Magnetic resonance image registration method based on adaptive regularization

Abstract

The invention discloses a magnetic resonance image registration method based on adaptive regularization. The method comprises the following steps: constructing a training set and a test set; building a self-adaptive regularization registration network; the floating image and the fixed image in the training set are spliced in a channel dimension and then input into the student registration network for bidirectional registration operation, floating image domain appearance disturbance is calculated, and a spatial transformation uncertainty graph is calculated according to a predicted deformation field output by the teacher registration network; and training the adaptive regularization registration network. According to the method, the spatial transformation uncertainty graph is obtained by adding the appearance disturbance of the floating image domain in the teacher registration network, and the weight of the regularization item in the loss function is automatically adjusted by using the spatial transformation uncertainty, so that the adaptive regularization is realized, and the magnetic resonance image registration precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic resonance image registration, and more specifically, to a magnetic resonance image registration method based on adaptive regularization, which is used for registering magnetic resonance images. Background Art

[0002] Magnetic Resonance Imaging (MRI) is a non-invasive, non-ionizing radiation-free imaging technology, and is a physical phenomenon of nuclear magnetic resonance. It uses radio frequency pulses to excite atomic nuclei with non-zero spins placed in a magnetic field. After the radio frequency pulse stops, the atomic nuclei relax, and during the relaxation process, the induction coil is used to collect signals, and a mathematical image is reconstructed according to a certain mathematical method. Compared with other imaging technologies, magnetic resonance imaging can accurately and clearly display the anatomical structure of the imaging site.

[0003] Image registration refers to solving the spatial transformation of two different images so that their corresponding points have spatial consistency. Usually, the image that needs to be spatially transformed is called a floating image, and the image that remains unchanged is called a fixed image. Specifically, image registration refers to the floating image after spatial transformation. After that, it can be spatially aligned with the fixed image, and its corresponding points have the same spatial position. Image registration can maximize the fusion of medical images of different modalities or time and improve information utilization.

[0004] Traditional magnetic resonance image registration methods have shown certain limitations, and generally have shortcomings such as complex calculations, long time consumption, and poor universality. Deep learning technology has been widely cited in the field of magnetic resonance image registration for its powerful learning ability and feature extraction capabilities. Among them, the Voxelmorph method [Balakrishnan G, Zhao A, Sabuncu M R. IEEE Trans Med Imaging, 2019, 38: 1788.] trains the registration network by maximizing the similarity between the deformed image and the reference image, avoiding the need for the gold standard deformation field for image registration, and realizing an end-to-end unsupervised registration model for 3D magnetic resonance images. However, in most deep learning registration methods, the deformation field regularization weight λ of all image pairs to be registered is set to a fixed value. This method of fixing weight parameters may not necessarily obtain the optimal solution, because the regularization strength of a specific image pair should be related to their image content and degree of alignment [Zhe X, Jie L, Donghuan L. MICCAI, 2022, pp: 14-24]. Therefore, it is necessary to find a suitable adaptive regularization weight method to improve the registration accuracy and increase the universality of the network model. Summary of the invention

[0005] The purpose of the present invention is to address the above-mentioned defects in the prior art and propose a magnetic resonance image registration method based on adaptive regularization. A teacher registration network-student registration network is used as the backbone network in the registration network. The spatial transformation uncertainty map is obtained by adding appearance perturbations in the floating image domain to the teacher registration network. The weight of the regularization term in the loss function is automatically adjusted using this spatial transformation uncertainty to achieve adaptive regularization, thereby improving the accuracy of magnetic resonance image registration.

[0006] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0007] A magnetic resonance image registration method based on adaptive regularization comprises the following steps:

[0008] Step 1: Preprocess the original MRI images with segmentation labels and select one of the original MRI images as the fixed image I F , the remaining original MRI image is used as the floating image I M , according to the floating image I M Construct training and test sets;

[0009] Step 2: Build an adaptive regularized registration network, which includes a teacher registration network, a spatial transformation network, and a student registration network;

[0010] Step 3: Floating images in the training set I M With fixed image I F After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation to obtain the positive predicted deformation field and the inverse predicted deformation field Forward-based predicted deformation field and floating image I M , get the positive predicted deformation image I WS , according to the inverse predicted deformation image I WS ′ Calculate the appearance perturbation T in the floating image domain, in the floating image I M After adding the floating image domain appearance perturbation T to the fixed image I F The splicing is performed and input into the teacher registration network. The spatial transformation uncertainty map is calculated based on the predicted deformation field output by the teacher registration network.

[0011] Step 4: Set the total loss function of the adaptive regularized registration network;

[0012] Step 5: According to the total loss function set in step 4, the training set preprocessed in step 1 and the fixed image I F Training the adaptive regularized registration network;

[0013] Step 6: Input the floating image to be registered into the student registration network trained in step 5 to obtain a positive predicted deformation field and a positive predicted deformation image.

[0014] As mentioned above, the pretreatment in step 1 includes the following steps:

[0015] The voxel intensity of the original MRI image is normalized; then the size of the original MRI image and the segmentation label after voxel intensity normalization is adjusted to a set fixed size by zero padding and cropping.

[0016] As described above, in step 2, the student registration network includes an encoder, a decoder, and a channel adjustment module. The encoder includes multiple convolution operation modules and a downsampling module; the decoder includes multiple convolution operation modules and an upsampling module, and the structure of the convolution operation module in the decoder is completely consistent with the structure of the convolution operation module in the encoder; the feature map output by each convolution operation module in the decoder is jump-connected with the feature map output by the corresponding convolution operation module in the encoder, and the feature map output by the last upsampling module in the decoder is input into the channel adjustment module, and the channel adjustment module outputs the displacement in the three coordinate directions of all voxel points of the MRI image, thereby obtaining a predicted deformation field.

[0017] As mentioned above, the convolution operation module includes two convolution layers and two LeakyReLU activation layers. The convolution kernel size of the convolution layer is set to 3×3×3, the convolution step size is set to 1, and the slope of the negative activation value of the LeakyReLU activation layer is set to 0.2.

[0018] As mentioned above, the structure of the teacher registration network is the same as that of the student registration network. The teacher network parameters θ are updated according to the exponential moving average of the student network parameters θ. ′ .

[0019] As mentioned above, the teacher network parameters θ ′ Calculated based on the following formula:

[0020] θ′ t =αθ′ t-1 +(1-α)θ t ,

[0021] where α is the update rate of the exponential moving average, θ′ t-1 is the teacher network parameter at time t-1, θ t is the student network parameter at time t.

[0022] As mentioned above, step 3 includes the following steps:

[0023] Step 3.1: Take the floating images I in the training set M With fixed image IF After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation to obtain the positive predicted deformation field and the inverse predicted deformation field Spatial Transformer Network uses forward prediction of deformation field For floating images I M Perform spatial transformation to obtain the positive predicted deformed image I WS , the spatial transformer network uses the inverse prediction deformation field For a fixed image I F Perform spatial transformation to obtain the inverse predicted deformed image I WS ′;

[0024] Step 3.2: Based on the inverse predicted deformation image I calculated in step 3.1 WS 'Calculate the floating image domain appearance perturbation T: T = β * (I WS ′-I M ), where β is the Beta distribution coefficient of random sampling;

[0025] Step 3.3: Use the Monte Carlo Dropout layer to filter the floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2 to the fixed image I F Splice and input into the teacher registration network, and repeatedly calculate the positive predicted deformation field of the teacher registration network N times A floating image domain appearance perturbation T is added to the input of each operation. The floating image domain appearance perturbation T corresponds to a machine-sampled Beta distribution coefficient β. The positive predicted deformation field output by the teacher registration network is calculated. The mean With variance Thus, we get the spatial transformation uncertainty graph

[0026] As mentioned above, the total loss function of the adaptive regularized registration network includes the positive predicted deformed image I output by the student registration network WS With fixed image I F The similarity measure loss L between sim (I F ,I WS ) and the forward predicted deformation field of the student registration network The regularization loss

[0027]

[0028] in, is the total loss function, is the adaptive regularization loss weight,

[0029]

[0030] in, is the indicator function, p represents each voxel point in the image domain, Represents the uncertainty value corresponding to each voxel point in the spatial transformation uncertainty map, Representing spatial transformation uncertainty graph , k is an empirical scalar value, and τ is the threshold for uncertain prediction of spatial transformation.

[0031] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the magnetic resonance image registration method when executing the computer program.

[0032] A computer-readable storage medium stores a computer program, which implements the steps of the magnetic resonance image registration method when executed by a processor.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention automatically adjusts the weight of the regularization term in the total loss function by registering the spatial uncertainty of the deformation field. The adaptive regularized registration network uses the teacher registration network and the student registration network as the backbone network, and uses the inverse predicted deformation image I generated by the student registration network. WS The difference between the ' and the floating image generates the floating image domain appearance perturbation T, which can make the registration network focus on learning the appearance difference of the unaligned area, so as to optimize the difficult-to-register area in a targeted manner; and by using the Monte Carlo Dropout layer to add the floating image domain appearance perturbation, a spatial transformation uncertainty map can be obtained to guide the student registration network to perceive the more suitable smooth constraint strength for each training step, so as to obtain the adaptive weight of the regularization term and encourage the student registration network to generate higher quality registration results. In view of the related problems of existing magnetic resonance image registration, the method of the present invention can realize fast and accurate registration calculation, improve the registration accuracy, and the deformation process is more consistent with the real topological structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the steps of the method of the present invention;

[0036] Figure 2 Schematic diagram of the overall framework of the adaptive regularized registration network;

[0037] Figure 3This is a schematic diagram of the structure of the teacher registration network and the student registration network. The number above the image block refers to the number of channels of the input image or feature map;

[0038] Figure 4 Schematic diagram of the results of the registration experiment of the fixed image and the floating image of the first embodiment performed by the trained adaptive regularized registration network. (a) is the fixed image I F , (b) is the floating image I M , (c) is the deformed image I of the Voxelmorph method W , (d) is the positive predicted deformation image I of the method of the present invention WS , (e) is the forward predicted deformation field of the method of the present invention

[0039] Figure 5 The schematic diagram of the result of the registration experiment of the fixed image and the floating image of the second embodiment performed by the trained adaptive regularized registration network (a) is a fixed image I F , (b) is the floating image I M , (c) is the deformed image I of the Voxelmorph method W , (d) is the positive predicted deformation image I of the method of the present invention WS , (e) is the forward predicted deformation field of the method of the present invention

[0040] Figure 6 Schematic diagram of the results of the registration experiment of the fixed image and the floating image of the third embodiment performed by the trained adaptive regularized registration network. (a) is the fixed image I F , (b) is the floating image I M , (c) is the deformed image I of the Voxelmorph method W , (d) is the positive predicted deformation image I of the method of the present invention WS , (e) is the forward predicted deformation field of the method of the present invention DETAILED DESCRIPTION

[0041] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0042] Embodiment 1:

[0043] like Figure 1 and Figure 2 As shown, a magnetic resonance image registration method based on adaptive regularization specifically includes the following steps:

[0044] Step 1: Perform a unified preprocessing operation on the original MRI images with segmentation labels, and then randomly select one of the original MRI images as the fixed image I that remains unchanged during the registration process. F , the remaining original MRI image is used as the floating image to be registered I M , and the floating image I M Divide into training set and test set according to a certain ratio.

[0045] In this embodiment, images in the MSD Hippocampus dataset are selected as original MRI images, and 260 original MRI images of the hippocampus with segmentation labels are selected for training and testing. First, the original MRI images are preprocessed, specifically including the following steps: normalizing the voxel intensity of the original MRI images, adjusting the voxel intensity value of the original MRI images to between [0,1]; then adjusting the size of the original MRI images and segmentation labels after voxel intensity normalization to a fixed size of 32×64×32 by 0 padding and cropping, which is convenient for subsequent registration network calculations. The first preprocessed original MRI image is selected as the fixed image I F The remaining 259 preprocessed original MRI images are used as floating images I M , the floating image I M It is divided into a training set and a test set, where the ratio of the number of floating images in the training set to the number of floating images in the test set is 180:79.

[0046] Step 2: Build an adaptive regularized registration network. The adaptive regularized registration network includes a teacher registration network and a student registration network. The network structures of the teacher registration network and the student registration network are the same. The parameters of the teacher registration network are updated by the exponential moving average of the student registration network parameters θ. A spatial transformation network is connected to the student registration network to generate a predicted deformed image. Specifically, the following steps are included:

[0047] Step 2.1, build the student registration network Student RegNet, which includes an encoder, a decoder, and a channel adjustment module. The encoder includes multiple convolution operation modules for extracting features and downsampling modules for reducing resolution. Each convolution operation module includes two convolution layers and two LeakyReLU activation layers. The convolution kernel size of the convolution layer is set to 3×3×3, the convolution step size is set to 1, and the slope of the negative activation value of the LeakyReLU activation layer is set to 0.2; the downsampling module halves the image resolution, reduces parameters and calculations, and speeds up the operation. The decoder includes multiple convolution operation modules and upsampling modules for restoring resolution. The structure of the convolution operation module in the decoder is exactly the same as that of the convolution operation module in the encoder; the upsampling module uses a bilinear interpolation method to double the image resolution for restoring the image size; the feature map output by each convolution operation module in the decoder is jump-connected with the feature map output by the corresponding convolution operation module in the encoder to solve the gradient disappearance problem. The channel adjustment module is a convolution layer with a convolution kernel size of 3×3×3, a convolution step size of 1, and an output channel number of 3. The feature map output by the last upsampling module in the decoder is input into the channel adjustment module, and the channel adjustment module outputs the displacements in the three coordinate directions of all voxel points of the MRI image, thereby obtaining a predicted deformation field (which can be a forward predicted deformation field or a reverse predicted deformation field). The encoder and decoder of this embodiment select a 5-layer pyramid structure, that is, 4 downsampling modules, 4 upsampling modules, and 4 skip connections.

[0048] Step 2.2: Build the teacher registration network Teacher RegNet. The structure of the teacher network is consistent with the student network. Update the teacher network parameters θ according to the exponential moving average (EMA) of the student network parameters θ ′ , the teacher network parameter θ′ at time t t The specific calculation formula is: t =αθ′ t-1 +(1-α)θ t , where α is the update rate of the exponential moving average EMA, which is empirically set to 0.99, and θ′ t-1 is the teacher network parameter at time t-1, θ t is the student network parameter at time t.

[0049] Step 2.3: Build a spatial transformation network. A spatial transformation network is connected to the student registration network. The spatial transformation network uses the positive predicted deformation field output by the student registration network. For floating images I M Perform spatial transformation operation to obtain the positive predicted deformed image I WS ;

[0050] Step 3: Floating images I in the training set M With fixed image I F After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation to obtain the positive predicted deformation field and the inverse predicted deformation field Forward-based predicted deformation field and floating image I M , get the positive predicted deformation image I WS , according to the inverse predicted deformation image I WS ′ Calculate the appearance perturbation T in the floating image domain, and use the Monte Carlo Dropout layer to perturb the floating image I M Add the floating image domain appearance perturbation T to the fixed image I F The splicing is performed and input into the teacher registration network. The spatial transformation uncertainty map is calculated based on the predicted deformation field output by the teacher registration network. Spatial uncertainty is used to filter out unreliable deformed pixel points, and the regularization loss weight of the deformation field is automatically adjusted based on reliable prediction results, so as to supervise the student registration network to learn meaningful and well-aligned areas, thereby generating high-quality registration results. The specific steps include:

[0051] Step 3.1: The floating images I in the training set M With fixed image I F After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation, that is, exchanging the floating image I M and a fixed image I F The order of can be used to obtain the positive predicted deformation field (by floating image I M To fixed image I F Aligned deformation field) and the inverse predicted deformation field (From the fixed image I F To floating image I M The spatial transformer network uses the positive predicted deformation field For floating images I M Perform spatial transformation to obtain the positive predicted deformed image I WS , the spatial transformer network uses the inverse prediction deformation field For a fixed image I F Perform spatial transformation to obtain the inverse predicted deformed image I WS ′. The forward predicted deformation field calculated by the student registration network The forward predicted deformed image I calculated by the spatial transformer network WS as a predicted result.

[0052] Step 3.2: Based on the inverse predicted deformation image I calculated in step 3.1 WS 'Calculate the floating image domain appearance perturbation T: T = β * (I WS ′-I M ), where β is the Beta distribution coefficient of random sampling, (I WS ′-I M ) is the reverse predicted deformed image I WS ′ and floating image I M The difference in appearance between.

[0053] Step 3.3: Monte Carlo Dropout layer on floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2 to the fixed image I F Splice and input into the teacher registration network, and repeatedly calculate the positive predicted deformation field of the teacher registration network N times Each operation input will add a floating image domain appearance perturbation T, which corresponds to a randomly sampled Beta distribution coefficient β. Therefore, the input of each repeated operation will change, resulting in different prediction outputs. In this embodiment, N is set to 6. After the loop is completed, the forward prediction deformation field of the teacher registration network output is calculated. The mean With variance Thus, we get the spatial transformation uncertainty graph Calculated according to the following formula:

[0054]

[0055] Step 4: Set the total loss function. The total loss function of the adaptive regularized registration network includes the positive predicted deformed image I output by the student registration network WS With fixed image I F The similarity measure loss L between sim (I F ,I WS ) and the forward predicted deformation field of the student registration network The regularization loss The specific calculation is expressed by the following formula:

[0056]

[0057] in, is the total loss function, the similarity measure loss L sim (I F ,I WS ) Select the positive predicted deformed image I output by the student registration networkWS With fixed image I F The negative local correlation coefficient between Select the positive predicted deformation field of the student registration network Diffusion regularization of spatial gradients in all directions is used as the loss function. is the adaptive regularization loss weight, calculated according to the following formula:

[0058]

[0059] in, is the indicator function, p represents each voxel point in the image domain, Represents the uncertainty value corresponding to each voxel point in the spatial transformation uncertainty map, Representing spatial transformation uncertainty graph The volume size of , k is an empirical scalar value, and τ is a threshold for uncertain prediction of spatial transformation. In this embodiment, k is set to 5 and τ is set to 0.1.

[0060] Step 5: According to the total loss function set in step 4, the training set preprocessed in step 1 and the fixed image I F The adaptive regularized registration network was trained with an initial learning rate of 0.0001, a batch size of 1, a maximum training epoch of 10,000, and the ADAM optimizer was used for optimization. After the training was completed, the parameter model of the adaptive regularized registration network was saved.

[0061] Step 6: Input the floating image to be registered into the student registration network trained in step 5 to obtain a positive predicted deformation field and a positive predicted deformation image.

[0062] Input the test set generated in step 1 into the student registration network trained in step 5 to obtain the positive predicted deformation field and the forward predicted deformed image I WS The registration results are quantitatively evaluated by calculating the Dice similarity coefficient and the Jacobian of the deformation field using the segmentation labels of the fixed image and the forward predicted deformed image. The number and percentage of voxel points are used to quantitatively analyze the registration performance.

[0063] Table 1 shows the quantitative analysis results of MRI image registration on the test set using the trained adaptive regularized registration network. The Dice similarity coefficient is used to measure the overlap between the deformable image and the fixed image. The Jacobian determinant The number and percentage of voxel points are used to evaluate the smoothness of the deformation field.

[0064] Table 1 Quantitative analysis results of image registration on the test set

[0065]

[0066] Figures 4 to 6 The following is a schematic diagram showing the results of the registration experiment of the fixed image and floating image of the three embodiments using the trained adaptive regularized registration network. Figures 4 to 6 In the figure, each column from left to right represents: fixed image I F Floating Image I M , Voxelmorph method deformation image I W , the forward prediction deformation image I of the method of the present invention WS And the forward predicted deformation field of the method of the present invention From Table 1 and Figures 4 to 6 It can be seen from the figure that the deformed image calculated by the present invention is visually similar to the fixed image I F The Dice similarity coefficient is higher than that of the Voxelmorph method, and the number of voxel points where the deformation field is folded is relatively small, which is consistent with the real deformation process, indicating that the method of the present invention can achieve accurate registration of magnetic resonance images.

[0067] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0068] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0069] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0070] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0071] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A magnetic resonance image registration method based on adaptive regularization, characterized in that: The following steps are involved: Step 1: Preprocess the original MRI images with segmentation labels and select one of the original MRI images as the fixed image I F , the remaining original MRI image is used as the floating image I M , according to the floating image I M Construct training and test sets; Step 2: Build an adaptive regularized registration network, which includes a teacher registration network, a spatial transformation network, and a student registration network; Step 3: Floating images in the training set I M With fixed image I F After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation to obtain the positive predicted deformation field and the inverse predicted deformation field Forward-based predicted deformation field and floating image I M , get the positive predicted deformation image I WS , according to the inverse predicted deformation image I WS ′ Calculate the appearance perturbation T in the floating image domain, in the floating image I M After adding the floating image domain appearance perturbation T to the fixed image I F The splicing is performed and input into the teacher registration network. The spatial transformation uncertainty map is calculated based on the predicted deformation field output by the teacher registration network. Step 4: Set the total loss function of the adaptive regularized registration network; Step 5: According to the total loss function set in step 4, the training set preprocessed in step 1 and the fixed image I F Training the adaptive regularized registration network; Step 6: Input the floating image to be registered into the student registration network trained in step 5 to obtain a positive predicted deformation field and a positive predicted deformation image.

2. The magnetic resonance image registration method based on adaptive regularization according to claim 1, characterized in that: The pretreatment in step 1 comprises the following steps: The voxel intensity of the original MRI image is normalized; then the size of the original MRI image and the segmentation label after voxel intensity normalization is adjusted to a set fixed size by zero padding and cropping.

3. The magnetic resonance image registration method based on adaptive regularization according to claim 1, characterized in that: In step 2, the student registration network includes an encoder, a decoder, and a channel adjustment module, the encoder includes a plurality of convolution operation modules and a downsampling module; the decoder includes a plurality of convolution operation modules and an upsampling module, and the structure of the convolution operation module in the decoder is completely consistent with the structure of the convolution operation module in the encoder; The feature map output by each convolution operation module in the decoder is jump-connected with the feature map output by the corresponding convolution operation module in the encoder. The feature map output by the last upsampling module in the decoder is input into the channel adjustment module. The channel adjustment module outputs the displacement in three coordinate directions of all voxel points of the MRI image, thereby obtaining the predicted deformation field.

4. The magnetic resonance image registration method based on adaptive regularization according to claim 3, characterized in that: The convolution operation module includes two convolution layers and two LeakyReLU activation layers. The convolution kernel size of the convolution layer is set to 3×3×3, the convolution step size is set to 1, and the slope of the negative activation value of the LeakyReLU activation layer is set to 0.

2.

5. The magnetic resonance image registration method based on adaptive regularization according to claim 3, characterized in that: The structure of the teacher registration network is the same as that of the student registration network, and the teacher network parameters θ′ are updated according to the exponential moving average of the student network parameters θ.

6. The magnetic resonance image registration method based on adaptive regularization according to claim 5, characterized in that: The teacher network parameter θ′ is calculated based on the following formula: θ′ t =θ′ t-1 +(1-α)θ t , where α is the update rate of the exponential moving average, θ′ t-1 is the teacher network parameter at time t-1, θ t is the student network parameter at time t.

7. The magnetic resonance image registration method based on adaptive regularization according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Take the floating images I in the training set M With fixed image I F After splicing in the channel dimension, it is input into the student registration network for bidirectional registration operation to obtain the positive predicted deformation field and the inverse predicted deformation field Spatial Transformer Network uses forward prediction of deformation field For floating images I M Perform spatial transformation to obtain the positive predicted deformed image I WS , the spatial transformer network uses the inverse prediction deformation field For a fixed image I F Perform spatial transformation to obtain the inverse predicted deformed image I WS ′; Step 3.2: Based on the inverse predicted deformation image I calculated in step 3.1 WS 'Calculate the floating image domain appearance perturbation T: T = β * (I WS ′-I M ), where β is the Beta distribution coefficient of random sampling; Step 3.3: Use the Monte Carlo Dropout layer to filter the floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2 to the fixed image I F Splice and input into the teacher registration network, and repeatedly calculate the positive predicted deformation field of the teacher registration network N times A floating image domain appearance perturbation T is added to the input of each operation. The floating image domain appearance perturbation T corresponds to a machine-sampled Beta distribution coefficient β. The positive predicted deformation field output by the teacher registration network is calculated. The mean With variance Thus, we get the spatial transformation uncertainty graph 8. The magnetic resonance image registration method based on adaptive regularization according to claim 7, characterized in that: The total loss function of the adaptive regularized registration network includes the positive predicted deformed image I output by the student registration network WS With fixed image I F The similarity measure loss L between sim (I F ,I WS ) and the forward predicted deformation field of the student registration network The regularization loss in, is the total loss function, is the adaptive regularization loss weight, in, is the indicator function, p represents each voxel point in the image domain, Represents the uncertainty value corresponding to each voxel point in the spatial transformation uncertainty map, Representing spatial transformation uncertainty graph , k is an empirical scalar value, and τ is the threshold for uncertain prediction of spatial transformation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the magnetic resonance image registration method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the magnetic resonance image registration method according to any one of claims 1 to 8 are implemented.

Images