpolation and a 3D convolution. Then the above steps are repeated to obtain multi-scale hierarchical fusion decoding features. Where s means the resolution is the original resolution times; the final fusion decoding features are used to participate in the subsequent estimation of the deformation field; Step (4) Prediction of deformation field The input of this step consists of three parts. The first part is the mobile decoding feature of the current stage obtained by step (3). The second part is the fused decoding feature of the current stage obtained by step (3). The third part is the feature obtained by the upsampling module of the fused decoding feature of the previous stage obtained by step (3). That is, in addition to the hidden feature of the minimum scale, the intermediate feature of the fused decoding feature process of the current stage is obtained. The intermediate feature is called the hidden feature. These three features are input into the gated update module to predict the deformation field at the current stage. The gated update module is improved from the gated recurrent unit and is a dual-input and single-output module. Specifically, the forget gate in the gated recurrent unit is removed, and the addition operation is replaced by the concatenation operation. The specific steps are as follows: First, and The concatenation is performed in the channel dimension as the current input of the gated update module, except for the smallest size feature map, denoted by H s for A feature map processed by an upsampling module including trilinear interpolation and 3D convolution is the current hidden state. Then the current input and the current hidden state are input to the gated update module to obtain the deformation field of the current scale. After obtaining the deformation field of the current scale, it is first upsampled by 2 times and then the deformation field is used to decode the motion features of the next stage. Distortion, the distortion operation is completed by the spatial transformer network STN, and then the above steps are repeated until the final deformation field is obtained; Step (5) Constructing an unsupervised loss function When training a neural network model, it is necessary to construct a loss function. An unsupervised loss function is used, and the registration loss L reg It consists of similarity and deformation field smoothing. The specific calculation formula is: where LNCC stands for Local Normalized Cross Correlation, ω is its window size, and F and Represent the fixed medical image and the deformed moving medical image respectively. represents the deformation field gradient, λ1 is a hyperparameter that balances similarity and smoothness; Step (6) Network training The entire network is implemented based on the Pytorch framework and uses the Adam optimization algorithm. The experiments are conducted on two public brain MRI datasets, LPBA and Mindboggle. The datasets are divided into training sets and test sets to evaluate the performance of the model. After the training is completed, the mobile medical image and the fixed medical image are input into the overall network constructed in step (1) to step (4), the deformation field is output, and the deformation field is used to generate the deformed image.

2. The medical image registration method based on multi-scale deformation decomposition according to claim 1, characterized in that: In step (6), the batch size of network training is set to 1, the initial learning rate is 0.0001, and the training is performed for a total of 80 epochs.

Images