A cross-modal medical image registration method and system based on Bayesian neural architecture search

Abstract

The present invention discloses a cross-modal medical image registration method and system based on Bayesian neural architecture search, including: a data preprocessing module, a search space module and a Bayesian optimization module; the data preprocessing module obtains the image data of the target cross-modal medical image registration task and performs image denoising, contrast adjustment and rigid registration; the search space module designs a full convolution double U-shaped parameterized neural architecture search space containing a feature adaptive fusion block; the Bayesian optimization module uses a loss function and a verification index to construct the optimization problem of the modality adaptive registration network architecture design under the constraints of the search space, and adopts Bayesian optimization to solve and obtain the optimal network architecture. The present invention generates an adaptive network architecture that extracts and fuses cross-modal image features with minimum deviation according to the modal characteristics of the target registration task, which can be automatically optimized without manual intervention, is suitable for cross-modal medical image registration, and achieves a significant improvement in registration performance.

Description

Technical Field

[0001] The present invention relates to the technical field of cross-modal medical image registration, and in particular to a cross-modal medical image registration method and system based on Bayesian neural architecture search. Background Art

[0002] Medical image registration, as a key technology for image-guided surgery and image-guided radiotherapy, focuses on achieving precise alignment between two or more images. In particular, multimodal images acquired from different sensors provide complementary organ and tissue structure information. Accurate alignment of multimodal information helps improve positioning accuracy, reduce treatment complications, and improve safety. Therefore, cross-modal medical image registration is crucial. Registration methods based on deep learning usually use neural networks to extract and fuse the features of floating images and reference images to predict the deformation field between the two images. Subsequently, the deformation field is applied to the floating image through a spatial transformation network to align it with the reference image.

[0003] However, existing deep learning registration algorithms perform poorly when registering multimodal medical images, and cross-modal lesion information is difficult to accurately align, which poses severe challenges to the accuracy and safety of surgery or radiotherapy. The fundamental reason is that there are huge differences between the image features of the same organ or tissue in different modalities, which leads to huge deviations in the extracted and fused multimodal features, and ultimately predicts incorrect deformation fields. With the explosive growth of modal types of medical images, the modal combinations of registration tasks are becoming more and more diverse. A single network architecture is difficult to adapt to the registration tasks of any modal combination because its performance is affected by the extraction, fusion and utilization of multimodal features. Customizing registration models for specific modal combinations is highly dependent on manual construction, which is very time-consuming and labor-intensive. Therefore, it is very important to construct a method that can adaptively design a network architecture that extracts and fuses cross-modal image features with minimal deviation based on the modal characteristics of the target registration task to enhance the accuracy and efficiency of surgery and radiotherapy. Summary of the invention

[0004] Purpose of the invention: The present invention provides a cross-modal medical image registration method and system based on Bayesian neural architecture search, which can generate an adaptive network architecture for extracting and fusing cross-modal image features with minimum deviation according to the modal characteristics of the target registration task in a fully convolutional dual U-parameter neural architecture search space containing a feature adaptive fusion block; the architecture can be automatically optimized without human intervention, is suitable for cross-modal medical image registration, and achieves a significant improvement in registration performance.

[0005] Technical solution: The cross-modal medical image registration method based on Bayesian neural architecture search described in the present invention comprises the following steps:

[0006] Step 1: Obtain image data for the target cross-modal medical image registration task and perform preprocessing;

[0007] Step 2: Design a fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks;

[0008] Step 3: Construct a loss function for training the neural network.

[0009] Step 4: Establish verification indicators for verifying the performance of the neural network;

[0010] Step 5, constructing the optimization problem of modality adaptive registration network architecture design according to the search space in step 2, the loss function in step 3 and the verification index in step 4;

[0011] Step 6: Use Bayesian optimization to solve the optimization problem in step 5 and obtain the optimal network architecture.

[0012] Furthermore, in step 1, the registration task aligns two images of different modalities at a time, and the images of the two modalities are from the same patient or from different patients, and then the image data is denoised, the contrast is adjusted, and finally the image data is pre-registered through a rigid registration method.

[0013] Furthermore, in step 2, the feature adaptive fusion block includes a feature fusion block, a feature selection block, and a correlation regularization term; in the feature fusion block, two feature map tensors that have been weighted by hyperparameters are concatenated in the channel dimension, and then the number of channels is halved using a three-dimensional convolution operation to achieve feature fusion; in the feature selection block, a spatial attention mechanism is used to implement the fused feature map tensor, and the fused features of different parts of the image are weighted by the attention mechanism to achieve feature selection. Subsequently, the fused and attention-weighted feature map tensor is added to the two original feature map tensors in a certain ratio to obtain the final fused feature map tensor; in the correlation regularization term, in order to gradually reduce the deviation between the cross-modal fused feature map tensors, a correlation regularization term is constructed. ,in is the correlation coefficient between the two fused feature map tensors, For bias.

[0014] Furthermore, in step 2, designing a fully convolutional double U-shaped parameterized neural architecture search space with a feature adaptive fusion block specifically includes the following steps:

[0015] Step 21: Construct a fully convolutional double U-shaped search space; use two U-shaped encoder-decoder networks with jump connections to process cross-modal floating images respectively. and reference image There is a feature adaptive fusion block between each level of the two U-shaped networks; in the encoder stage, the two input images are respectively passed through the convolution layer, the normalization layer and the activation function to obtain two feature map tensors, and then the feature adaptive fusion module takes these two feature map tensors as input, and the output two fused weighted feature map tensors are respectively reduced in size by the maximum pooling operation to enter the next level; in the decoder stage, the two fused feature map tensors output by the feature adaptive fusion block are expanded in size by trilinear interpolation to enter the next level, and after the two U-shaped networks respectively output the final mixed feature map tensors, the two mixed feature map tensors are tensor-added, and then pass through a three-dimensional convolution and normalization layer with an output channel number of 3 to output the predicted deformation field , and finally, the deformation field is transformed through the space transformation network Acting on floating images Get the reference image Aligned registered images ;

[0016] Step 22, parameterize the search space; use discrete parameters to represent the 3D convolution kernel size of each layer in the double U-shaped network, as well as the normalization type, activation function type and 3D convolution kernel size of the feature fusion block in the feature adaptive fusion block; take the network layer as a module, and perform integer encoding on each module; use continuous parameters to represent the 3D convolution kernel size of the feature adaptive fusion block in the feature adaptive fusion block. , , , The four weighting coefficients all range in value from [0,1].

[0017] Furthermore, in step 3, the loss function for training the neural network is constructed as follows: in is a modality-independent neighborhood descriptor, is an image or The voxel set white The surrounding non-local area, yes white The displacement vector at Represents the absolute value of the input, Indicates input norm, is the gradient operator, It is The correlation regularization term of the feature adaptive fusion block, is the total number of feature adaptive fusion blocks in the network, and are regularization coefficients.

[0018] Furthermore, in step 4, the verification index for verifying the performance of the neural network is established: the Dice score DS or the target registration error TRE is used as the verification index; the Dice score measures the overlap of two sets of segmentation labels, and the larger the Dice score, the more accurate the registration is, which is defined as: in and are the first Individual elements, is the total number of voxels, and the target registration error is the Euclidean distance between the corresponding landmarks in the reference image and the registered floating image. The smaller the distance, the more accurate the registration. It is defined as: in and They are the reference image and the registered floating image. The three-dimensional coordinates of the landmark points, Represents the total number of landmark points on an image.

[0019] Furthermore, in step 5, the optimization problem of constructing the modality adaptive registration network architecture design is specifically: in is the fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks in step 2, represents the neural network architecture parameters, are the neural network parameters, is the neural network parameter space, yes The optimal neural network parameters obtained after the corresponding network is trained, is the loss function in step 3, is the accuracy loss after the registration network training, corresponding to step 4 or .

[0020] Furthermore, in step 6, Bayesian optimization is used to solve the optimization problem in step 5 to obtain the optimal network architecture, which specifically includes the following steps:

[0021] Step 61: Initialization; Initialization; Neural network architecture parameters Random Sampling times, we get a set of initial observation point samples , where the neural network corresponding to each observation point sample is trained with the loss function in step 3, and then its accuracy loss is obtained , thus forming the initial observation set ;

[0022] Step 62: Construct a proxy model; define a posterior probability distribution: in Represents a threshold value, which is determined by a certain quantile Determine, among which represents the probability, Divide the network architecture parameters in the observation set into advantage sets and disadvantage set , kernel density estimation is used to estimate from the advantage set and disadvantage set and : in is the Gaussian kernel function;

[0023] Step 63: construct the acquisition function EI: in Used to return the maximum value of two numbers. Represents the mathematical expectation of the input;

[0024] Step 64, maximize the acquisition function EI; calculate the corresponding value of the current observation space through the kernel density estimation in step 62 and , then Random sampling times, obtained Candidate parameters ; Then, bring the candidate parameters into and Calculate the corresponding Right now , and finally select the element with the largest value Corresponding network architecture parameters That is the new observation point for the next iteration;

[0025] Step 65: Add the new observation point Included in the observation set And update the probability distribution , after continuous iterative updates of the probability distribution, the optimal neural network architecture parameters are finally obtained .

[0026] Accordingly, a cross-modal medical image registration system based on Bayesian neural architecture search includes: a data preprocessing module, a search space module, and a Bayesian optimization module; the data preprocessing module obtains image data of the target cross-modal medical image registration task and performs image denoising, contrast adjustment and rigid registration; the search space module designs a fully convolutional double U-shaped parameterized neural architecture search space containing a feature adaptive fusion block; the Bayesian optimization module uses a loss function and a verification indicator to construct the optimization problem of the modality adaptive registration network architecture design under the constraints of the search space, and adopts Bayesian optimization to solve and obtain the optimal network architecture.

[0027] Furthermore, there is a feature adaptive fusion block between each level of the two U-shaped networks; in the encoder stage, the two input images are respectively passed through the convolution layer, the normalization layer and the activation function to obtain two feature map tensors, and then the feature adaptive fusion module takes these two feature map tensors as input, and the output two fused weighted feature map tensors are respectively reduced in size by the maximum pooling operation to enter the next level; in the decoder stage, the two fused feature map tensors output by the feature adaptive fusion block are expanded in size by trilinear interpolation to enter the next level, and after the two U-shaped networks respectively output the final mixed feature map tensors, the two mixed feature map tensors are tensor-added, and then pass through a three-dimensional convolution and normalization layer with an output channel number of 3 to output the predicted deformation field , and finally, the deformation field is transformed through the space transformation network Acting on floating images Get the reference image Aligned registered images .

[0028] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The cross-modal medical image registration method based on Bayesian neural architecture search of the present invention is based on the modal characteristics of the target registration task, and adaptively designs a network architecture that extracts and fuses cross-modal image features with minimum deviation. A neural network architecture with significant advantages can be obtained without manual intervention, and can be applied to cross-modal medical image registration; (2) A fully convolutional dual U-shaped parameter neural architecture search space with a feature adaptive fusion block is also designed: two U-shaped encoder-decoder networks with jump connections are used to extract image features of the two modalities respectively; the spatial attention mechanism in the feature adaptive fusion block cooperates with the correlation regularization term to adaptively reduce the difference between features, so that the cross-modal features gradually tend to be consistent in the progressive network layer, thereby reducing the deviation of the fused features and enhancing the registration performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0030] Figure 2 Schematic diagram of the double U-shaped neural architecture search space of the present invention.

[0031] Figure 3 Schematic diagram of the feature adaptive fusion block of the present invention. DETAILED DESCRIPTION

[0032] like Figure 1 As shown, a cross-modal medical image registration method based on Bayesian neural architecture search includes the following steps:

[0033] Step 1: Obtain image data for the target cross-modal medical image registration task and perform preprocessing;

[0034] Step 2: Design a fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks;

[0035] Step 3: Construct a loss function for training the neural network.

[0036] Step 4: Establish verification indicators for verifying the performance of the neural network;

[0037] Step 5, constructing the optimization problem of modality adaptive registration network architecture design according to the search space in step 2, the loss function in step 3 and the verification index in step 4;

[0038] Step 6: Use Bayesian optimization to solve the optimization problem in step 5 and obtain the optimal network architecture.

[0039] The step 1 obtains the image data of the target cross-modal medical image registration task and performs preprocessing, specifically: the registration task aligns two images of different modalities at a time, and the images of the two modalities can come from the same patient or from different patients, and then denoises the image data, adjusts the contrast, and finally pre-registers the image data through a rigid registration method.

[0040] The step 2 of designing a parameterized neural architecture search space includes the following sub-steps:

[0041] Step 21: An example of constructing a double U-shaped search space is as follows: Figure 2 As shown in Figure 2, two U-shaped encoder-decoder networks with skip connections are used to process cross-modal floating images respectively. and reference image Each U-shaped network has seven layers, each of which consists of a convolutional layer, a normalization layer, and a nonlinear activation function. The number of channels in each layer is 8, 16, 16, 32, 16, 16, and 8. There is a feature adaptive fusion block between each layer of the two U-shaped networks, which takes the feature map after the activation function as input, and the output is reduced in size after the maximum pooling and enters the next layer. Figure 3 As shown in the figure, the 3D convolution layer fuses the two input feature maps, and then forms spatial attention through maximum pooling and average pooling and a 7*7*7 convolution, so that the network can better understand the spatial relationship between cross-modal features. After the two U-shaped networks output the mixed features, the two mixed features are tensored and then passed through a 3-channel 3D convolution and normalization layer to output the predicted deformation field. Finally, the deformation field is transformed through the spatial transformation network. Acting on floating images Get the reference image Aligned ;

[0042] Step 22: Parameterize the search space as follows: The parameterized content includes: Figure 2 The 3D convolution kernel size, normalization, and activation function type of each layer in the double U-shaped network; Figure 3 The 3D convolution kernel size and , , , Four weighting coefficients. The 3D convolution kernel size is 3, which corresponds to 0, and the size is 5, which corresponds to 1; the normalization layer is 0 for batch normalization and 1 for instance normalization; the activation function is LeakyReLU, which corresponds to 0, and Mish, which corresponds to 1. , , , The four weighting coefficients are represented by four continuous parameters, and their value intervals are all [0,1], as shown in Table 1.

[0043] Table 1 Correspondence between discrete parameters and network layers

[0044] Discrete parameters Binary encoding Convolution kernel size, normalization layer, activation function 0 000 3*3*3 convolution, batch normalization, LeakyReLU 1 001 3*3*3 convolution, batch normalization, Mish 2 010 3*3*3 convolution, instance normalization, LeakyReLU 3 011 3*3*3 convolution, instance normalization, Mish 4 100 5*5*5 convolution, batch normalization, LeakyReLU 5 101 5*5*5 convolution, batch normalization, Mish 6 110 5*5*5 convolution, instance normalization, LeakyReLU 7 111 5*5*5 convolution, instance normalization, Mish .

[0045] The loss function of the step 3 to construct the training neural network is specifically: in is a modality-independent neighborhood descriptor, is an image or The voxel set white The surrounding non-local area, yes white The displacement vector at Represents the absolute value of the input, Indicates input norm, is the gradient operator, It is The correlation regularization term of the feature adaptive fusion block, is the total number of feature adaptive fusion blocks in the network, and are regularization coefficients.

[0046] The verification index for verifying the performance of the neural network in step 4 is specifically: using the Dice score DS or the target registration error TRE as the verification index. The Dice score measures the overlap of two sets of segmentation labels (the segmentation labels of the reference image and the segmentation labels of the registered floating image). The larger the Dice score, the more accurate the registration. It is defined as: in and are the first Individual elements, is the total number of voxels, and the target registration error is the Euclidean distance between the corresponding landmarks in the reference image and the registered floating image. The smaller the distance, the more accurate the registration. It is defined as: in and They are the reference image and the registered floating image. The three-dimensional coordinates of the landmark points, Represents the total number of landmark points on an image.

[0047] In step 5, the optimization problem of constructing the modality adaptive registration network architecture design is specifically: in is the fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks in step 2, represents the neural network architecture parameters, are the neural network parameters, is the neural network parameter space, yes The optimal neural network parameters obtained after the corresponding network is trained, is the loss function in step 3, is the accuracy loss after the registration network training, corresponding to step 4 or .

[0048] The step 6, using Bayesian optimization to solve the optimization problem in step 5 to obtain the optimal network architecture, includes the following sub-steps:

[0049] Step 61: Initialization: Neural network architecture parameters Random Sampling times, we get a set of initial observation point samples , where the neural network corresponding to each observation point sample is trained with the loss function in step 3, and then its performance loss is taken , thus forming the initial observation set ;

[0050] Step 62: Construct a proxy model; define a posterior probability distribution: in Represents a threshold value, which is determined by a certain quantile Determine, among which represents the probability, Divide the network architecture parameters in the observation set into advantage sets and disadvantage set , kernel density estimation is used to estimate from the advantage set and disadvantage set and : in is the Gaussian kernel function;

[0051] Step 63: construct the acquisition function EI: in Used to return the maximum value of two numbers. Represents the mathematical expectation of the input;

[0052] Step 64, maximize the acquisition function EI; calculate the corresponding value of the current observation space through the kernel density estimation in step 62 and , then Random sampling times, obtained Candidate parameters ; Then, bring the candidate parameters into and Calculate the corresponding Right now , and finally select the element with the largest value Corresponding network architecture parameters That is the new observation point for the next iteration;

[0053] Step 65: Add the new observation point Included in the observation set And update the probability distribution , after continuous iterative updates of the probability distribution, the optimal neural network architecture parameters are finally obtained .

[0054] The present invention designs a cross-modal medical image registration method based on Bayesian neural architecture search. Since the objective function belongs to a black-box optimization problem and each iteration needs to solve the internal optimization problem of neural network training, the Bayesian optimization algorithm is used to solve it. In order to solve this optimization problem, a parameterized search space is first established to generate an initial observation set, and a Bayesian optimization algorithm is used to obtain new network architecture parameters based on the observation set and update the observation set. By continuously updating the observation set, the optimal network architecture is finally iterated.

[0055] Multiple experiments were conducted on two datasets, including the TCIA dataset for abdominal MRI-CT registration and the EASY-RESECT dataset for brain MRI-US. The TCIA dataset uses the Dice score to evaluate the registration accuracy, and the EASY-RESECT dataset uses the target registration error to evaluate the registration accuracy. As shown in Table 2, VoxelMorph is a registration model based on a fully convolutional neural network, while XMorpher and TransMorph both use visual transformers. Specifically, the encoder of TransMorph is implemented using a visual transformer, and the decoder is implemented using a fully convolutional neural network, while XMorpher is a registration model based entirely on a visual transformer. It can be seen from the experimental results that the method of the present invention has better accuracy and better robustness than existing advanced algorithms. In terms of model parameters, the method of the present invention is more than ten times better than XMorpher and TransMorph.

[0056] Table 2 Comparative experiment of the system of the present invention and the existing advanced cross-modal medical image registration method on two data

[0057] method TCIA (DS) EASY-RESECT (TRE) Model parameters Initial Accuracy 0.7645 5.56 none VoxelMorph 0.7891 5.22 300,000 XMorpher 0.7959 5.54 15.09 million TransMorph 0.8352 5.50 11.64 million Method of the present invention 0.8404 4.83 1.02 million

Claims

1. A cross-modal medical image registration method based on Bayesian neural architecture search, characterized in that: The steps include: Step 1: Obtain image data for the target cross-modal medical image registration task and perform preprocessing; Step 2: Design a fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks; Step 3: Construct a loss function for training the neural network. Step 4: Establish verification indicators for verifying the performance of the neural network; Step 5, constructing the optimization problem of modality adaptive registration network architecture design according to the search space in step 2, the loss function in step 3 and the verification index in step 4; Step 6: Use Bayesian optimization to solve the optimization problem in step 5 and obtain the optimal network architecture; specifically, the following steps are included: Step 61: Initialization: Neural network architecture parameters Random Sampling times, we get a set of initial observation point samples , where the neural network corresponding to each observation point sample is trained with the loss function in step 3, and then its accuracy loss is obtained , thus forming the initial observation set ; Step 62: Construct a proxy model; define a posterior probability distribution: in Represents a threshold value, which is determined by a certain quantile Determine, among which represents the probability, Divide the network architecture parameters in the observation set into advantage sets and disadvantage set , kernel density estimation is used to estimate from the advantage set and disadvantage set and : in is the Gaussian kernel function; Step 63: construct the acquisition function EI: in Used to return the maximum value of two numbers. Represents the mathematical expectation of the input; Step 64, maximize the acquisition function EI; calculate the corresponding value of the current observation space through the kernel density estimation in step 62 and , then Random sampling times, obtained Candidate parameters ; Then, bring the candidate parameters into and Calculate the corresponding Right now , and finally select the element with the largest value Corresponding network architecture parameters That is the new observation point for the next iteration; Step 65: Add the new observation point Included in the observation set And update the probability distribution , after continuous iterative updates of the probability distribution, the optimal neural network architecture parameters are finally obtained .

2. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 1, the registration task aligns two images of different modalities at a time. The images of the two modalities are from the same patient or from different patients. Then the image data is denoised, the contrast is adjusted, and finally the image data is pre-registered through the rigid registration method.

3. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 2, the feature adaptive fusion block includes a feature fusion block, a feature selection block, and a correlation regularization term; In the feature fusion block, two feature map tensors that have been weighted by hyperparameters are concatenated in the channel dimension, and then the number of channels is halved using a three-dimensional convolution operation to achieve feature fusion. In the feature selection block, the spatial attention mechanism is used to weight the fused features of different parts of the image through the attention mechanism to achieve feature selection. Subsequently, the fused and attention-weighted feature map tensor is added to the two original feature map tensors in a certain proportion to obtain the final fused feature map tensor. In the correlation regularization term, in order to gradually reduce the deviation between the cross-modal fused feature map tensors, a correlation regularization term is constructed. ,in is the correlation coefficient between the two fused feature map tensors, For bias.

4. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 2, designing a fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks specifically includes the following steps: Step 21: Construct a fully convolutional double U-shaped search space; use two U-shaped encoder-decoder networks with jump connections to process cross-modal floating images respectively. and reference image There is a feature adaptive fusion block between each level of the two U-shaped networks; in the encoder stage, the two input images are respectively passed through the convolution layer, the normalization layer and the activation function to obtain two feature map tensors, and then the feature adaptive fusion module takes these two feature map tensors as input, and the output two fused weighted feature map tensors are respectively reduced in size by the maximum pooling operation to enter the next level; in the decoder stage, the two fused feature map tensors output by the feature adaptive fusion block are expanded in size by trilinear interpolation to enter the next level, and after the two U-shaped networks respectively output the final mixed feature map tensors, the two mixed feature map tensors are tensor-added, and then pass through a three-dimensional convolution and normalization layer with an output channel number of 3 to output the predicted deformation field , and finally, the deformation field is transformed through the space transformation network Acting on floating images Get the reference image Aligned registered images ; Step 22, parameterize the search space; use discrete parameters to represent the three-dimensional convolution kernel size of each layer in the double U-shaped network, as well as the normalization type, activation function type and the three-dimensional convolution kernel size of the feature fusion block in the feature adaptive fusion block; take the network layer as a module, and perform integer encoding on each module; Using continuous parameters to represent the feature adaptive fusion block , , , The four weighting coefficients all range in value from [0,1].

5. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 3, the loss function for training the neural network is constructed as follows: in is a modality-independent neighborhood descriptor, is an image or The voxel set white The surrounding non-local area, yes white The displacement vector at Represents the absolute value of the input, Indicates input norm, is the gradient operator, It is The correlation regularization term of the feature adaptive fusion block, is the total number of feature adaptive fusion blocks in the network, and are regularization coefficients.

6. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 4, the verification indicators for verifying the performance of the neural network are established as follows: the Dice score DS or the target registration error TRE is used as the verification indicator; the Dice score measures the overlap of two sets of segmentation labels. The larger the Dice score, the more accurate the registration. It is defined as: in and are the first Individual elements, is the total number of voxels, and the target registration error is the Euclidean distance between the corresponding landmarks in the reference image and the registered floating image. The smaller the distance, the more accurate the registration. It is defined as: in and They are the reference image and the registered floating image. The three-dimensional coordinates of the landmark points, Represents the total number of landmark points on an image.

7. The cross-modal medical image registration method based on Bayesian neural architecture search according to claim 1, characterized in that: In step 5, the optimization problem of constructing the modality adaptive registration network architecture design is specifically: in is the fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks in step 2, represents the neural network architecture parameters, are the neural network parameters, is the neural network parameter space, yes The optimal neural network parameters obtained after the corresponding network is trained, is the loss function in step 3, is the accuracy loss after the registration network training, corresponding to step 4 or .

8. A system for implementing the cross-modal medical image registration method based on Bayesian neural architecture search as claimed in claim 1, characterized in that: include: Data preprocessing module, search space module, and Bayesian optimization module; The data preprocessing module obtains the image data of the target cross-modal medical image registration task and performs image denoising, contrast adjustment and rigid registration; the search space module designs a fully convolutional double U-shaped parameterized neural architecture search space with feature adaptive fusion blocks; the Bayesian optimization module uses the loss function and verification indicators to construct the optimization problem of modality adaptive registration network architecture design under the constraints of the search space, and uses Bayesian optimization to solve it and obtain the optimal network architecture.

9. The cross-modal medical image registration system based on Bayesian neural architecture search according to claim 8, characterized in that: There is a feature adaptive fusion block between each level of the two U-shaped networks; in the encoder stage, the two input images are respectively passed through the convolution layer, the normalization layer and the activation function to obtain two feature map tensors, and then the feature adaptive fusion module takes these two feature map tensors as input, and the output two fused weighted feature map tensors are respectively reduced in size by the maximum pooling operation to enter the next level; in the decoder stage, the two fused feature map tensors output by the feature adaptive fusion block are expanded in size by trilinear interpolation to enter the next level, and after the two U-shaped networks respectively output the final mixed feature map tensors, the two mixed feature map tensors are tensor-added, and then pass through a three-dimensional convolution and normalization layer with an output channel number of 3 to output the predicted deformation field , and finally, the deformation field is transformed through the space transformation network Acting on floating images Get the reference image Aligned registered images .

Images