A method for magnetic resonance image registration based on adaptive regularization

Abstract

The application discloses a magnetic resonance image registration method based on adaptive regularization, constructs a training set and a test set; an adaptive regularization registration network is built; the floating image and the fixed image in the training set are spliced in the channel dimension and then input into the student registration network for bidirectional registration operation, the appearance disturbance of the floating image domain is calculated, the spatial transformation uncertainty map is calculated according to the predicted deformation field output by the teacher registration network; and the adaptive regularization registration network is trained. The appearance disturbance of the floating image domain is added in the teacher registration network to obtain the spatial transformation uncertainty map, the weight of the regularization term in the loss function is automatically adjusted by utilizing the spatial transformation uncertainty, adaptive regularization is realized, and the magnetic resonance image registration precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetic resonance image registration, and more particularly to a magnetic resonance image registration method based on adaptive regularization for registering magnetic resonance images. BACKGROUND

[0002] Magnetic Resonance Imaging (MRI) is a non-invasive, non-ionizing radiation imaging technology, which is a physical phenomenon of nuclear magnetic resonance. It uses radio frequency pulses to excite atomic nuclei with non-zero spins in a magnetic field, and after the radio frequency pulses stop, the atomic nuclei relax. During the relaxation process, the signal is collected by an induction coil, and a mathematical image is formed by reconstruction according to a certain mathematical method. Compared with other imaging techniques, magnetic resonance images can accurately and clearly show the anatomical structure of the imaging site.

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

[0004] Traditional magnetic resonance image registration methods have certain limitations and generally have the disadvantages of complex calculation, long time consumption and poor universality. Deep learning technology has been widely used in magnetic resonance image registration due to its strong learning ability and feature extraction ability. 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 a 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 regularization weight λ of the deformation field of all image pairs to be registered is set to a fixed value. This fixed weight parameter method may not necessarily obtain the optimal solution, because the regularization strength of a specific image pair should be related to its image content and alignment degree [Zhe X, Jie L, Donghuan L. MICCAI, 2022, pp: 14-24]. Therefore, a suitable adaptive regularization weight method is needed to improve registration accuracy and increase the universality of the network model. SUMMARY

[0005] The present application aims at the above-mentioned defects existing in the prior art, and proposes a magnetic resonance image registration method based on adaptive regularization, which uses a teacher registration network-student registration network as a backbone network in a registration network, adds appearance disturbance of a floating image domain in the teacher registration network to obtain a spatial transformation uncertainty map, and automatically adjusts the weight of a regularization term in a loss function by using the spatial transformation uncertainty, so as to realize adaptive regularization and improve the magnetic resonance image registration precision.

[0006] The above-mentioned object of the present application is realized by the following technical scheme:

[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, select one of the original MRI images as a fixed image I F , and the remaining original MRI images as floating images I M , construct a training set and a test set according to the floating images I M ;

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

[0010] Step 3: The floating images I M in the training set are spliced with the fixed image I F in the channel dimension and then input into the student registration network for bidirectional registration operation, to obtain a forward predicted deformation field and a reverse predicted deformation field Based on the forward predicted deformation field and the floating image I M , a forward predicted deformation image I WS is obtained, the floating image domain appearance disturbance T is calculated according to the reverse predicted deformation image I WS ′, and the floating image I M is spliced with the fixed image I F after adding the floating image domain appearance disturbance T and then input into the teacher registration network, and the spatial transformation uncertainty map is calculated according to the predicted deformation field output by the teacher registration network.

[0011] Step 4: Set the total loss function of the adaptive regularization 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 are used to train the adaptive regularization registration network.

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

[0014] As described above, the preprocessing in step 1 includes the following steps:

[0015] The original MRI image is voxel intensity normalized; then, the size of the original MRI image and segmentation labels after voxel intensity normalization and segmentation labels are 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. The structure of the convolution operation modules in the decoder is completely consistent with the structure of the convolution operation modules in the encoder. The feature map output by each convolution operation module in the decoder is skipped to 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 to the channel adjustment module. The channel adjustment module outputs the displacement of all voxel points in the MRI image in three coordinate directions, thereby obtaining the predicted deformation field.

[0017] As described above, the convolution operation module includes two convolutional layers and two LeakyReLU activation layers. The kernel size of the convolutional layers is set to 3×3×3, the stride is set to 1, and the slope of the negative activation values ​​of the LeakyReLU activation layers 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 based on the exponential moving average of the student network parameters θ. ′ .

[0019] As mentioned above, the teacher network parameter 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 Let θ be the teacher network parameters at time t-1. t Let t be the student network parameters.

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

[0023] Step 3.1: Floating images I from the training set M With fixed image IF After being stitched along the channel dimension, the data is input into the student registration network for bidirectional registration, yielding the positive predicted deformation field. And the inverse prediction of deformation field Spatial transformation networks use forward prediction of deformation fields. For floating image I M A spatial transformation is performed to obtain a positive predicted deformation image I. WS Spatial transformation networks use inverse prediction of deformation fields. For a fixed image I F Perform a spatial transformation to obtain the inverse predicted deformation image I. WS ′;

[0024] Step 3.2: Calculate the inverse predicted deformation image I based on step 3.1. WS Calculate the appearance perturbation T in the floating image domain: T = β*(I) WS ′-I M ), where β is the Beta distribution coefficient of the random sample;

[0025] Step 3.3: Apply the Monte Carlo Dropout layer to the floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2, and then combine it with the fixed image I. F The data is then stitched together and input into the teacher registration network. The positive predicted deformation field of the teacher registration network is calculated N times. Each operation adds a floating image domain appearance perturbation T to the input. This perturbation T corresponds to a machine-sampled Beta distribution coefficient β. The positive predicted deformation field output by the teacher registration network is then calculated. mean With variance Thus, the spatial transformation uncertainty diagram is obtained.

[0026] As described above, the total loss function of the adaptive regularization registration network includes the positive predicted deformed image I output by the student registration network. WS With fixed image I F Similarity measure loss L sim (I F ,I WS The positive prediction of deformation field by the student registration network. Regularization loss

[0027]

[0028] in, For the total loss function, It is an adaptive regularization loss weight.

[0029]

[0030] wherein, is an 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, represents the volume size of the spatial transformation uncertainty map k is an empirical scalar value, and tau is a threshold value of the spatial transformation uncertainty prediction.

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

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

[0034] The present application automatically adjusts the weight of the regularization term in the total loss function by registering the spatial uncertainty of the deformation field. Among them, the adaptive regularization registration network uses the teacher registration network and the student registration network as the backbone network, and uses the inverse prediction deformation image I WS generated by the student registration network to generate the appearance disturbance T of the floating image domain between the floating image and the difference, so that the registration network can focus on learning the appearance difference of the misaligned region, thereby optimizing the difficult registration region; and by using the Monte Carlo Dropout layer (Monte Carlo Dropout) to add the appearance disturbance of the floating image domain, the spatial transformation uncertainty map can be obtained to guide the student registration network to perceive the more suitable smooth constraint strength at 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. For the related problems of existing magnetic resonance image registration, the method of the present application can realize fast and accurate registration calculation, can improve the registration accuracy, and the deformation process is more consistent with the real topological structure. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is the step flow chart of the method of the present application;

[0036] Figure 2 is the overall framework schematic diagram of the adaptive regularization registration network;

[0037] Figure 3The structure diagram of the teacher matching network and the student matching network, the numbers above the image blocks refer to the channel number of the input image or feature map;

[0038] Figure 4 The result diagram of the trained adaptive regularization matching network in the registration experiment of the fixed image and the floating image of the first embodiment, (a) is the fixed image I F , (b) is the floating image I M , (c) is the deformation image I W of the Voxelmorph method, (d) is the predicted deformation image I WS of the forward direction of the method of the application, (e) is the predicted deformation field of the forward direction of the method of the application

[0039] Figure 5 The result diagram of the trained adaptive regularization matching network in the registration experiment of the fixed image and the floating image of the second embodiment, (a) is the fixed image I F , (b) is the floating image I M , (c) is the deformation image I W of the Voxelmorph method, (d) is the predicted deformation image I WS of the forward direction of the method of the application, (e) is the predicted deformation field of the forward direction of the method of the application

[0040] Figure 6 The result diagram of the trained adaptive regularization matching network in the registration experiment of the fixed image and the floating image of the third embodiment, (a) is the fixed image I F , (b) is the floating image I M , (c) is the deformation image I W of the Voxelmorph method, (d) is the predicted deformation image I WS of the forward direction of the method of the application, (e) is the predicted deformation field of the forward direction of the method of the application DETAILED DESCRIPTION

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

[0042] Embodiment 1:

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

[0044] Step 1, uniform preprocessing operation is performed on the original MRI image with segmentation label, and then one of the original MRI images is randomly selected as the fixed image I which remains unchanged in the registration process F , and the remaining original MRI images are used as the floating images I M , and the floating images I M are divided into training set and test set according to a certain proportion.

[0045] In this embodiment, the images in the MSD Hippocampus dataset are selected as the original MRI images, and the original MRI images with segmentation labels of 260 hippocampi are selected for training and testing. First, the original MRI images are preprocessed, which specifically includes the following steps: the voxel intensity of the original MRI image is normalized, and the voxel intensity value of the original MRI image is adjusted to [0, 1]; then the size of the original MRI image and the segmentation label after voxel intensity normalization is adjusted to a fixed size of 32x64x32 through 0 padding and cropping, which is convenient for subsequent registration network calculation. The first preprocessed original MRI image is selected as the fixed image I F , and the remaining 259 preprocessed original MRI images are used as the floating images I M , the floating images I M are divided into training set and test set, and 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 regularization registration network, the adaptive regularization registration network includes a teacher registration network and a student registration network, the network structure of the teacher registration network and the student registration network is the same, and the parameters of the teacher registration network are updated through the exponential moving average of the student registration network parameters θ. A spatial transformation network is connected behind the student registration network, which is used to generate a predicted deformation image. Specifically, it includes the following steps:

[0047] Step 2.1: Construct the Student RegNet registration network, which includes an encoder, a decoder, and a channel adjustment module. The encoder includes multiple convolutional operation modules for feature extraction and a downsampling module to reduce resolution. Each convolutional operation module includes two convolutional layers and two LeakyReLU activation layers. The kernel size of the convolutional layers is set to 3×3×3, the stride is set to 1, and the slope of the negative activation values ​​of the LeakyReLU activation layers is set to 0.2. The downsampling module halves the image resolution, reducing parameters and computational load, and speeding up the operation. The decoder includes multiple convolutional operation modules and an upsampling module to restore resolution. The structure of the convolutional operation modules in the decoder is completely identical to that in the encoder. The upsampling module uses bilinear interpolation to double the image resolution to restore the image size. The feature maps output by each convolutional operation module in the decoder are skip-connected to the corresponding feature maps output by the convolutional operation modules in the encoder to solve the gradient vanishing problem. The channel adjustment module is a convolutional layer with a kernel size of 3×3×3, a stride of 1, and 3 output channels. The feature map output from the last upsampling module in the decoder is input into the channel adjustment module. The channel adjustment module outputs the displacement of all voxels in the MRI image in three coordinate directions, thereby obtaining the predicted deformation field (which can be a forward or inverse predicted deformation field). In this embodiment, the encoder and decoder adopt a 5-layer pyramid structure, that is, using 4 downsampling modules, 4 upsampling modules, and 4 skip connections.

[0048] Step 2.2: Construct the Teacher RegNet, with the same structure as the student network. Update the teacher network parameters θ based on the exponential moving average (EMA) of the student network parameters θ. ′ Teacher network parameters θ′ at time t t The specific calculation formula is: θ′ 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, θ′ t-1 Let θ be the teacher network parameters at time t-1. t Let t be the student network parameters.

[0049] Step 2.3: Construct a spatial transformation network. A spatial transformation network is connected after the student registration network. This spatial transformation network utilizes the positive predicted deformation field output by the student registration network. For floating image I M Perform a spatial transformation operation to obtain a positive predicted deformation image I. WS ;

[0050] Step 3, the floating image I in the training set M with the fixed image I F After concatenation in the channel dimension, input into the student registration network for bidirectional registration operation, to obtain the forward predicted deformation field and the reverse predicted deformation field Based on the forward predicted deformation field and the floating image I M , obtain the forward predicted deformed image I WS , calculate the floating image domain appearance disturbance T according to the reverse predicted deformed image I WS , add the floating image domain appearance disturbance T on the floating image I M through the Monte Carlo Dropout layer (Monte Carlo Dropout), and then concatenate with the fixed image I F and input into the teacher registration network, and calculate the spatial transformation uncertainty map according to the predicted deformation field output by the teacher registration network Filter out unreliable deformation voxel points using spatial uncertainty, automatically adjust the regularization loss term weight of the deformation field from reliable prediction results, supervise the student registration network to learn meaningful and well-aligned regions, and thus generate high-quality registration results. Specifically, the following steps are included:

[0051] Step 3.1, the floating image I in the training set M and the fixed image I F After concatenation in the channel dimension, input into the student registration network for bidirectional registration operation, i.e. exchange the order of the floating image I M and the fixed image I F , to obtain the forward predicted deformation field (by the floating image I M to the fixed image I F alignment deformation field) and the reverse predicted deformation field (by the fixed image I F to the floating image I M alignment deformation field), the spatial transformation network uses the forward predicted deformation field to perform spatial transformation on the floating image I M to obtain the forward predicted deformed image I WS , and the spatial transformation network uses the reverse predicted deformation field to perform spatial transformation on the fixed image I F to obtain the reverse predicted deformed image I WS '. Take the forward predicted deformation field calculated by the student registration network and the forward predicted deformed image I WS calculated by the spatial transformation network as the prediction result.

[0052] Step 3.2: Calculate the inverse predicted deformation image I based on step 3.1. WS Calculate the appearance perturbation T in the floating image domain: T = β*(I) WS ′-I M ), where β is the Beta distribution coefficient of the random sample, (I WS ′-I M (I) is the inverse predicted deformation image. WS ′ with floating image I M The differences in appearance between them.

[0053] Step 3.3: Apply Monte Carlo Dropout layer to the floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2, and then combine it with the fixed image I. F The data is then stitched together and input into the teacher registration network. The positive predicted deformation field of the teacher registration network is calculated N times. Each operation adds a floating image domain appearance perturbation T to the input. This perturbation T corresponds to a randomly sampled Beta distribution coefficient β. Therefore, the input changes with each repeated operation, leading to different predicted outputs. In this embodiment, N is set to 6. After the loop ends, the positive predicted deformation field output by the teacher registration network is calculated. mean With variance Thus, the spatial transformation uncertainty diagram is obtained. Calculate using the following formula:

[0054]

[0055] Step 4: Define the total loss function. The total loss function of the adaptive regularized registration network includes the positive predicted deformation image I output by the student registration network. WS With fixed image I F Similarity measure loss L sim (I F ,I WS The positive prediction of deformation field by the student registration network. Regularization loss The specific calculation is expressed by the following formula:

[0056]

[0057] in, Let L be the total loss function, and L be the similarity measure loss. sim (I F ,I WS Select the positive predicted deformation image I output from the student registration network.WS between the fixed image I F and the negative local cross-correlation coefficient between the moving image I The forward predicted deformation field of the student registration network is selected The diffusion regularization of the spatial gradient in each direction is taken as the loss function. is the adaptive regularization loss weight, which is calculated according to the following formula:

[0058]

[0059] wherein, is an 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, represents the volume size of the spatial transformation uncertainty map , k is an empirical scalar value, and τ is a threshold value of the spatial transformation uncertainty prediction. In the present 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 are input into the adaptive regularization registration network to train the network, the initial learning rate of the network is 0.0001, the batch size is 1, the maximum training epoch is 10000, the ADAM optimizer is used for optimization, and the parameter model of the adaptive regularization registration network is saved after the training is completed.

[0061] Step 6: The floating image to be registered is input into the student registration network trained in step 5 to obtain the forward predicted deformation field and the forward predicted deformation image.

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

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

[0064] Table 1 shows the quantitative analysis results of image registration of the test set

[0065]

[0066] Figures 4 to 6 Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, Figures 4 to 6 Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, F Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, M Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, W Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, WS Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments, From Table 1 and Figures 4 to 6 It can be seen from Table 1 and F Fig. 6 shows a result diagram of the registration experiment of the trained adaptive regularization registration network on the fixed image and the floating image of three embodiments,

[0067] In one embodiment, a computer device is also provided, which comprises a memory and a processor, the memory stores a computer program, 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, which stores a computer program, and the computer program implements the steps in the above-mentioned method embodiments when executed by a processor.

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

[0070] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments.

[0071] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or replace them with similar ways without departing from the spirit of the present application or exceeding the scope defined by the appended claims.

Claims

1. A magnetic resonance image registration method based on adaptive regularization, characterized in that, Includes the following steps: Step 1: Preprocess the raw MRI images with segmentation labels, and select one of the raw MRI images as the fixed image I. F The remaining original MRI images are used as floating images I M According to floating image I M Construct training and testing sets; Step 2: Construct an adaptive regularized registration network, which includes a teacher registration network, a spatial transformation network, and a student registration network; Step 3: Floating images I in the training set M With fixed image I F After being stitched along the channel dimension, the data is input into the student registration network for bidirectional registration, yielding the positive predicted deformation field. And the inverse prediction of deformation field Based on positive prediction of deformation field and floating image I M Obtain a positive predicted deformation image I WS Based on inverse prediction of deformation field and fixed image I F Obtain the inverse predicted deformation image I WS ′, using the student registration network to generate the inverse predicted deformation image I WS The difference between the floating image and the floating image generates a perturbation T in the floating image domain, in the floating image I. M After adding a floating image domain appearance perturbation T, and then comparing it with a fixed image I F The data is spliced ​​and input into the teacher registration network. Based on the predicted deformation field output by the teacher registration network, the spatial transformation uncertainty diagram is calculated. and The mean of the positive predicted deformation field output by the teacher registration network is respectively... With variance Step 4: Set the total loss function of the adaptive regularization registration network, determine the adaptive regularization loss weight based on the uncertainty value corresponding to each voxel point in the spatial transformation uncertainty map, and determine the total loss function based on the adaptive regularization loss weight; Step 5: Based on the total loss function set in Step 4, use the training set preprocessed in Step 1 and the fixed image I F Train 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 the positive predicted deformation field and the positive predicted deformation image.

2. The magnetic resonance image registration method based on adaptive regularization according to claim 1, characterized in that, The preprocessing in step 1 includes the following steps: The original MRI image is voxel intensity normalized; then, the size of the original MRI image and segmentation labels after voxel intensity normalization and segmentation labels are 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 multiple convolution operation modules and a downsampling module; the decoder includes multiple convolution operation modules and an upsampling module. 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 skipped to the feature map output by the corresponding convolution operation module in the encoder. The feature map output by the last layer upsampling module in the decoder is input to the channel adjustment module. The channel adjustment module outputs the displacement of all voxel points in the MRI image in three coordinate directions, 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 convolutional layers and two LeakyReLU activation layers. The kernel size of the convolutional layers is set to 3×3×3, the stride is set to 1, and the slope of the negative activation values ​​of the LeakyReLU activation layers 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 parameter θ′ is updated based on the exponential moving average of the student network parameter 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 Let θ be the teacher network parameters at time t-1. t Let θ′ be the student network parameters at time t. t Let t be the teacher network parameters at time t.

7. The magnetic resonance image registration method based on adaptive regularization according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Floating images I from the training set M With fixed image I F After being stitched along the channel dimension, the data is input into the student registration network for bidirectional registration, yielding the positive predicted deformation field. And the inverse prediction of deformation field Spatial transformation networks use forward prediction of deformation fields. For floating image I M A spatial transformation is performed to obtain a positive predicted deformation image I. WS Spatial transformation networks use inverse prediction of deformation fields. For a fixed image I F Perform a spatial transformation to obtain the inverse predicted deformation image I. WS ′; Step 3.2: Calculate the inverse predicted deformation image I based on step 3.

1. WS Calculate the appearance perturbation T in the floating image domain: T = β*(I) WS ′-I M ), where β is the Beta distribution coefficient of the random sample; Step 3.3: Apply the Monte Carlo Dropout layer to the floating image I M Add the floating image domain appearance perturbation T obtained in step 3.2, and then combine it with the fixed image I. F The data is then stitched together and input into the teacher registration network. The positive predicted deformation field of the teacher registration network is calculated N times. Each operation adds a floating image domain appearance perturbation T to the input. This perturbation T corresponds to a machine-sampled Beta distribution coefficient β. The positive predicted deformation field output by the teacher registration network is then calculated. mean With variance Thus, the spatial transformation uncertainty diagram is obtained.

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 Similarity measure loss L sim (I F ,I WS The positive prediction of deformation field by the student registration network. Regularization loss in, For the total loss function, It is an adaptive regularization loss weight. in, It is an indicator function, where p represents each voxel point in the image domain. This represents the uncertainty value corresponding to each voxel point in the spatial transformation uncertainty graph. Representing the uncertainty diagram of spatial transformation The volume size is given by k, which 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, characterized in that, When the processor executes the computer program, it implements the steps of the magnetic resonance image registration method according to any one of claims 1 to 8.

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

Images