Medical image registration method and device based on three-dimensional lightweight convolutional neural network

Abstract

The application discloses a medical image registration method and device based on a three-dimensional lightweight convolutional neural network, which inputs a three-dimensional medical image to be registered and a target three-dimensional medical image into the three-dimensional lightweight convolutional neural network, constructs a lightweight feature fusion module based on a cross-channel low-dimensional convolution kernel after the original image is split into three-dimensional images with smaller sizes, embeds the lightweight feature fusion module as a basic processing unit into a backbone network based on a three-dimensional residual dense connection structure, lets the network learn a deformation field between input images to realize three-dimensional image registration. The application uses a high-precision three-dimensional residual dense network and an efficient lightweight feature fusion module, realizes fast and accurate three-dimensional medical image registration with a small amount of parameters, and can effectively reduce the demand for the data volume of medical images and the time cost required in actual tests.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application and artificial intelligence, and in particular, to a medical image registration method and device based on a three-dimensional lightweight convolutional neural network. BACKGROUND

[0002] Medical image registration is a process of combining two or more medical images based on image representation information to obtain more information. As an indispensable preprocessing step in other advanced medical image processing tasks, medical image registration seeks to find a suitable spatial deformation field to best align the underlying anatomical structures of the images to be registered and the target image.

[0003] Compared with supervised deformation field prediction methods, unsupervised methods can effectively overcome the problem of lack of training data. In previous image registration methods based on unsupervised learning, a processing method mainly composed of a convolutional neural network and a spatial transformation function has been basically formed. For example, Balakrishnan et al. proposed a deep learning model VoxelMorph for large-scale distortion of the entire three-dimensional image. Later, inspired by FlowNet and spatial transformer networks (STN), Zhao et al. proposed an end-to-end convolutional neural network model to perform voxel-level three-dimensional medical image registration, called volume tweening network (VTN). Based on the cascade and deformation system established by VTN, Zhao et al. also proposed a recursive cascaded network (RCN) for unsupervised image registration. In existing methods, the network based on a single model often has fewer layers, so the performance is poor when registering two images with large deformation amplitudes. In the method based on the cascade model, how to determine the number of cascade layers is a difficult problem, and the computational cost is very high. In addition, training these models with a large number of hyperparameters requires a large amount of medical image data.

[0004] In summary, the existing medical image registration technology based on deep learning only considers the improvement of accuracy and does not consider the use of lightweight technology to reduce the medical image data required for training the network model and the time cost required for processing registration, so it is difficult to apply in actual clinical tasks. SUMMARY

[0005] The present application aims at the deficiencies of the prior art and proposes a medical image registration method and device based on a three-dimensional lightweight convolutional neural network considering the actual application.

[0006] The purpose of the application is achieved by the following technical solutions: in a second aspect, the application provides a medical image registration method based on a three-dimensional lightweight convolutional neural network, comprising the following steps:

[0007] Step one: obtain the three-dimensional medical images to be registered and target three-dimensional medical images, and pre-process the images;

[0008] Step two: construct a three-dimensional lightweight convolutional neural network, whose backbone model is a three-dimensional residual dense connection network, that is, all standard two-dimensional convolutions in the original residual dense connection network are replaced by standard three-dimensional convolutions, and a lightweight feature fusion module based on cross-channel low-dimensional convolution kernels is constructed, feature fusion and channel merging are performed using cross-channel two-dimensional convolution kernels, the number of feature maps is adjusted, and the lightweight feature fusion module is used as a basic convolution processing unit to replace all standard three-dimensional convolutions except the first convolution layer and the last convolution layer of the backbone model;

[0009] Step three: input the three-dimensional medical images to be registered and the target three-dimensional medical images into the three-dimensional lightweight convolutional neural network; learn and predict the three-dimensional deformation field between the three-dimensional medical images to be registered and the target three-dimensional medical images using the three-dimensional lightweight convolutional neural network established in step two;

[0010] Step four: based on the three-dimensional deformation field, perform non-rigid deformation on the images to be registered to achieve medical image registration.

[0011] Further, in step one, the pre-processing includes image gray value truncation, adjusting the gray value range to 0-255, and a rigid registration process.

[0012] Further, the lightweight feature fusion module specifically implements the following feature fusion process:

[0013] 1) using point-by-point convolution, the number of feature maps S output by the previous layer is reduced to μS, μ is the reduction coefficient;

[0014] 2) using a cross-channel two-dimensional convolution kernel, all feature maps output by step 1) are fused in the X direction;

[0015] 3) using a cross-channel two-dimensional convolution kernel, all feature maps output by step 1) are fused in the Y direction;

[0016] 4) using a cross-channel two-dimensional convolution kernel, all feature maps output by step 1) are fused in the Z direction;

[0017] 5) using a cross-channel two-dimensional convolution kernel, all feature maps output by step 2) are fused in the X direction;

[0018] 6) Channel merging of the outputs of step 2) and step 3), and then feature fusion in the Y direction of all feature maps by using a two-dimensional convolution kernel across channels;

[0019] 7) Channel merging of the outputs of step 2), step 3), and step 4), and then feature fusion in the Z direction of all feature maps by using a two-dimensional convolution kernel across channels;

[0020] 8) Channel merging of the outputs of step 5), step 6), and step 7), and then feature fusion in the X direction of all feature maps by using a two-dimensional convolution kernel across channels;

[0021] 9) Channel merging of the outputs of step 5), step 6), and step 7), and then feature fusion in the Y direction of all feature maps by using a two-dimensional convolution kernel across channels;

[0022] 10) Channel merging of the outputs of step 5), step 6), and step 7), and then feature fusion in the Z direction of all feature maps by using a two-dimensional convolution kernel across channels;

[0023] 11) Channel merging of the outputs of step 8), step 9), and step 10), and then adjusting the number of feature maps to T by using point-wise convolution, and by default, T is equal to S.

[0024] Further, the reduction factor μ in the point-wise convolution ranges from 0 to 1, excluding 0 and 1, and μS is rounded.

[0025] Further, the three-dimensional lightweight convolutional neural network further comprises a whole decomposition module and a pixel shuffling module, the whole decomposition module is used to split the original image into eight three-dimensional images with a length, a width, and a thickness of half of the original image, and then input to the backbone model, the whole decomposition operation can be regarded as a rearrangement of the pixel points in the digital image, which can reduce the GPU memory occupied during training the model and retain the brightness information of the original image; the pixel shuffling module is the inverse process of the whole decomposition module and is connected after the backbone model, which is used to splice the eight three-dimensional images with the same size into a three-dimensional image with a length, a width, and a thickness of twice the original size, and the pixel shuffling operation can ensure that the size of the deformation field is equal to the input image, and the operation does not increase the model parameters, avoiding increasing the calculation time.

[0026] In a second aspect, the application further provides a medical image registration device based on a three-dimensional lightweight convolutional neural network, comprising a memory and one or more processors, the memory stores executable code, and the processor executes the executable code to realize the medical image registration method based on the three-dimensional lightweight convolutional neural network.

[0027] In a third aspect, the present application further provides a computer readable storage medium, which stores a program, and the program is executed by a processor to implement the medical image registration method based on the three-dimensional lightweight convolutional neural network.

[0028] The present application has the advantages of high efficiency and high precision in three-dimensional medical image registration, which can reduce the demand for GPU memory on the one hand, and greatly reduce the parameter amount of the model on the other hand, thereby reducing the demand for the number of training samples and reducing the cost of three-dimensional medical image registration based on deep learning method. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 It is a three-dimensional lightweight convolutional neural network model framework diagram;

[0030] Figure 2 It is a lightweight feature fusion module schematic diagram;

[0031] Figure 3 It is a medical image registration process schematic diagram based on three-dimensional lightweight convolutional neural network;

[0032] Figure 4 It is a three-dimensional image decomposition operation and pixel shuffling schematic diagram;

[0033] Figure 5 It is a structure diagram of a medical image registration device based on three-dimensional lightweight convolutional neural network provided by the present application. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0035] The present application provides a highly efficient and practical method based on unsupervised deep learning and three-dimensional lightweight convolutional neural network. A three-dimensional residual dense network is selected as the backbone network and combined with an overall decomposition layer, which makes it possible to train a very deep three-dimensional convolutional neural network on a single block graphics card (24GB of video memory). In addition, according to the existing lightweight method and model acceleration method, a lightweight module based on low-dimensional convolution kernel is designed as the basic unit of feature extraction and feature fusion in the model, replacing the standard three-dimensional convolution operation in the traditional three-dimensional convolutional neural network, which is used to accelerate medical image registration and reduce the required training samples. Specifically, the following steps are included:

[0036] Step one: obtaining the three-dimensional medical images to be registered and the fixed images (i.e. the target three-dimensional medical images for registration), and preprocessing (including gray value truncation, adjusting the gray value range to 0-255, and rigid registration);

[0037] Step two: using a two-dimensional residual dense connection network as the backbone model, the model includes two three-dimensional convolutional layers (using 3x3x3 convolutional kernels) for preliminary extraction of image features, five three-dimensional residual dense connection modules for further extraction of image potential features, a channel merging layer for merging the feature maps of each three-dimensional residual dense connection module, a three-dimensional convolutional layer (using 3x3x3 convolutional kernels) for adjusting the number of feature maps, a residual connection layer (i.e. adding the feature maps of the previous layer to the feature maps of the first three-dimensional convolutional layer) for reducing the training difficulty, an upsampling layer for enlarging the size of the feature maps (improving the spatial resolution). In the three-dimensional residual dense connection module, the feature maps will pass through 1 standard three-dimensional convolutional layer and 1 rectified linear unit (ReLU), and the output feature maps are merged with the previous output feature maps and then enter the next standard three-dimensional convolutional layer and rectified linear unit. This operation needs 6 times, and then all the output feature maps of 6 times are merged in the channel, using 1x1x1 convolutional kernel, adjusting the number of feature maps to the same as the input feature maps of the three-dimensional residual dense connection module, and adding the input feature maps of the three-dimensional residual dense connection module as the output of the three-dimensional residual dense connection module;

[0038] Step three: replacing all standard two-dimensional convolutions of the backbone model with standard three-dimensional convolutions, then constructing a lightweight feature fusion module based on cross-channel low-dimensional convolution kernels, and replacing all standard three-dimensional convolutions of the backbone model except the first convolutional layer and the last convolutional layer with the lightweight feature fusion module as the basic convolution processing unit to obtain the final backbone model;

[0039] Step four: inputting the three-dimensional medical image to be registered and the fixed image into the three-dimensional lightweight convolutional neural network, which includes the overall decomposition module, the backbone model and the pixel shuffling module;

[0040] Step five: decomposing the three-dimensional medical image to be registered and the fixed image through the overall decomposition module respectively (i.e. rearranging the pixels of the digital image), so that the two images are respectively divided into eight three-dimensional images with half the length, width and thickness of the original image;

[0041] Step five: using the backbone model in the three-dimensional lightweight convolutional neural network established in step two to perform feature fusion of the images and improving the spatial resolution through pixel shuffling to obtain the predicted three-dimensional deformation field between the three-dimensional medical image to be registered and the fixed image;

[0042] Step six: performing non-rigid deformation on the image to be registered based on the deformation field to realize medical image registration.

[0043] Further, as Figure 2As shown, the lightweight feature fusion module specifically implements the following feature fusion process:

[0044] (1.1) The number of feature maps output by the previous layer is reduced from S to R (R is an integer greater than 0 and less than S) by using point-by-point convolution (R 1x1x1xS convolution kernels), so as to reduce the model memory occupation, inference time and training difficulty; the reduction factor μ of the point-by-point convolution ranges from 0 to 1, and does not include 0 and 1, and μS is rounded off;

[0045] (1.2) The X direction of all feature maps output by (1.1) is convolved by using a cross-channel two-dimensional convolution kernel (R 3x1x1xR convolution kernel), so as to complete the output of R feature maps;

[0046] (1.3) The Y direction of all feature maps output by (1.1) is convolved by using a cross-channel two-dimensional convolution kernel (R 1x3x1xR convolution kernel), so as to output R feature maps;

[0047] (1.4) The Z direction of all feature maps output by (1.1) is convolved by using a cross-channel two-dimensional convolution kernel (R 1x1x3xR convolution kernel), so as to output R feature maps;

[0048] (1.5) In order to expand the receptive field in the X direction, the X direction of all feature maps output by (1.2) is convolved by using a cross-channel two-dimensional convolution kernel (R 3x1x1xR convolution kernel), so as to output R feature maps;

[0049] (1.6) In order to expand the receptive field in the Y direction, the outputs of (1.2) and (1.3) are merged in the channel, and then a cross-channel two-dimensional convolution kernel (R 1x3x1x2R convolution kernel) is used to convolve the Y direction of all feature maps, so as to output R feature maps;

[0050] (1.7) In order to expand the receptive field in the Z direction, the outputs of (1.2), (1.3) and (1.4) are merged in the channel, and then a cross-channel two-dimensional convolution kernel (R 1x1x3x3R convolution kernel) is used to convolve the Z direction of all feature maps, so as to output R feature maps;

[0051] (1.8) In order to further expand the receptive field in the X direction, the outputs of (1.5), (1.6) and (1.7) are merged in the channel, and then a cross-channel two-dimensional convolution kernel (R 3x1x1x3R convolution kernel) is used to convolve the X direction of all feature maps, so as to output R feature maps;

[0052] (1.9) To further expand the receptive field in the Y direction, the outputs of (1.5), (1.6), and (1.7) are channel-merged, and then a cross-channel two-dimensional convolution kernel (R 1×3×1×3R convolution kernels) is used to perform convolution operations on the Y direction of all feature maps, and the number of output feature maps is R.

[0053] (1.10) To further expand the receptive field in the Z direction, the outputs of (1.5), (1.6), and (1.7) are channel-merged, and then a cross-channel two-dimensional convolution kernel (R 1×1×3×3R convolution kernels) is used to perform convolution operations on all feature maps in the Z direction, and the number of output feature maps is R.

[0054] (1.11) The outputs of (1.8), (1.9), and (1.10) are merged, and then the number of feature maps is adjusted to T by using pointwise convolution (T 1×1×1×R convolution kernels).

[0055] like Figure 1 The diagram shown illustrates the framework of the three-dimensional lightweight convolutional neural network model proposed in this invention. The fixed image and the image to be registered first undergo... Figure 4 (a) After overall decomposition, the resulting smaller image first extracts features using a standard 3D convolutional layer. Then, it passes through a lightweight module and several 3D residual dense connection modules to further extract latent features. Channel merging and compression (1×1×1 convolution) are performed, and the extracted features are added to the initial features. Finally, pixel shuffling is used to improve spatial resolution, yielding the predicted deformation field. Each 3D residual dense connection module consists of several lightweight modules, modified linear units, and channel merging and compression (1×1×1 convolution). This framework can extract feature maps from different receptive fields and fully integrate the features from each receptive field. Multiple dense connections and residual connections effectively reduce gradient vanishing caused by model depth and complexity.

[0056] like Figure 2As shown in the figure, the schematic diagram of the lightweight feature fusion module provided by the application is shown. From top to bottom, the convolution operations in the single lightweight feature fusion module are sequentially performed. Here, the number of input feature maps of the module is S, the number of feature maps is reduced to R by using a 1x1x1 convolution kernel, and then R Xx1x1, 1xYx1, and 1x1xZ convolution kernels are used respectively to perform cross-channel convolution in the X, Y, and Z directions on the feature maps, and the output feature maps are combined as shown in the figure and input to the subsequent network layer to perform cross-channel convolution in the X, Y, and Z directions. A 1x1x1 convolution kernel is used in the last layer to adjust the number of feature maps to the required number T. The module connects low-dimensional convolution layers in parallel, which can increase the receptive field and avoid training difficulties caused by too deep models. Assuming that "X-Y-Z" represents cross-channel convolution in the X direction, cross-channel convolution in the Y direction, and cross-channel convolution in the Z direction, the connection mode shown in the figure can simultaneously realize the following 10 connection modes: X-X-X, X-X-Y, X-X-Z, X-Y-Y, X-Y-Z, X-Z-Z, Y-Y-Y, Y-Y-Z, Y-Z-Z, and Z-Z-Z.

[0057] As shown in the figure, Figure 3 As shown in the figure, the schematic diagram of the medical image registration process based on the three-dimensional lightweight convolutional neural network is shown. The loss in the dashed box is the loss function required when training the network, wherein the similarity loss is the similarity between the deformed image after the deformed field learned by the neural network and the target image, and the smoothing loss is the gradient smoothing constraint term added to prevent the deformed image from being deformed locally.

[0058] As shown in the figure, Figure 4 As shown in the figure, Figure 4 (a) in the figure is a schematic diagram of a holistic decomposition operation, Figure 4 (b) in the figure is a schematic diagram of a pixel shuffling operation, that is, a schematic diagram of an inverse operation of holistic decomposition. Holistic decomposition can be regarded as rearranging and combining the voxels of a digital image as shown in the figure, so it does not need to be calculated, which can effectively avoid increasing the parameter quantity and inference time of the model.

[0059] Compared with rigid registration and deep learning methods in recent years, the method has higher accuracy, close processing speed and fewer model parameters in the problem of medical image registration. Table 1 is the three-dimensional image registration error test result on the liver tumor segmentation data set The Liver Tumor Segmentation Benchmark (LiTS). In the experiment of Table 1, the method adopts 10 three-dimensional residual dense connection modules, the channel number of the standard 3D convolution is 32, the output channel of the last layer of each lightweight feature fusion module is 32, and the output channel of the middle layer of each lightweight feature fusion module is 16. In this experiment, the DICE coefficient, the average surface distance and the Jaccard similarity coefficient are used to measure the registration accuracy of the registered image and the target image by measuring the overlapping degree of the liver part before and after registration. From the table, it can be seen that the method proposed in the application has obvious improvement in accuracy compared with the results obtained by rigid registration and deep learning-based methods (Voxel Morph, VTN and RCN), which proves the effectiveness of the method.

[0060] Table 1 Comparison of three-dimensional medical image registration algorithm error

[0061]

[0062] Corresponding to the foregoing embodiment of the medical image registration method based on the three-dimensional lightweight convolutional neural network, the application also provides an embodiment of a medical image registration device based on the three-dimensional lightweight convolutional neural network.

[0063] Referring to Figure 5 The medical image registration device based on the three-dimensional lightweight convolutional neural network provided in the embodiment of the application includes a memory and one or more processors, the memory stores executable code, and the processor executes the executable code to implement the medical image registration method based on the three-dimensional lightweight convolutional neural network in the foregoing embodiment.

[0064] The embodiment of the medical image registration device based on the three-dimensional lightweight convolutional neural network of the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, hardware or a combination of software and hardware. Taking software realization as an example, as a logically meaningful device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability where it is located. From the hardware level, as shown in Figure 5 As shown in the figure, it is a hardware structure diagram of the device with data processing capability where the medical image registration device based on the three-dimensional lightweight convolutional neural network of the application is located, exceptFigure 5 In addition to the processor, the memory, the network interface, and the nonvolatile memory, any data processing capable device in which the apparatus of the embodiments is located can also include other hardware according to the actual functions of the data processing capable device, which will not be described here.

[0065] The implementation processes of the functions and roles of the units in the apparatus are specifically described in the implementation processes of the corresponding steps in the above method, which will not be described here.

[0066] For the apparatus embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The apparatus embodiment described above is only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Some or all of the modules can be selected to achieve the purpose of the present application according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0067] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the medical image registration method based on the three-dimensional lightweight convolutional neural network in the above embodiment.

[0068] The computer readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of any data processing capable device. The computer readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0069] The above embodiments are used to explain and illustrate the present application, but not to limit the present application. Any modification and change made to the present application within the spirit and protection scope of the claims of the present application, falls into the protection scope of the present application.

Claims

1. A medical image registration method based on a three-dimensional lightweight convolutional neural network, characterized in that, Includes the following steps: Step 1: Acquire the 3D medical image to be registered and the target 3D medical image, and preprocess the images; Step 2: Construct a 3D lightweight convolutional neural network. Its backbone model is a 3D residual dense connection network. Replace all the standard 2D convolutions in the original residual dense connection network with standard 3D convolutions. Then, construct a lightweight feature fusion module based on cross-channel low-dimensional convolution kernels. Use cross-channel 2D convolution kernels to perform feature fusion and channel merging, adjust the number of feature maps, and use the lightweight feature fusion module as the basic convolutional processing unit to replace all standard 3D convolutions except the first and last convolutional layers of the backbone model. The lightweight feature fusion module specifically implements the following feature fusion process: 1) By using pointwise convolution, the number of feature maps S output by the previous layer is reduced to μS, where μ is the reduction factor; 2) Using cross-channel 2D convolution kernels, feature fusion is performed on the X-direction of all feature maps output in step 1); 3) Using cross-channel 2D convolution kernels, perform feature fusion on the Y direction of all feature maps output in step 1); 4) Using cross-channel 2D convolution kernels, perform feature fusion on the Z-direction of all feature maps output in step 1); 5) Using cross-channel 2D convolution kernels, perform feature fusion on the X-direction of all feature maps output in step 2); 6) Merge the outputs of steps 2) and 3) by channel merging, and then use cross-channel 2D convolution kernels to perform feature fusion on the Y direction of all feature maps; 7) Merge the outputs of steps 2), 3), and 4) into channels, and then use cross-channel 2D convolution kernels to fuse features in the Z direction of all feature maps. 8) Merge the outputs of steps 5), 6), and 7) into channels, and then use cross-channel 2D convolution kernels to fuse features in the X direction of all feature maps. 9) Merge the outputs of steps 5), 6), and 7) into channels, and then use cross-channel 2D convolution kernels to fuse features in the Y direction of all feature maps. 10) Merge the outputs of steps 5), 6), and 7) by channel merging, and then use cross-channel 2D convolution kernels to fuse features in the Z direction of all feature maps. 11) Merge the channels of the outputs from steps 8), 9), and 10), and then use pointwise convolution to adjust the number of feature maps to T, which is equal to S by default; Step 3: Input the 3D medical image to be registered and the target 3D medical image into a 3D lightweight convolutional neural network; use the 3D lightweight convolutional neural network established in Step 2 to learn and predict the 3D deformation field between the 3D medical image to be registered and the target 3D medical image; Step 4: Perform non-rigid deformation on the image to be registered based on the three-dimensional deformation field to achieve medical image registration.

2. The medical image registration method based on a three-dimensional lightweight convolutional neural network according to claim 1, characterized in that, In step one, the preprocessing includes truncating the grayscale values ​​of the image, adjusting the grayscale value range to 0-255, and performing rigid registration.

3. The medical image registration method based on a three-dimensional lightweight convolutional neural network according to claim 1, characterized in that, The reduction factor μ during pointwise convolution is in the range of 0-1, excluding 0 and 1, and μS is rounded to the nearest integer.

4. The medical image registration method based on a three-dimensional lightweight convolutional neural network according to claim 1, characterized in that, The lightweight 3D convolutional neural network also includes a global decomposition module and a pixel shuffling module. The global decomposition module is used to split the original image into eight 3D images with lengths, widths, and thicknesses equal to half that of the original image, and then input them into the backbone model. The global decomposition operation can be regarded as a rearrangement of pixels in a digital image. This operation can reduce the GPU memory occupied during model training and retain the brightness information of the original image. The pixel shuffling module is the reverse process of the global decomposition module and is connected after the backbone model. It is used to stitch eight 3D images of the same size into a 3D image with twice the length, width, and thickness of the original. The pixel shuffling operation can ensure that the size of the deformation field is equal to that of the input image. This operation does not increase model parameters and avoids increasing computation time.

5. A medical image registration device based on a three-dimensional lightweight convolutional neural network, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that... When the processor executes the executable code, it implements the medical image registration method based on a three-dimensional lightweight convolutional neural network as described in any one of claims 1-4.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the medical image registration method based on a three-dimensional lightweight convolutional neural network as described in any one of claims 1-4.

Images