Brain medical image deformable registration method and system

Abstract

The invention discloses a brain medical image deformable registration method and system, and relates to the technical field of image registration, and the method comprises the following steps: splicing a fixed image and a moving image in a channel dimension, and obtaining a first tensor; performing multi-stage encoding and decoding processing on the first tensor to obtain a plurality of encoding features and corresponding decoding features; carrying out convolution on the decoding feature output in the final stage to obtain a first deformation field; performing interpolation processing on the moving image based on the deformation information of the first deformation field to obtain a deformation image; splicing the fixed image and the deformed image in the channel dimension to obtain a second tensor; performing multiple stages of encoding and decoding processing on the second tensor; carrying out convolution on the decoding features output in the final stage to obtain a second deformation field; and performing interpolation processing on the deformation image based on the deformation information of the second deformation field to obtain a registration image. According to the method, through two times of registration, the precision of registration between patients is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of image registration technology, and in particular to a deformable registration method and system for brain medical images. Background Technology

[0002] Utilizing medical images for disease diagnosis and treatment has always been a major development direction in the field of medical image processing. Deformable registration technology for brain images has numerous applications in medicine, such as brain tumor localization, neurosurgical assistance, and treatment response assessment. Deformable registration refers to establishing an optimal spatial correspondence between a fixed image and a moving image within the same space, aligning the differing regions between the two images. Depending on the medical scenario, it can perform atlas-to-patient registration, patient-to-patient registration, multimodal image registration, and multi-view image registration. Furthermore, the results of image registration can be applied to many other medical image processing techniques, such as image segmentation and image fusion.

[0003] In recent years, deep learning-based methods have been widely applied in image analysis and processing, demonstrating advantages and potential not possessed by traditional algorithms. Since AlexNet achieved outstanding performance, Convolutional Neural Networks (CNNs) have become one of the most successful models in deep learning. CNNs, with their ability to efficiently process highly structured data, have also made significant breakthroughs in medical image registration. CNN-based medical image registration methods can be divided into two categories: iterative registration methods that use CNNs for similarity prediction, and methods that use CNNs for transform parameter prediction. The former can be called deep similarity-based methods, which mainly use CNNs to measure the similarity between multimodal images. Traditional intensity-based similarity measures, such as cross-correlation (CC), sum-of-squares distance (SSD), and mean squared distance (MSD), are very effective for registering single-modal images with the same intensity distribution. However, for multimodal image registration, the performance of traditional similarity measures is not ideal. Therefore, some researchers have considered using CNNs to measure the similarity between images and have achieved good results. However, these methods still need to be combined with traditional image registration algorithms for iterative optimization, and the problem of low registration efficiency still exists. The second category of CNN-based registration methods can be further divided into supervised learning-based registration methods and unsupervised learning-based registration methods. Unlike deep similarity-based methods, these two methods infer transformation parameters in a single step during forward prediction, eliminating the iterative optimization process and significantly improving registration efficiency. The difference between these two methods lies in the fact that supervised models require label information to aid learning during training. However, obtaining this necessary real information or true values ​​is often difficult. This data often requires manual labeling by experts or generation through traditional registration algorithms or random transformations. The true values ​​obtained through these methods may contradict actual physiological changes and are not realistic enough. Furthermore, this data also limits the upper limit of registration. In contrast, unsupervised learning inherits the advantages of supervised learning and is not limited by true values, thus gaining favor among researchers. Increasing research is exploring how to utilize unsupervised methods to improve registration performance.

[0004] Accurate registration between patients is extremely challenging due to the diverse brain structures and shapes. Traditional registration algorithms, based on mathematical models and optimization methods, excel in maintaining local structure and overall consistency, but they are slow in large-scale data or real-time applications and suffer from low accuracy when handling complex deformations or nonlinear transformations. Deep learning-based registration algorithms, by learning features and transformation patterns from large datasets, achieve faster registration and higher accuracy, outperforming traditional algorithms in both speed and accuracy. Deep learning models learn deformation patterns from training data; however, if the training data lacks samples of extreme deformations or complex topological changes, the model may fail to correctly handle unseen complex deformations, leading to low registration accuracy. Summary of the Invention

[0005] Based on the shortcomings of the existing technology, the present invention provides a deformable registration method and system for brain medical images, which solves the problem of low registration accuracy of deep learning-based registration algorithms when dealing with complex and detailed targets.

[0006] The present invention adopts the following technical solution: In a first aspect, the present invention provides a deformable registration method for brain medical images, comprising the following steps: Baseline brain medical images and brain medical images to be registered were acquired from different patients. The baseline brain medical images were used as fixed images, and the brain medical images to be registered were used as moving images. The fixed and moving images are stitched together along the channel dimension to obtain the first tensor. The first tensor undergoes multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features. The decoded features output in the final stage are convolved to obtain the first deformation field. In each stage of decoding, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features. The weighted features are then fused with the decoded features of the previous stage to obtain the decoded features of the current stage. The moving image is interpolated based on the deformation information of the first deformation field to obtain a deformed image; The fixed image and the deformed image are concatenated along the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field. The deformation image is interpolated based on the deformation information of the second deformation field to obtain the registration image.

[0007] Preferably, the first deformation field is obtained through a first registration sub-network, which includes an encoder and a decoder; The encoder includes a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer. The first coding layer includes a first convolutional layer and a first stride convolution. The second coding layer includes a first dilated convolutional module and a second stride convolution. The third coding layer includes a second dilated convolutional module and a third stride convolution. The fourth coding layer includes a second convolutional layer and a fourth stride convolution. The first and second dilated convolutional modules are both dilated convolutions with different dilation rates, concatenated together. The decoder includes decoding layers at different stages, multiple attention weighting units, and multiple feature fusion units. The first, second, third, and fourth encoding layers perform multiple stages of encoding processing on the first tensor to obtain multiple encoded features; the decoding layers at different stages are used to output multiple decoded features; during the decoding process at each stage, the attention weighting unit is used to weight the encoded features of the current stage with the decoded features of the previous stage to obtain weighted features; the feature fusion unit is used to fuse the weighted features with the decoded features of the previous stage to obtain the decoded features of the current stage.

[0008] Preferably, the first convolutional layer is used to perform 3×3×3 convolutional feature extraction on the first tensor to obtain a shallow feature T1 with 16 channels; the first long convolution is used to downsample the shallow feature T1 to obtain a feature T1 with a resolution of ½ and still 16 channels. The first dilated convolution module applies a 3×3×3 convolution with a dilation rate d=1 / 2 / 3 to the downsampled feature T1 to obtain a 32-channel feature T2; the second stride convolution downsamples the feature T2 to obtain a feature T2 with a resolution of ¼ and 32 channels. The second dilated convolution module is used to apply a 3×3×3 dilated convolution to the downsampled feature T2 to obtain a 64-channel feature T3; the third stride convolution is used to downsample the feature T3 to obtain a feature T3 with a resolution of ⅛ and 64 channels. The second convolutional layer performs a 3×3×3 convolution on the downsampled feature T3 to obtain feature T4; the fourth stride convolution is used to downsample feature T4 to obtain a resolution of ¹ / ². 16 The feature T4 with 64 channels is used as the input to the decoder.

[0009] Preferably, the second deformation field is obtained through a second registration sub-network, and the first registration sub-network and the second registration sub-network have the same structure.

[0010] Preferably, interpolation processing is performed on the moving image and the deformed image using two spatial transformation networks respectively; The first registration subnetwork, the second registration subnetwork, and two spatial transformation networks are interconnected to obtain a deformable image registration model. The deformable image registration model is trained using similarity loss and regularization constraints. The specific similarity loss is as follows: ; In the formula, L sim For similarity loss, NCC For normalized cross-correlation, F M is a fixed image, and M is a moving image. p It represents the position of a specific pixel within the entire image. p i For a certain region in the image i1 pixel For the image domain, and For fixed images and deformed images p The average voxel intensity within the local window centered on the value. It is a compound symbol. For deformation field; The specific regularization constraints are as follows: ; in, ; ; In the formula, R For regularization constraints, and The tradeoff coefficient between the two regularization terms. For diffusion regularization, For bending energy regularization, For gradient operators, x , y , z These are the coordinate axes in three-dimensional space. For the Laplace operator.

[0011] Preferably, the acquisition of the deformed image is specifically as follows: ; In the formula, I 1 is a deformed image. U 1 represents the first registration subnetwork. This is the first deformation field; The acquisition of the registration image is as follows: ; In the formula, I2 is the registered image. U 2 represents the first registration sub-network. For the first deformation field, This is the second deformation field.

[0012] Preferably, before stitching the fixed image and the moving image in the channel dimension, the fixed image and the moving image are preprocessed by skull stripping, normalization, cropping and affine alignment.

[0013] In a second aspect, the present invention provides a deformable registration device for brain medical images, comprising: The acquisition module is used to acquire baseline brain medical images and brain medical images to be registered from different patients. The baseline brain medical images are used as fixed images, and the brain medical images to be registered are used as moving images. The first deformation module is used to stitch the fixed image and the moving image in the channel dimension to obtain the first tensor; the first tensor is subjected to multiple stages of encoding and decoding processing to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the first deformation field; wherein, in the decoding process of each stage, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features; the weighted features are fused with the decoded features of the previous stage to obtain the decoded features of the current stage; The first registration module is used to perform interpolation processing on the moving image based on the deformation information of the first deformation field to obtain a deformed image; The second deformation module is used to stitch the fixed image and the deformed image in the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field. The second registration module is used to interpolate the deformed image based on the deformation information of the second deformation field to obtain a registered image.

[0014] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention uses a reference brain medical image as a fixed image and the brain medical image to be registered as a moving image. The moving image is registered to the fixed image through a deformation field. This invention performs image registration by generating two deformation fields. In the first registration process, the fixed image and the moving image are used as input and concatenated to obtain a first tensor. Then, the first tensor undergoes multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features. The decoded features output in the final stage are convolved to obtain a first deformation field. Based on the deformation information of the first deformation field, the moving image is interpolated to obtain a deformed image. In each stage of decoding, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features. Weighting can suppress the saliency of irrelevant regions, enhance the response to key regions, identify key deformed regions in the image, and improve the final registration accuracy. Then, the weighted features are fused with the decoded features of the previous stage to obtain the decoded features of the current stage. In the second registration process, the fixed image and the deformed image are used as inputs, and they are stitched together to obtain a second tensor. The same operations as in the first registration process are then performed to obtain a second deformation field. Based on the deformation information of the second deformation field, the deformed image is interpolated to obtain the registered image. This invention significantly improves the accuracy of registration between patients through two registration processes.

[0015] This invention employs a first registration sub-network to obtain a first deformation field. The first registration sub-network includes a first dilated convolution module and a second dilated convolution module. Both the first and second dilated convolution modules are connected in series with dilated convolutions of different dilation rates. The combination of dilated convolutions can expand the receptive field of the network without increasing too much computational cost and network parameters, thereby obtaining multi-scale features. This enables the network to better understand the global structure and local details of the image and improve its ability to accurately detect deformed regions. Attached Figure Description

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 This is a schematic diagram of the network architecture of the multi-constraint cascaded attention network of the present invention; Figure 2 This is a structural diagram of the extended self-attention network of the present invention; Figure 3 This is a structural diagram of the attention gating module of the present invention; Figure 4 This is a schematic diagram showing the visualization results of several registration methods used in this invention; in, Figure 4 (a): Fixed image Figure 4 (b): The moving image registered using the SyN algorithm. Figure 4 (c): The moving image registered using the LDDMM algorithm. Figure 4 (d): The moving image registered using the VoxelMorph-1 algorithm. Figure 4 (e): The moving image registered using the VoxelMorph-2 algorithm. Figure 4 f: The moving image registered using the FAIM algorithm. Figure 4 (g): The moving image registered using the MCANet algorithm. Figure 4 (h): moving image, Figure 4 (i): The deformable field generated using the SyN algorithm. Figure 4 (j): The deformable field generated using the LDDMM algorithm. Figure 4 (k): The deformable field generated using the VoxelMorph-1 algorithm. Figure 4 (l): Deformation field generated using the VoxelMorph-2 algorithm. Figure 4(m): The deformation field generated using the FAIM algorithm, Figure 4 (n): The deformation field generated using the MCANet algorithm; Figure 5 The Dice scores and folded voxel percentages of MCANet under different β values of the present invention; Figure 6 For the present invention, under different weight coefficients The visualized deformation fields; Among them, Figure 6 (a): The deformation field when β = 0, Figure 6 (b): The deformation field when β = 0.4, Figure 6 (c): The deformation field when β = 0.8, Figure 6 (d): The deformation field when β = 1.2, Figure 6 (e): The deformation field when β = 1.4, Figure 6 (f): The deformation field when β = 1.6, Figure 6 (g): The deformation field when β = 2, Figure 6 (h): The registration error map when β = 0, Figure 6 (i): The registration error map when β = 0.4, Figure 6 (j): The registration error map when β = 0.8, Figure 6 (k): The registration error map when β = 1.2, Figure 6 (l): The registration error map when β = 1.4, Figure 6 (m): The registration error map when β = 1.6, Figure 6 (n): The registration error map when β = 2. Detailed implementation manners

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0019] Related background: In the absence of supervision information, the loss function for training a registration network based on unsupervised learning has always been a research issue. With the emergence of the Spatial Transformer Network (STN), this problem has been perfectly solved. As a differentiable module, STN can be arbitrarily inserted into any position of the neural network and can perform gradient backpropagation along with the entire network. The insertion of STN enables the deformation of the moving image during the training process to generate the deformed image and perform similarity calculation.

[0020] Accurate registration between patients is extremely challenging due to the diverse brain structures and shapes. Furthermore, voxel folding in deformation fields severely impacts image topology and the realism of deformation. While traditional registration algorithms can effectively handle voxel folding and deformation smoothness based on mathematical models and optimization methods, deep learning models, which typically learn constraints implicitly through loss functions, may not strictly guarantee these properties. Therefore, deep learning-based registration algorithms may be insufficient in handling voxel folding and deformation smoothness.

[0021] Example 1 To improve registration speed and accuracy, and to enhance voxel folding and registration smoothness, this invention proposes a deformable registration method for brain medical images, specifically comprising the following steps: S1: Acquire baseline brain medical images and brain medical images to be registered from different patients. The baseline brain medical images are used as fixed images, and the brain medical images to be registered are used as moving images.

[0022] The data used in the brain MRI registration experiment of this invention comes from the Mindboggle101 dataset, which is freely available and contains T1-weighted brain MR images of 101 participants. This invention selects four subsets of the Mindboggle101 dataset: MMRR-21, NKI-RS-22, NKI-TRT-20, and OASIS-TRT-20, totaling 83 3D brain MR images. For image preprocessing, all images underwent skull stripping and were strictly aligned to the MNI152 template space. Thirty-one cortical regions were manually labeled according to the Desikan-Killiany-Tourville (DKT) protocol. Furthermore, this invention normalizes image intensity and crops the image size from the original 181×217×181 to 160×208×160 to fit the proposed network. Simultaneously, linear affine alignment of all images was performed using the ANTs package. The three subsets, MMRR-21, NKI-RS-22, and NKI-TRT-20, containing a total of 63 brain images (3906 pairs), were used as training data for the model. The 20 images (380 pairs) from OASIS-TRT-20 were used as test data.

[0023] The input to a deformable image registration model is typically a pair of images that have already been affine aligned. The reference image is called the fixed image, which serves as the basis for spatial alignment. The image to be registered is called the moving image, which is aligned to the fixed image using a deformation field (DVF).

[0024] S2: The fixed and moving images are stitched together along the channel dimension to obtain the first tensor; the first tensor undergoes multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the first deformation field. The moving image is interpolated based on the deformation information of the first deformation field to obtain the deformed image.

[0025] S3: The fixed image and the deformed image are stitched together in the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field; the deformed image is interpolated based on the deformation information of the second deformation field to obtain the registration image.

[0026] Reference Figure 2 This invention proposes a fully convolutional, multi-constraint cascaded attention network (MCANet) to achieve the above steps. This network architecture consists of two structurally identical registration subnetworks, referred to as the dilated self-attention network (DSNet). This invention introduces an attention-gating-based attention mechanism into the dilated self-attention network, which improves the network's sensitivity to regions with significant differences in patient images by suppressing feature activation in the alignment region. This allows for better adaptation to image differences among different patients and improves registration accuracy. The cascading of the two registration subnetworks achieves coarse-to-fine registration of input image pairs, significantly improving the registration accuracy of the convolutional neural network between patients. Applying additional regularization constraints reduces noise and discontinuities in the registration process, resulting in smoother and more realistic deformation results.

[0027] For a pair of fixed images and moving images The registration subnet can perform feature extraction and deformation prediction on moving images. M Generate a displacement vector field (DVF). This is used for image deformation. In the registration task, the training process of the registration subnet is a process of continuously updating network parameters to obtain the optimal Deformation Video Function (DVF). After obtaining the DVF, the Spatial Transformation Network (STN) can be used to perform trilinear interpolation on the moving image based on the deformation information in the DVF, thereby obtaining the deformed image. For the MCANet proposed in this invention, the moving image needs to go through two deformation stages in total, such as... Figure 1 As shown.

[0028] In the first phase, and These are input together into the registration sub-network as initial inputs. The registration sub-network generates the deformation field. Afterwards, The first deformation is performed to obtain a rough deformation image. This stage can be represented by formula (1):

[0029] (1); In obtaining deformation images After that, it will enter the second stage of registration. At this time, As a moving image, with Input together into the registration sub-network In the middle. After The prediction can generate and Deformation field between .according to Voxel displacement in the image, for coarse deformation The deformation is performed again to obtain a fine-grained deformation image. This process can be represented by formula (2):

[0030] (2); because This is the second stage of the moving image, after the spatial transformation in the first stage. Original moving images may be lost. Partial spatial information. Therefore, in this invention, the deformation field... Input as auxiliary information The end, to assist in generation Through the above process, the network architecture completes the coarse-to-fine registration of the original input image.

[0031] The deformable registration network of this invention consists of two structurally identical encoder-decoder networks, which this invention calls the Extended Self-Attention Network (DSNet), as follows: Figure 2As shown in the diagram, this invention employs dilated convolution to expand the receptive field of the convolutional layers, thereby improving network performance. In deep neural networks, downsampling is typically used to increase the receptive field. However, due to the limited number of downsampling operations in shallow networks, the receptive field of low-level features cannot be effectively expanded, resulting in extracted features that fail to adequately reflect the differences between input images. Therefore, some studies have increased the receptive field of low-level features by adding an Inception module to the network. However, for 3D convolutions, especially in shallow layers (where the feature map size is large), using any large-scale convolution incurs significant memory and computational costs. Dilated convolution bypasses this problem. Dilated convolution explicitly expands the receptive field of the feature map without adding additional parameters, allowing the network to capture more contextual information during convolution. Simultaneously, since there is no downsampling, the loss of image detail information is avoided, preserving the ability of low-level features to perceive differences in image details. In computer vision, acquiring multi-scale features is crucial for target region detection. By adjusting the dilation rate of dilated convolution, the size of the receptive field can be directly changed, enabling the network to capture features at different scales.

[0032] Previous research has shown that dilated convolution uses sparse sampling, and blindly using a single dilation rate can lead to a grid problem in the pixel receptive field. On the one hand, voxels in deep feature maps can only view the underlying information in a checkerboard pattern, resulting in the loss of a large amount of local information; on the other hand, the information between distant voxels is not directly related, and the grid problem weakens the continuity and consistency of local information. Therefore, this invention uses three different dilation rates. The dilated convolutions are concatenated and combined into a module. By concatenating the outputs of three dilated convolutions with different dilation rates, the receptive field of each layer's output can be expanded without increasing the number of parameters. Furthermore, the parameters can be used more efficiently to capture long-term dependencies and global information in the input data, thus reducing model complexity to some extent. Dilated convolutions with different dilation rates can capture features at different scales; combining them can better extract multi-scale features, helping to improve the model's ability to represent input data.

[0033] Reference Figure 2The data flow begins with the input image and passes through a series of convolutional layers, dilated convolutions, strided convolutions, transposed convolutions, and attention gating modules. Each convolutional layer (blue) takes the output feature map of the previous layer as input and outputs 16, 32, or 64 feature maps. Dilated convolutions (green) also take the output feature map of the previous layer as input and output 16 or 32 feature maps. Stride convolutions (orange) take the output feature map of the previous layer as input and output 32 or 64 feature maps. Transposed convolutions (blue) take the output feature map of the previous layer as input and output 32 or 64 feature maps. Deep feature fusion (purple circle) takes feature maps from different layers as input and outputs the fused feature map. Attention gating (yellow) takes the output feature map of the previous layer as input and outputs an attention-enhanced feature map. Finally, after processing by these modules, the network outputs 64 feature maps.

[0034] The encoder includes a first coding layer, a second coding layer, a third coding layer, and a coding layer. The first coding layer includes a first convolutional layer and a first long convolution. The first convolutional layer is used to extract features from the first tensor using a 3×3×3 convolution, resulting in a 16-channel shallow feature T1. The first long convolution is used to downsample the shallow feature T1, resulting in a shallow feature T1 with a resolution of ½ and still 16 channels.

[0035] The second encoding layer includes a first dilated convolution and a second stride convolution. The first dilated convolution is used to apply a 3×3×3 convolution with an dilation rate of d=1 / 2 / 3 to the downsampled shallow feature T1 to obtain a 32-channel context feature T2. The second stride convolution downsamples the context feature T2 to obtain a context feature T2 with a resolution of ¼ and 32 channels.

[0036] The third encoding layer includes a second dilated convolution and a third stride convolution. The second dilated convolution is used to apply a 3×3×3 dilated convolution to the downsampled context feature T2 to further capture long-range dependencies and obtain a 64-channel feature T3. The third stride convolution is used to downsample the feature T3 to obtain a feature T3 with a resolution of ⅛ and 64 channels. The fourth encoding layer consists of a second convolutional layer and a fourth stride convolution. The second convolutional layer is used to refine the downsampled feature T3 using a 3×3×3 convolution, resulting in the bottleneck feature T4. The fourth stride convolution is used to downsample the bottleneck feature T4, resulting in a feature with a resolution of ¹ / ². 16 The bottleneck characteristic T4 is 64 channels.

[0037] The bottleneck feature T4 is refined using a second 3×3×3 convolution. Further integration is performed on ¹ / 16Global semantic information at resolution. As the gating feature (g) of the attention gate, i.e. the decoder-side guiding signal, AG uses these 64-channel features to "gate" the output of the fourth coding layer (also 64 channels) to generate weighted features. At the same time, it feeds itself into the transposed convolution to start the first step of upsampling in the decoder.

[0038] The decoder consists of four symmetrical upsampling units, each containing: ① Transposed convolution or nearest neighbor upsampling amplifies the current feature by a factor of 2; ②Attention gate (AG): Receives the corresponding encoder features and the current decoder features, and generates weighted features; ③ The “C” concatenation operation concatenates the upsampled features and weighted features, and then fuses them through a 3×3×3 convolution, restoring the number of channels to 64→32→16 in sequence.

[0039] The third convolutional layer (deformation field prediction head, including a 16-channel convolutional layer and a 3-channel transposed convolution) is used to perform 3×3×3 convolution on the 16-channel features at the end of the decoding process to obtain a 3-channel deformation field Flow=(Δx,Δy,Δz), which is then passed to a differentiable spatial transformer to deform the source image M to align with the target image F.

[0040] Attention gating has proven successful in dense image prediction tasks. Network models trained with attention gating can implicitly learn to focus on task-relevant regions in an image and suppress responses to task-irrelevant features. This invention incorporates attention gating into the design of the registration sub-network DSNet. Figure 3 The structure of the attention gating module is illustrated. Attention gating takes feature maps from two adjacent scales as input, allowing selective learning of relevant regions in the input image and suppression of the salience of irrelevant regions. This avoids introducing additional human supervision during network construction. Attention gating enhances the model's response to key regions, improving the final registration accuracy. Deep feature maps along the decoding path serve as the gating signal input to the network, containing rich contextual information. Attention gating dynamically adjusts the model's focus on input features, filtering image features closely related to the registration task and identifying key deformation regions in the image. For the gating feature map and weighted feature map, this invention employs an additive attention mechanism to calculate attention weights. Compared to multiplicative attention, although additive attention requires more computational cost, its performance is superior. Therefore, this invention uses an additive attention mechanism to calculate attention weights. The attention mechanism continuously optimizes the attention weights, helping the network identify task-related salient features. The formula for calculating attention weights is:

[0041] (3); (4); in, and They represent the first in the network. Low-level feature maps and gated feature maps of the layer. , , The linear transformation parameters represent the convolution. Low-level feature map. Downsampling via stride convolution enables Size and gating feature map They are the same size. and It is the bias term of the operation. It is the ReLU activation function. It is the Sigmoid activation function, used to enhance nonlinear expressive power. Then, trilinear interpolation is applied... Resampling is performed to obtain the attention weight coefficients. Finally, by... and Element-wise multiplication yields the attention-weighted feature map. .

[0042] (5); By adding attention gating to the skip connections at each scale, multi-scale imaging information can be aggregated in the gated signal. This helps the network achieve more accurate feature selection and weight allocation, resulting in better deformation prediction. Furthermore, attention gating is also a differentiable module whose parameters can be updated during backpropagation.

[0043] In this invention, two registration sub-networks are trained as a whole, and the loss function is minimized through end-to-end training to achieve the highest registration accuracy. The loss function typically consists of two parts: one is the similarity loss between the fixed image and the moving image, and the other is the regularization of the DVF (displacement vector field). In this invention, normalized cross-correlation (NCC) is used to measure the similarity between the deformed image and the fixed image, and its negative value is used as the similarity loss. The correlation calculation can be performed according to formula (6):

[0044] (6); In the formula, Represents voxel points in an image. Represents the image domain. and Representing the fixed image and the deformed image respectively, with The average voxel intensity within a local window centered on the target. In this invention, the window size is 9. 3 .

[0045] If the network is optimized solely through similarity loss, the network parameters will be updated to make the deformed image infinitely close to the fixed image. However, in this case, the DVF (displacement vector field) predicted by the CNN will be excessively distorted, and even voxels in the DVF may be folded. Therefore, it is essential to apply some spatial regularization to the DVF. Considering that the twin-network registration model used in this invention may produce large deformation predictions, applying only a single regularization constraint is insufficient to meet the smoothness requirements of the deformation field; therefore, a dual regularization constraint is implemented. The first regularization term is diffusion regularization, which penalizes the spatial gradient of the DVF:

[0046] (7); in, For gradient operators, x , y , z These are the coordinate axes in three-dimensional space. For the Laplace operator.

[0047] The second regularization is the bending energy, which penalizes severe bending deformation by imposing a loss on the second derivative of the voxel displacement in the DVF. It is defined as:

[0048] (8); in, For the Laplace operator.

[0049] The dual regularization constraint in this invention can be expressed as: (9); in, and Represents the trade-off coefficients between the two regularization terms. The diffusion regularization term penalizes the spatial gradient of the deformation field to promote a smoother deformation field, which helps avoid overly complex local deformations and maintain the continuity of the overall deformation. The bending energy regularization term not only promotes the smoothness of the deformation field while suppressing severe bending deformation, but also effectively controls the complexity of the model and reduces the risk of overfitting. The bending energy weight coefficients... With the increase of , the degree of voxel folding is significantly reduced, but the similarity coefficient Dice also decreases. Therefore, when using double regularization constraints, the voxel folding problem can be effectively alleviated by adjusting the weight coefficients and ensuring smooth deformation.

[0050] MCANet consists of two identical registration subnetworks cascaded together, each incorporating an extended self-attention design. Since downsampling to enhance the receptive field can lead to the loss of local image details in deep feature maps, MCANet uses convolutional combinations with different dilation rates to explicitly expand the receptive field without adding extra parameters, allowing the network to capture more contextual information during convolution. A self-attention mechanism, primarily in the form of attention gating, is incorporated; shallow feature maps are progressively downsampled through convolutions to make their size match that of the gating feature maps. Figure 1 The attention gate, through a gating mechanism, combines the attention-adjusted features with the original features. This effectively expands the network's receptive field without adding a large number of parameters, while suppressing feature activity in aligned regions and enhancing the network's sensitivity to discrepancies, thus implicitly improving the deformation field smoothing problem. Subsequently, transposed convolution is used for upsampling, restoring the obtained feature map to the original size of the input image. For the two-sub-network cascaded structure, double regularization constraints can explicitly solve the voxel folding problem, comprehensively improving registration quality.

[0051] Example 2 In this invention, the Dice formula is used to calculate the volumetric overlap between identical anatomical tissues in fixed and deformed images, and the average Dice score of all labeled tissues is calculated to quantitatively assess registration accuracy. The calculation formula is shown below:

[0052] (10); in and These represent the corresponding brain anatomical tissues in the fixed and deformed images, respectively. Simultaneously, to quantify the regularization of the DVF (displacement vector field) during registration, this invention also calculates the number of folded voxels in the DVF, which is obtained by calculating the number of voxels with non-positive Jacobian determinants in the DVF (i.e.,...). ).

[0053] To visually demonstrate the performance of MCANet, this invention introduces several other advanced registration methods as baselines, including SyN, LDDMM, VoxelMorph-1, VoxelMorph-2, and FAIM. SyN, as an excellent representative of traditional registration algorithms, has been proven to effectively perform brain registration between patients. This invention uses the ANTS software package for SyN experiments, employing cross-correlation as the similarity metric and a gradient step size of 0.2. The second traditional algorithm baseline is LDDMM, designed to address the problem of large-deformation image registration and capable of handling registration between patients with different brain structures well. In the experiments of this invention, mean squared error (MSE) is used as the objective function, with a smoothing kernel size of 5, a smoothing kernel power of 2, a matching term coefficient of 4, a regularization term coefficient of 8, and 500 iterations. Furthermore, this invention also selects VoxelMorph, based on unsupervised learning, as a baseline. This model has two variants, VoxelMorph-1 and VoxelMorph-2. In this invention, the model is retrained based on the optimal parameters given in the literature, and registration evaluation is performed on the test dataset of this invention. FAIM is a lightweight unsupervised registration model that can achieve high-precision registration with fewer network parameters. Furthermore, FAIM incorporates a regularization term that directly penalizes folded voxels during training. According to the literature, the weight coefficient of the regularization term is set to 10⁻⁵ in this invention to ensure a balance between registration accuracy and deformation field smoothness.

[0054] The MCANet proposed in this invention is implemented using the deep learning framework PyTorch. All experiments were conducted on a single NVIDIA Tesla P100 GPU and Intel Xeon Silver 4210 CPU. All models were trained for 500 epochs (63,000 iterations) using the Adam gradient optimizer to minimize training loss and update network parameters, with a batch size of 1. During training, the initial learning rate was set to 1 × 10⁻⁴ and automatically decreased with increasing training iterations. Based on experimental experience, the weight coefficients... It is set to 1.

[0055] To evaluate the registration performance of MCANet and other advanced registration methods on the MindBoggle101 dataset, this invention conducted a series of comparative experiments and calculated their average Dice scores and the degree of deformation field folding (i.e., the number and percentage of folded voxels) on the test dataset. The registration performance metrics of these methods are listed in Table 1. The data shows that MCANet achieved the highest registration accuracy among all methods. Compared to traditional methods, MCANet improved the Dice score by 0.252 compared to the Affine method, 0.129 compared to SyN, and 0.145 compared to LDDMM. MCANet significantly outperformed SyN and LDDMM in terms of registration speed. Compared to deep learning-based methods, MCANet improved the Dice score by 0.018 compared to FAIM, 0.056 compared to VoxelMorph-1, and 0.034 compared to VoxelMorph-2.

[0056] While traditional algorithms show advantages in the smoothness of deformed fields, this requires careful parameter selection. Experiments show that ANTSSyN's registration accuracy gradually improves with increasing gradient step size, but the folding rate of the deformed field also increases linearly. This means that with increasing parameters, traditional algorithms may produce overly complex and difficult-to-interpret deformation results. Compared to traditional algorithms, MCANet demonstrates a clear advantage in time efficiency due to its faster registration speed, and exhibits a significant efficiency advantage among these methods. To more intuitively illustrate the registration effects of these methods, Figure 4 The visualization of the registration results on the same cross-sectional slices of the brain is shown. Clearly, MCANet produces the most accurate and natural registration results.

[0057] Table 1. Registration results of various methods In summary, MCANet demonstrates excellent registration performance on the Mindboggle101 dataset. Compared to traditional algorithms and other deep learning-based methods, MCANet has significant advantages in registration accuracy and time efficiency, and is able to generate more accurate and natural deformation results.

[0058] Example 3 To verify the impact of each network component on registration performance, this invention conducted effectiveness experiments on each structure of the registration sub-network DSNet and the cascaded structure of the two sub-networks. This experiment tested the registration capability of DSNet and performed an ablation study on DSNet. Table 2 lists three variants of DSNet, where "DC" represents a combination of dilated convolutions with different dilation rates, and "AG" represents an attention-gated component. In "DSNet (without DC)," the combination of dilated convolutions is removed, and ordinary 3D convolutions are used instead. This change also reflects the impact of a large receptive field on registration performance. In "DSNet (without AG)," four attention-gated components at different scales are replaced by skip connections. The "DSNet (without DC+AG)" variant removes the combination of dilated convolutions and the attention mechanism, retaining only the basic network structure. This variant aims to verify the compatibility of the combination of dilated convolutions and the attention-gated component. The experimental data in Table 2 show that the design added in the encoder-decoder architecture of this invention is fully effective.

[0059] Table 2 Registration results of the MCANet ablation study Example 4 This invention imposes dual regularization constraints on the model during training to avoid excessive deformation and voxel folding problems that may be caused by two-stage registration. In this section, the invention analyzes the impact of the additional bending energy loss on registration performance. Table 3 records the changes in bending energy weighting coefficients. The registration accuracy and percentage of folded voxels of MCANet were calculated for different values, and the data were presented as a line graph. Figure 5 It is displayed in the middle.

[0060] Table 3 Regularization Coefficients Impact on registration performance To visually demonstrate the constraint effect of the bending energy regularization term on the deformation field, Figure 6 In different The deformation field was visualized at a certain value. By observing the visualization results of the deformation field, it can be found that as... As the value increases, the number of folded mesh lines (white areas) in the mesh deformation field decreases significantly, and the distortion of the mesh lines becomes more natural and smooth. Observation Figure 5 As can be seen from the trend of the red line in this invention, when the weight of bending energy is slightly increased, the registration accuracy does not decrease significantly (except for...). The Dice score between the deformed and fixed images can still be maintained at a high level. Registration accuracy is only significantly affected when the number of folded voxels is significantly reduced. Therefore, A value of 1.4 is an ideal and compromise choice, which can ensure the smoothness and realism of deformation while maintaining high registration accuracy.

[0061] In the comparative experiments, MCANet significantly improved the Dice score compared to traditional registration methods Affine and SyN, as well as existing deep learning-based registration methods FAIM and VoxelMorph. Compared to traditional algorithms, MCANet demonstrates an absolute advantage in time efficiency due to its faster registration speed. MCANet's clear advantage in time efficiency confirms the feasibility of the algorithm proposed in this invention.

[0062] In the ablation experiments, this invention demonstrates the impact of each module in the network. Adding extended convolutional groups improves registration accuracy by increasing the receptive field, while adding attention gating components enhances the model's response to key regions, further improving registration accuracy. Using both methods simultaneously expands the receptive field without significantly increasing computational cost or network parameters, enabling the acquisition of multi-scale features, mitigating voxel folding issues, and addressing over-deformation. The network can better understand the global structure and local details of images, improving its ability to accurately detect deformed regions.

[0063] In the regularization analysis section, this invention applies double regularization constraints to the model during training and analyzes the bending energy weight coefficients. The impact on registration accuracy and smoothness, when the bending energy weighting coefficient is used. As the number of voxels increases, the degree of voxel folding decreases, and the Dice score also decreases to some extent. Through numerous experiments, this invention selects weighting coefficients... The value is 1.4, which not only ensures the smoothness and realism of the deformation, but also maintains a high registration accuracy.

[0064] 3D brain MRI registration is crucial for observing structural brain injuries and for brain health treatment. Traditional deformable registration algorithms require significant time to register image pairs and necessitate manual adjustment of multiple parameters to achieve optimal results. Furthermore, the optimal registration parameters differ between different images, making traditional methods highly dependent on manual operation, which is both cumbersome and time-consuming. With improvements in hardware performance and the increasing size of network models, traditional methods have become increasingly unsatisfactory in terms of registration time and accuracy. In contrast, the model proposed in this invention can predict the optimal parameters for deformability through continuous optimization learning and can be directly applied to unseen images, completing registration at near real-time speeds without human intervention. The network can automatically predict deformability based on the features of the input image, which is both efficient and accurate. While some traditional algorithms still have advantages in deformability smoothness, FAIM and MCANet have also made significant progress in this area. When the weights of the regularization term are sufficiently large, they can achieve very smooth deformability without unnecessary sacrifices.

[0065] As a classic unsupervised registration model, VoxelMorph has long been a key reference for researchers conducting comparative experiments due to its excellent registration accuracy, speed, and robustness. In terms of registration accuracy, VoxelMorph-1 and VoxelMorph-2 still lag behind DSNet, while MCANet has a clear advantage. Furthermore, VoxelMorph lacks attention to deformation smoothness during training and is deficient in voxel folding. Without additional regularization constraints, DSNet's folding rate is also significantly lower than other learning-based methods, which is significantly related to the addition of attention gating. In the experiments of this invention, the voxel folding problem is improved by adding an attention gating module, and the attention gating module also implicitly improves the over-deformation problem because it pays close attention to regions of difference. When a model possesses a large receptive field and the ability to concentrate deformation, the deformation of its motion images becomes more natural and accurate, eliminating the need for excessive deformation to improve registration accuracy. This can be seen from the ablation experiment data in Table 2, where using dilated convolution combined with attention gating reduces the proportion of folded voxels by approximately 0.3% and improves the smoothness of the deformation field by approximately 38%. Furthermore, adding bending energy regularization can further improve the voxel folding problem, but at the cost of sacrificing a small amount of registration accuracy, as shown in Table 3. FAIM, a lightweight unsupervised network model that directly uses the values ​​of non-orthogonal comparable determinants as part of the loss function, has also made some progress. Although MCANet is larger than FAIM in terms of network architecture and number of parameters, it significantly outperforms the latter in registration accuracy, which remains one of the most important metrics for evaluating models and algorithms. In summary, the model proposed in this invention achieves progress in registration accuracy, speed, and robustness, and its advantages are verified through ablation experiments and comparative analysis.

[0066] This invention proposes a multi-constraint cascaded attention network (MCANet) based on unsupervised learning for deformable registration of brain MRI images. The sub-network design incorporates dilated convolutional combinations and attention mechanisms, which effectively improves several factors affecting constrained registration performance. MCANet was tested on the Mindboggle101 dataset, and its registration performance on 3D brain MRI was quantitatively evaluated by calculating the Dice score, the number of folded voxels, and registration time. Experimental results show that MCANet achieves the highest registration accuracy compared to several existing state-of-the-art registration methods, including traditional algorithms and learning-based methods, and can smoothly deform motion images. This demonstrates MCANet's excellent performance in aligning two brain MR images, representing substantial progress in this research. Future research will continue to improve the smoothness of deformation and collect more 3D images of other organs for registration to enhance applicability.

[0067] Image registration in brain medicine is a fundamental prerequisite for computer-aided treatment of brain diseases. The brain is one of the most important and complex organs in the human body, making accurate and rapid registration extremely challenging. To address the voxel folding problem in deformation fields and the low registration accuracy when dealing with complex and detailed targets, this invention proposes a fully convolutional multi-constraint cascaded attention network (MCANet). This network consists of two cascaded registration sub-networks that iteratively register input image pairs from coarse to fine. The registration sub-network, called the Dilated Self-attention Network (DSNet), introduces a combination of dilated convolutions with different dilation rates and an attention gating module. During the training of MCANet, dual regularization constraints are applied to specifically penalize excessive deformation, enabling the network to generate relatively smooth deformations while achieving high registration accuracy. Experimental results on the Mindboggle101 dataset show that MCANet's registration accuracy significantly outperforms several existing state-of-the-art registration methods, and the network can achieve relatively smooth registration.

[0068] The combination of dilated convolutions and attention mechanisms has been successfully applied to 3D brain MRI image registration. Dilated convolutions can expand the receptive field of the network without increasing computational cost and network parameters, thereby acquiring multi-scale features and enabling the network to better understand the global structure and local details of the image, improving its ability to accurately detect deformed regions.

[0069] By employing a cascaded network architecture, deformation registration of the input image is achieved from coarse to fine, significantly improving registration accuracy. This method of progressively refining the registration results effectively enhances the accuracy of the registration and can handle registration tasks of varying scales and complexities. Simultaneously, dual regularization constraints are used to ensure the smoothness and realism of image registration, avoiding excessive deformation and discontinuities in the registration results, making the results more consistent with reality.

[0070] The proposed network model employs an unsupervised end-to-end training method, which is not limited by the true values ​​and can achieve near real-time registration of brain MRI between patients.

[0071] Compared with traditional algorithms and existing deep learning algorithms, the algorithm of this invention can perform alignment efficiently and accurately under double regularization constraints, promote alignment smoothness, suppress severe bending deformation, and comprehensively improve performance.

[0072] Based on the same concept, the present invention also provides a deformable registration system for brain medical images, including an acquisition module, a first deformation module, a first registration module, a second deformation module, and a second registration module.

[0073] The acquisition module is used to acquire baseline brain medical images and brain medical images to be registered from different patients. The baseline brain medical images are used as fixed images, and the brain medical images to be registered are used as moving images.

[0074] The first deformation module is used to stitch the fixed image and the moving image in the channel dimension to obtain the first tensor; the first tensor is subjected to multiple stages of encoding and decoding processing to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the first deformation field; wherein, in the decoding process of each stage, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features; the weighted features are fused with the decoded features of the previous stage to obtain the decoded features of the current stage.

[0075] The first registration module is used to perform interpolation processing on the moving image based on the deformation information of the first deformation field to obtain a deformed image.

[0076] The second deformation module is used to stitch the fixed image and the deformed image together in the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field.

[0077] The second registration module is used to interpolate the deformed image based on the deformation information of the second deformation field to obtain a registered image.

[0078] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0079] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deformable registration method for brain medical images, characterized in that, The following steps are involved: Baseline brain medical images and brain medical images to be registered were acquired from different patients. The baseline brain medical images were used as fixed images, and the brain medical images to be registered were used as moving images. The fixed and moving images are stitched together along the channel dimension to obtain the first tensor. The first tensor undergoes multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features. The decoded features output in the final stage are convolved to obtain the first deformation field. In each stage of decoding, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features. The weighted features are then fused with the decoded features of the previous stage to obtain the decoded features of the current stage. The moving image is interpolated based on the deformation information of the first deformation field to obtain a deformed image; The fixed image and the deformed image are concatenated along the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field. The deformation image is interpolated based on the deformation information of the second deformation field to obtain the registration image.

2. The deformable registration method for brain medical images as described in claim 1, characterized in that, The first deformation field is obtained through a first registration sub-network, which includes an encoder and a decoder; The encoder includes a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer. The first coding layer includes a first convolutional layer and a first stride convolution. The second coding layer includes a first dilated convolutional module and a second stride convolution. The third coding layer includes a second dilated convolutional module and a third stride convolution. The fourth coding layer includes a second convolutional layer and a fourth stride convolution. The first and second dilated convolutional modules are both dilated convolutions with different dilation rates, concatenated together. The decoder includes decoding layers at different stages, multiple attention weighting units, and multiple feature fusion units. The first, second, third, and fourth coding layers perform multiple stages of coding processing on the first tensor to obtain multiple coded features; the decoding layers at different stages are used to output multiple decoded features. During the decoding process at each stage, the attention weighting unit is used to weight the encoded features of the current stage with the decoded features of the previous stage to obtain weighted features; The feature fusion unit is used to fuse the weighted features with the decoding features from the previous stage to obtain the decoding features for the current stage.

3. The deformable registration method for brain medical images as described in claim 2, characterized in that, The first convolutional layer is used to perform 3×3×3 convolutional feature extraction on the first tensor to obtain a shallow feature T1 with 16 channels; the first long convolution is used to downsample the shallow feature T1 to obtain a feature T1 with a resolution of ½ and still 16 channels. The first dilated convolution module applies a 3×3×3 convolution with a dilation rate d=1 / 2 / 3 to the downsampled feature T1 to obtain a 32-channel feature T2; the second stride convolution downsamples the feature T2 to obtain a feature T2 with a resolution of ¼ and 32 channels. The second dilated convolution module is used to apply a 3×3×3 dilated convolution to the downsampled feature T2 to obtain a 64-channel feature T3; the third stride convolution is used to downsample the feature T3 to obtain a feature T3 with a resolution of ⅛ and 64 channels. The second convolutional layer performs a 3×3×3 convolution on the downsampled feature T3 to obtain feature T4; the fourth stride convolution is used to downsample feature T4 to obtain a resolution of ¹ / ². 16 The feature T4 with 64 channels is used as the input to the decoder.

4. The deformable registration method for brain medical images as described in claim 2, characterized in that, The second deformation field is obtained through the second registration sub-network, and the first and second registration sub-networks have the same structure.

5. The deformable registration method for brain medical images as described in claim 4, characterized in that, Two spatial transformation networks are used to interpolate the moving and deformed images respectively. The first registration subnetwork, the second registration subnetwork, and two spatial transformation networks are interconnected to obtain a deformable image registration model. The deformable image registration model is trained using similarity loss and regularization constraints. The specific similarity loss is as follows: ； In the formula, L sim For similarity loss, NCC For normalized cross-correlation, F M is a fixed image, and M is a moving image. p It represents the position of a specific pixel within the entire image. p i For a certain region in the image i 1 pixel For the image domain, and For fixed images and deformed images p The average voxel intensity within the local window centered on the value. It is a compound symbol. For deformation field; The specific regularization constraints are as follows: ； in, ； ； In the formula, R For regularization constraints, and The tradeoff coefficient between the two regularization terms. For diffusion regularization, For bending energy regularization, For gradient operators, x , y , z These are the coordinate axes in three-dimensional space. For the Laplace operator.

6. The deformable registration method for brain medical images as described in claim 5, characterized in that, The deformation image is obtained as follows: ； In the formula, I 1 is a deformed image. U 1 represents the first registration subnetwork. This is the first deformation field; The acquisition of the registration image is as follows: ； In the formula, I2 is the registered image. U 2 represents the first registration subnetwork. For the first deformation field, This is the second deformation field.

7. The deformable registration method for brain medical images as described in claim 1, characterized in that, Before stitching the fixed and moving images in the channel dimension, the fixed and moving images undergo preprocessing such as skull stripping, normalization, cropping, and affine alignment.

8. A deformable registration system for brain medical images, characterized in that, include: The acquisition module is used to acquire baseline brain medical images and brain medical images to be registered from different patients. The baseline brain medical images are used as fixed images, and the brain medical images to be registered are used as moving images. The first deformation module is used to stitch the fixed image and the moving image in the channel dimension to obtain the first tensor; the first tensor is subjected to multiple stages of encoding and decoding processing to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the first deformation field; wherein, in the decoding process of each stage, the encoded features of the current stage are weighted with the decoded features of the previous stage to obtain weighted features; the weighted features are fused with the decoded features of the previous stage to obtain the decoded features of the current stage; The first registration module is used to perform interpolation processing on the moving image based on the deformation information of the first deformation field to obtain a deformed image; The second deformation module is used to stitch the fixed image and the deformed image in the channel dimension to obtain the second tensor; the second tensor is subjected to multiple stages of encoding and decoding to obtain multiple encoded features and corresponding decoded features; the decoded features output in the final stage are convolved to obtain the second deformation field. The second registration module is used to interpolate the deformed image based on the deformation information of the second deformation field to obtain a registered image.

Images