Dual-stream attention registration method and system based on shared weights of large kernel convolution LKA

Abstract

The present invention discloses a dual-stream attention registration method and system based on shared weights of large kernel convolution (LKA), specifically relating to the field of medical image registration technology. The input fixed and moving brain images are respectively fed into the dual-stream structure of DELCA-Net. DELCA-Net combines the advantages of convolutional neural networks (CNN) and Transformer, and deeply mines the semantic information of the image through a self-attention mechanism. With the help of a shared weight encoder design, the network can learn more general and rich feature representations. On this basis, using the cross-self-attention mechanism, DELCA-Net achieves precise feature matching of these shared features, significantly improving the accuracy of registration. To reduce computational complexity, DELCA-Net decomposes the large convolution kernel into cascaded deep convolution and deep dilated convolution modules. In addition, through a multi-scale attention optimization mechanism, DELCA-Net effectively integrates spatial correspondences and anatomical semantic associations at different scales.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image registration, and in particular to a dual-stream attention registration method and system based on shared weights of large kernel convolution (LKA). Background Art

[0002] With the development of deep learning technology, deformable registration methods based on convolutional neural networks (CNNs) have shown significant advantages. These methods effectively circumvent the iterative computational bottleneck of traditional methods by learning the mapping relationship between image pairs and deformation fields in an end-to-end manner. The CNN architecture, with its local receptive field and hierarchical feature extraction mechanism, can efficiently model the structural features of medical images. However, due to the local characteristics of conventional convolutional kernels, these methods have obvious shortcomings in capturing the global spatial correlation of large-scale anatomical structures, especially for organ registration tasks with significant deformation or anatomical heterogeneity.

[0003] In recent years, the Transformer architecture, with its global self-attention mechanism, has provided a new technology for medical image registration, leveraging its ability to model long-range dependencies. Unlike CNNs, which rely on local receptive fields for feature extraction, the Transformer, through sequential processing and weighted self-attention, can theoretically model contextual relationships across the entire image domain. Representative works, such as TransMorph, have achieved superior performance in brain MRI registration by constructing a Transformer-based U-Net architecture. However, simple attention mechanisms lack the constraints required to model the inherent spatial locality of medical images and may focus on anatomically irrelevant long-range voxel associations. This is particularly true in registration scenarios dominated by small displacements and deformations. This global, unbiased attention distribution is prone to noise and incurs significant computational overhead. Therefore, a new paradigm for deformable medical image registration based on large kernel attention (LKA) is urgently needed to synergize the complementary strengths of existing technologies. Summary of the Invention

[0004] To this end, the present invention provides a dual-stream attention registration method and system based on shared weights of large kernel convolution LKA to solve the problems raised in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a two-stream attention registration method based on shared weights of large kernel convolution LKA, in which fixed and moving brain images are input into the two-stream structure of DELCA-Net respectively, and features are extracted through multiple modules composed of large kernel convolution LKA blocks in the encoding path. The shared weights of the two-stream structure help maintain the consistency and comparability of medical image feature extraction;

[0006] In each large kernel convolution LKA Block, downsampling is first performed through the Down operation to reduce the image resolution to expand the receptive field, allowing the DELCA-Net network to perceive the spatial distribution and large-scale displacement of brain structures. Subsequently, the local structural features of the image, including gyri, sulci, and cortical boundaries, are extracted through depthwise separable convolution (DW-Conv) and depthwise separable dilated convolution (DW-D-Conv). These structures have high anatomical significance in brain MRI. At the same time, this design significantly reduces the number of model parameters and computational complexity, improving registration efficiency.

[0007] Next, 1×1 convolution is used to adjust the number of channels, and the PReLU activation function is introduced to enhance the nonlinear representation ability, further improving the adaptability of typical medical image problems, including blurred boundaries and intensity changes. Among them, the additional depth-separable convolution DC-Conv is used to enhance the perception of detailed areas of the ventricles and hippocampus. Multiple LKA Blocks are used in series to effectively achieve layer-by-layer abstraction from shallow low-level features of edges and textures to deep semantic features of structural shapes and functional area distributions. The features extracted from the i-th layer of each LKA Block are recorded as and , respectively represent the weighted feature representations of the moving image and the fixed image;

[0008] After feature extraction, and It is sent to CARM Block for feature refinement. LKCA is first introduced in CARM to and Perform cross attention modeling to get the cross attention of layer i , which can enhance the saliency of corresponding structures between images and improve the accuracy of feature matching in the case of non-rigid changes such as tumor and brain tissue offset or loss.

[0009] Then, in the Correct Layer, attention is paid to the current layer. Perform upsampling operation UP to match the upper attention Keeping the resolution consistent, the subsequent LKA Block and Conv 1×1 further enhance the guiding effect of the current layer global semantic information on the upper layer attention map, thereby correcting the spatial offset caused by downsampling and obtaining the corrected upper layer attention map. , making the attention map more accurately focused on key anatomical structures, including the corpus callosum boundary or lesion edge.

[0010] In the Fusion Layer, the upper attention map Downsample and focus with the current layer After alignment and channel splicing, convolution and activation operations are introduced to fuse spatial details and structural semantic information at different scales to obtain the refined current layer attention map. , which combines high-level semantic structures (such as functional areas) with low-level texture information (such as gray-white matter boundaries) to form a refined attention map with fine-grained structural information, providing accurate spatial guidance for subsequent registration;

[0011] In the Decoder Block, 、 , refined attention feature map , and the decoder output of layer i+1 , together constitute the input of Decoder Block; among them, and It provides the structure and texture information of the original image, refines the attention map to enhance the sensitivity of feature alignment, and the output of the i+1 layer decoder contains contextual information at a higher semantic level. After splicing the four in the channel dimension, the input convolution layer is fused with the ReLU activation function to effectively achieve the unity of structure preservation and alignment guidance, and finally output the fused feature map. , used to estimate the deformation field and complete image registration.

[0012] Preferably, in the LKA Block, the depthwise separable dilated convolution DW-D-Conv obtains a large spatial receptive field by integrating the information of the depthwise separable convolution DW-Conv, and finally completes the cross-channel feature integration through 1×1 convolution;

[0013] Decompose a large kernel convolution of size K*K into a convolution with a dilation rate of d Depthwise separable dilated convolution DW-D-Conv, depthwise separable convolution DW-Conv with a kernel size of 2d-1, and a 1×1 convolution;

[0014] ;

[0015] ;

[0016] represents the input feature map, represents 1×1 convolution, represents depth-wise separable dilated convolution, Represents depth-separable convolution, Attention represents the attention after convolution extraction, represents element-wise multiplication, represents the output feature map;

[0017] After obtaining the long-term dependency information, the importance of each point is estimated and the corresponding attention map is generated. The number of parameters of the decomposed large kernel convolution is and the amount of computation The calculation is as follows:

[0018] ;

[0019] ;

[0020] Where C represents the number of input channels; H and W represent the size of the input feature map, which are height and width respectively; d represents the expansion rate, and K represents the convolution kernel size.

[0021] Preferably, in CARM Block, enter the mobile feature With fixed features , respectively, through the parameter-independent 1×1×1 convolutional layer to integrate cross-channel information, and through the adaptive gating mechanism of the GELU activation function to improve feature discrimination to obtain feature enhancement. and , fixed features Further expand the receptive field through DC Conv to capture a wide range of context information and obtain features containing global context information , and then the moving feature is realized by Hadamard product Features containing global context information Feature matching to obtain the cross attention of the current layer moving image and fixed image , as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] in Represents the Hadamard product. The cross-scale attention optimization strategy downsamples the attention information of the high-resolution layer through the Fusion Layer and passes it to the current layer. Its purpose is to integrate high-level fine spatial detail information into the feature representation of the current layer, thereby significantly enhancing the model's ability to capture and understand the corresponding relationship between voxel positions. At the same time, the attention map of the current layer is upsampled through the Correct Layer, using the global information of the current layer's attention to correct the deviation in the attention map of the upper layer. The specific implementation of the Fusion Layer and Connect Layer is as follows:

[0027] ;

[0028] ;

[0029] represents two independent convolutional activation layers, represents the attention map of the previous layer, 、 represents the attention map after refinement and adjustment, Cat represents the channel splicing operation, Down represents downsampling, and UP represents upsampling;

[0030] The cross-scale attention optimization strategy leverages the fine spatial details of high-resolution layers through a combination of fusion and correction mechanisms, ensuring the consistency and accuracy of attention maps across scales. This mechanism continuously fine-tunes and optimizes the distribution of attention weights, enabling the model to more accurately focus on important feature areas.

[0031] The present invention also discloses a dual-stream attention registration system with shared weights based on large kernel convolution LKA, which realizes image registration based on the above method.

[0032] This paper proposes a two-stream shared-weight registration network (DELCA-Net) based on large kernel attention (LKA). DELCA-Net combines the advantages of convolutional neural networks (CNNs) and transformers, using a self-attention mechanism to deeply mine image semantic information. Leveraging a shared-weight encoder design, the network learns a more general and rich feature representation. Furthermore, utilizing a cross-self-attention mechanism, DELCA-Net achieves precise feature matching of these shared features, significantly improving registration accuracy. To reduce computational complexity, DELCA-Net decomposes the large convolution kernel into a cascade of depthwise convolution and depthwise atrous convolution modules. Furthermore, through a multi-scale attention optimization mechanism, DELCA-Net effectively incorporates spatial correspondences and anatomical semantic associations at different scales. Experimental results on two publicly available 3D brain magnetic resonance (MR) datasets demonstrate that DELCA-Net achieves state-of-the-art registration results while also demonstrating good performance in terms of model parameters and computational load. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 An overview of the DELCA-Net provided by the present invention;

[0034] Figure 2Schematic diagram of the structure of LKA Block and LKCA Block provided by the present invention;

[0035] Figure 3 A comparison chart of visual results obtained using different registration methods on the IXI dataset provided by the present invention;

[0036] Figure 4 A comparison chart of the Dice score statistical distribution based on the box plot provided by the present invention, the DELKA-NET framework, the traditional method, and the CNN-based image registration method in the registration accuracy of multiple subregions of the brain;

[0037] Figure 5 A comparison chart of the Dice score statistical distribution based on the box plot provided by the present invention, the DELKA-NET framework, the traditional method, and the Transformer-based image registration method in the registration accuracy of multiple subregions of the brain;

[0038] Figure 6 This is a visualization of the test results of dice scores for various channel settings on the OASIS-2D dataset provided by the present invention;

[0039] Figure 7 A graph showing the relationship between dice scores, convolution kernel size, and number of channels on the OASIS-2D dataset provided by the present invention. DETAILED DESCRIPTION

[0040] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0041] In recent years, deep learning registration methods have demonstrated advantages in efficiency and performance over traditional methods. Feature extraction and matching between input images are key challenges in deep learning registration. Traditional deep learning registration networks typically use a channel-by-channel splicing approach to input image pairs to be registered in order to extract features and capture the relationships between images. However, this approach lacks in-depth modeling of individual image features, and the feature matching process between images is less intuitive.

[0042] like Figure 1As shown in the figure, the present invention proposes a dual-stream attention registration method based on shared weights of large kernel convolution LKA. The fixed image and moving image of the input brain are respectively fed into the dual-stream structure of DELCA-Net. In the encoding path, features are extracted through multiple modules composed of large kernel convolution LKA Blocks (LKA modules).

[0043] In each large kernel convolution LKA Block, downsampling is first performed through the Down operation to enable the DELCA-Net network to perceive the spatial distribution and large-scale displacement of the brain structure; then, the local structural features of the image, including gyri, sulci and cortical boundaries, are extracted through depthwise separable convolution DW-Conv and depthwise separable dilated convolution DW-D-Conv. The number of channels is adjusted using 1×1 convolution, and the PReLU activation function is introduced to enhance the nonlinear representation ability, further improving the adaptability of typical medical image problems, including boundary blur and intensity changes; among them, the additional depthwise separable convolution DC-Conv is used to enhance the perception of detailed areas of the ventricles and hippocampus; multiple LKA Blocks are used in series to effectively realize the layer-by-layer abstraction from shallow low-level features of edges and textures to deep semantic features of structural shapes and functional area distributions; the features extracted in the i-th layer of each LKA Block are recorded as and , respectively represent the weighted feature representations of the moving image and the fixed image;

[0044] After feature extraction, and It is sent to CARM Block for feature refinement. LKCA is first introduced in CARM to and Perform cross attention modeling to get the cross attention of layer i , so as to enhance the saliency of corresponding structures between images in the case of non-rigid changes in tumors, brain tissue offset or missing.

[0045] Then, in the Correct Layer, attention is paid to the current layer. Perform upsampling operation UP to match the upper attention Keeping the resolution consistent, the subsequent LKA Block and Conv 1×1 further enhance the guiding effect of the current layer global semantic information on the upper layer attention map, thereby correcting the spatial offset caused by downsampling and obtaining the corrected upper layer attention map. , making the attention map more accurately focused on key anatomical structures, including the corpus callosum boundary or lesion edge.

[0046] In the Fusion Layer, the upper attention map Downsample and focus with the current layer After alignment and channel splicing, convolution and activation operations are introduced to fuse spatial details and structural semantic information at different scales to obtain the refined current layer attention map. , which combines high-level semantic structure with low-level texture information to form a refined attention map with fine-grained structural information, providing accurate spatial guidance for subsequent registration;

[0047] In the Decoder Block, 、 , refined attention feature map , and the decoder output of layer i+1 , together constitute the input of the Decoder Block (decoder); among them, and It provides the structure and texture information of the original image, refines the attention map to enhance the sensitivity of feature alignment, and the output of the i+1 layer decoder contains contextual information at a higher semantic level. After splicing the four in the channel dimension, the input convolution layer is fused with the ReLU activation function to effectively achieve the unity of structure preservation and alignment guidance, and finally output the fused feature map. , used to estimate the deformation field and complete image registration.

[0048] In the LKA Block, the depthwise separable dilated convolution DW-D-Conv integrates the information of the depthwise separable convolution DW-Conv to obtain a large spatial receptive field, and finally completes the cross-channel feature integration through 1×1 convolution;

[0049] Decompose a large kernel convolution of size K*K into a convolution with a dilation rate of d Depthwise separable dilated convolution DW-D-Conv, depthwise separable convolution DW-Conv with a kernel size of 2d-1, and a 1×1 convolution;

[0050] ;

[0051] ;

[0052] represents the input feature map, represents 1×1 convolution, represents depth-wise separable dilated convolution, Represents depth-separable convolution, Attention represents the attention after convolution extraction, represents element-wise multiplication, represents the output feature map;

[0053] After obtaining the long-term dependency information, the importance of each point is estimated and the corresponding attention map is generated. The number of parameters of the decomposed large kernel convolution is and the amount of computation The calculation is as follows:

[0054] ;

[0055] ;

[0056] Where C represents the number of input channels; H and W represent the size of the input feature map, which are height and width respectively; d represents the expansion rate, and K represents the convolution kernel size.

[0057] In CARM Block, enter the moving feature With fixed features , respectively, through the parameter-independent 1×1×1 convolutional layer to integrate cross-channel information, and through the adaptive gating mechanism of the GELU activation function to improve feature discrimination to obtain feature enhancement. and , fixed features Further expand the receptive field through DC Conv to capture a wide range of context information and obtain features containing global context information , and then the moving feature is realized by Hadamard product Features containing global context information Feature matching to obtain the cross attention of the current layer moving image and fixed image , as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] in Represents the Hadamard product. The cross-scale attention optimization strategy downsamples the attention information of the high-resolution layer through the Fusion Layer and passes it to the current layer. Its purpose is to integrate high-level fine spatial detail information into the feature representation of the current layer, thereby significantly enhancing the model's ability to capture and understand the corresponding relationship between voxel positions. At the same time, the attention map of the current layer is upsampled through the Correct Layer, using the global information of the current layer's attention to correct the deviation in the attention map of the upper layer. The specific implementation of the Fusion Layer and Connect Layer is as follows:

[0063] ;

[0064] ;

[0065] represents two independent convolutional activation layers, represents the attention map of the previous layer, 、 represents the attention map after refinement and adjustment, Cat represents the channel splicing operation, Down represents downsampling, and UP represents upsampling;

[0066] The cross-scale attention optimization strategy leverages the fine spatial details of high-resolution layers through a combination of fusion and correction mechanisms, ensuring the consistency and accuracy of attention maps across scales. This mechanism continuously fine-tunes and optimizes the distribution of attention weights, enabling the model to more accurately focus on important feature areas.

[0067] The training strategy of the present invention is based on the unsupervised registration paradigm, whose loss function is constrained by image similarity and deformation field regularization constraints Mathematically, it can be reduced to the following optimization problem:

[0068] ;

[0069] in Represents a fixed image, Represents a moving image, represents the deformation field, Represents a deformation operation, ( ) represents the image of the moving figure after being distorted by the deformation field, Represents a hyperparameter. By using the energy function in the traditional image registration algorithm, the accuracy of image alignment is ensured and the deformation field is predicted. Spatial regularization is performed to smooth the deformation field to reduce irregular deformation and noise. In this work, Local Normalized Cross-Correlation and Mean Squared Error are used as similarity metrics:

[0070] ;

[0071] in represents the image domain, Representatives A cubic neighborhood with a diameter of N as the center, 、 represents the average value of the voxels in the neighborhood. In this embodiment, N=9.

[0072] ;

[0073] calculate and The image similarity is evaluated by the average of the squared difference of the voxel values ​​between them, where p represents the voxel position. In this embodiment, diffusion regularization is used as a smooth loss function to ensure reasonable deformation, where Represents the spatial gradient of the displacement field:

[0074] ;

[0075] This embodiment carried out the following experiments according to the technical solution of the present invention:

[0076] 1. Dataset and Experimental Settings

[0077] This example validates the model's effectiveness on two public datasets, IXI and OASIS. For the IXI dataset, this example uses the preprocessed version provided by TransMorph (including skull stripping, affine alignment, and subcortical structure segmentation). The dataset contains 576 T1-weighted brain MRI images with an image size of 160×190×224. For the IXI dataset, this example performs atlas-to-patient registration. The dataset is divided into a training set (403 images), a validation set (58 images), and a test set (115 images) in a 7:1:2 ratio. Label maps containing 30 anatomical structures are used as the evaluation criteria. This example uses the OASIS dataset, which has been preprocessed using Hypermorph (including skull stripping, affine alignment, and normalization). The dataset contains 414 T1-weighted 3D MRI images with an image size of 160×192×224. For the OASIS dataset, this example performs inter-patient registration, dividing 375 images into a training set, 19 images into a validation set, and 19 image pairs into a test set. Evaluation is performed using a label map containing 35 anatomical structures. For the OASIS-3D dataset, this example additionally uses DiceLoss. To quickly test different methods and hyperparameters, this example performs ablation experiments on the OASIS 2D data, which contains 414 2D brain MRI images with a size of 160×192, all of which are slices from the corresponding 3D images. This example divides the dataset into 200 training sets, 14 validation sets, and 200 test sets.

[0078] The size of the model in this example is primarily determined by the hyperparameters C and K. In the experiment, C=4 and K=9 were set as the defaults. The network in this example was deployed in the Pytorch framework, using ADAM as the optimizer, a learning rate of 0.0001, and a batch size of 1.

[0079] like Figure 2 As shown in the figure, in the LKA Block module, we use the self-attention mechanism to capture semantic information, and then use LKCA to learn the position correspondence between the moving graphics and the fixed images. Moreover, since the moving graphics and the fixed images use the same feature representation, the cross-attention module achieves more efficient information fusion and processing between the moving images and the fixed images.

[0080] Specifically, image A passes through a convolutional layer with a kernel size of K=1, a GELU activation function, and a decomposed large-kernel convolution operation. The self-attention mechanism is used to weight the central pixel according to the neighboring pixels, thereby effectively capturing local semantic information. Subsequently, in the LKCA Block, image A and image B, which have been weighted by the LKA Block, are combined, the receptive field is increased through large-kernel convolution, and Hadamard multiplication is used for weighting. Feature search and matching is performed for each pixel in image A within the wide-area neighborhood of the corresponding position in image B, thereby achieving high-precision feature alignment and fusion across images. And because images A and B use the same feature representation, the LKCA Block achieves more efficient information fusion and processing between moving and fixed images.

[0081] Figure 3 The first row shows the fixed image, the moving image, and the distorted moving image obtained by various methods. The second row shows the image fused with the segmentation labels, including the fixed image, the moving image, and the distorted moving image obtained by different methods. The third row shows the deformation fields predicted by various methods.

[0082] 2. Atlas-to-patient brain MR registration:

[0083] This example conducts a comprehensive study of atlas-to-subject registration on the IXI dataset. Table 1 shows the quantitative evaluation results of various methods, covering key metrics such as registration accuracy (Dice coefficient), deformation field regularity (percentage of negative Jacobians), model complexity (number of parameters), and computational efficiency (number of operations). To ensure experimental fairness, all compared methods (including SyN, VoxelMorph, TransMorph, LKU-Net, etc.) use the same preprocessing process and data partitioning strategy.

[0084] Table 1. Experimental results compared with other methods on the IXI dataset

[0085]

[0086] Experimental results demonstrate that the DELCA-Net architecture proposed in this example demonstrates significant advantages. By constructing network structures of different sizes (DELCA-Net (K=9, C=4) and DELCA-Net-Large (K=9, C=8)), DELCA-Net and DELCA-Net-Large achieved average Dice coefficients of 0.766±0.129 and 0.768±0.122, respectively, in terms of registration accuracy. This significantly outperforms the traditional SyN method (0.645±0.152) and mainstream deep learning models VoxelMorph (0.732±0.123) and TransMorph (0.754±0.124). DELCA-Net-Large achieved the highest Dice coefficient, outperforming all compared methods in the table.

[0087] In terms of deformation field regularization, DELCA-Net demonstrates excellent topology preservation, with a negative Jacobian fraction of only 0.094±0.051%. This property is crucial for ensuring appropriate deformation of anatomical structures, particularly in the registration of fine structures such as the ventricles and hippocampus, effectively reducing the risk of deformation field collapse and topological disruption. From the perspective of model efficiency, DELCA-Net achieves optimal performance with only 0.91M parameters, a 95.1% and 56.3% reduction compared to TransMorph (46.77M) and LKU-Net (2.08M), respectively. While some convolutional neural network-based models have smaller parameters, their Dice coefficients decrease significantly, DELCA-Net successfully achieves an optimal balance between accuracy and efficiency.

[0088] In terms of computational complexity, DELCA-Net utilizes mechanisms such as self-attention and cross-attention during inference, which involve key-value dot product operations. This improves registration accuracy but also incurs a computational overhead of 275.63G FLOPs. Therefore, further reducing computational complexity and continuously improving accuracy while maintaining high accuracy and elegant deformation regularization remains a promising area for future research.

[0089] Figure 4 and Figure 5The Dice score distribution of various image registration methods for various brain structures is shown. The method proposed in this example achieves higher Dice scores than other comparison methods for the vast majority of anatomical structures. Furthermore, the model of this example demonstrates excellent performance on key and complex brain structures such as the hippocampus, third ventricle, and amygdala, demonstrating its ability to accurately identify and precisely register fine brain structures.

[0090] 3. Brain MR registration between patients:

[0091] In order to compare the performance of different methods in the inter-patient image registration task, this embodiment selected the OASIS dataset for systematic experimental verification. As shown in Table 2, among the learning-based methods, the DELCA-NET model proposed in this embodiment achieved the highest Dice coefficient of 0.891, and also ranked second in the Hausdorff distance (HD-95) indicator, reaching 1.334. In the dataset with labeled learning, the model of this embodiment achieved an even more obvious lead. This advantage may be attributed to the fact that after incorporating segmentation label information into the model, the self-attention mechanism can more effectively capture and identify common features between images, while the cross-attention mechanism further promotes the precise matching of key points. In terms of SDLOGJ, DELCA-Net produces fewer smooth deformations, which may be related to the weight selection of different losses during training.

[0092] Table 2. Experimental results compared with other methods on the OASIS dataset

[0093]

[0094] 4. Ablation and parameter studies:

[0095] This example addresses the scale selection and channel capacity expansion issues of large kernel convolutions. A systematic ablation experiment is conducted on the OASIS 2D dataset, quantitatively analyzing the synergistic effects of different convolution kernel sizes (3×3 to 21×21) and channel numbers (C=4 / 8) on registration accuracy.

[0096] Table 3. Hyperparameter ablation experiment on convolution kernel size and number of channels on the OASIS 2D dataset

[0097]

[0098] As shown in Table 3, as the convolution kernel size gradually expands from 3×3 to 9×9, the registration performance shows an improving trend: when C=4, the Dice score increases by 0.09 and 0.16 when the convolution kernel size increases from 3×3 to 6×6 and 9×9, respectively. This phenomenon confirms the advantage of a large receptive field in capturing long-range spatial correlation. However, when the convolution kernel size continues to expand to 15×15 and 21×21, the Dice coefficient shows a downward trend. This embodiment speculates that this is due to two factors: (1) an excessively large receptive field in a two-dimensional slice may span multiple anatomical structures, causing local features to be overwhelmed by global information; (2) the local deformation characteristics of brain images make the detailed information of local areas more discriminative. Based on this, this embodiment sets the default size of the convolution kernel to 9×9, achieving the optimal balance between global context modeling and local feature preservation.

[0099] Experiments demonstrating a positive correlation between performance and capacity revealed an increase in channel dimension when the ERF is fixed at 9×9. The Dice coefficient increased by 0.61 and 0.96 as the number of channels increased from 4 to 8 and 16, respectively. This suggests that by appropriately increasing the channel dimension, the model can more fully encode multi-scale feature representations, demonstrating stronger representational power in the high-dimensional feature space. If computational resources permit, performance can be further improved by moderately increasing the channel dimension. Furthermore, the trend in the convolution kernel size and the Dice coefficient remained consistent across different channel sizes.

[0100] Figure 6 and Figure 7 Different color bars in the figure represent different convolution kernel sizes.

[0101] To further validate the effectiveness of the convolutional attention and cross-scale attention optimization strategies, this example conducted ablation experiments on OASIS 2D. As shown in Table 4 (which evaluates the impact of the Large Kernel Attention (LKA) block, the Large Kernel Cross-Attention (LKCA) block, and the Cross-Resolution Attention Refinement (CRAR) mechanism on model performance), the network's Dice coefficient significantly decreased when the LKA block or the LKCA block were used alone, compared to when both mechanisms were used simultaneously, from 76.84% to 76.60% and 76.65%, respectively. This demonstrates the effectiveness of this "commonality-property" collaborative learning mechanism. The weight-sharing mechanism forces the network to learn common features of anatomical structures, while the cross-attention mechanism focuses on establishing correspondences between specific anatomical landmarks. Furthermore, the application of the cross-scale attention refinement mechanism allows the model to dynamically calibrate feature maps at different resolutions, improving the Dice coefficient from 76.84% to 76.99%.

[0102] Table 4. Ablation experimental study based on the OASIS-2D multimodal brain MRI dataset

[0103]

[0104] This example leverages the Transformer concept to transform traditional convolutional networks and proposes a novel deformable image registration framework, DELCA-Net. This framework leverages a self-attention mechanism to deeply understand the semantic information of images and achieves precise feature matching between images through a cross-self-attention mechanism. Furthermore, this example establishes an effective interaction mechanism between attention at different resolution levels, enabling a detailed characterization of attention.

[0105] To validate the effectiveness of DELCA-Net, this example conducted experiments on two 3D brain MR datasets and an ablation experiment on a 2D brain MR dataset. The experimental results show that the model in this example achieves state-of-the-art registration performance, fully demonstrating its superiority and practicality.

[0106] Although DELCA-Net has demonstrated good performance, it still has certain limitations. Therefore, in future work, this embodiment plans to further expand the model to multimodal registration tasks to handle more complex and diverse image registration scenarios. In addition, this embodiment will further optimize the model's parameters and computational complexity, introducing more efficient algorithms and architectural designs to reduce the model's complexity and operating costs, making it more suitable for practical application scenarios.

[0107] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A two-stream attention registration method based on shared weights of large kernel convolution LKA, characterized by: The input fixed and moving brain images enter the dual-stream structure of DELCA-Net respectively, and pass through multiple modules composed of large kernel convolution LKA blocks in the encoding path to extract features; Multiple LKA Blocks are used in series to abstract the shallow low-level features of edges and textures to the deep semantic features of structural shapes and functional area distribution layer by layer; the features extracted by the i-th layer of each LKA Block are recorded as and , respectively represent the weighted feature representations of the moving image and the fixed image; After feature extraction, and It is sent to CARM Block for feature refinement. LKCA is first introduced in CARM to and Perform cross attention modeling to get the cross attention of layer i , so as to enhance the saliency of corresponding structures between images in the case of non-rigid changes in tumors, brain tissue offsets or missing; In the Correct Layer, attention is paid to the i-th layer Make corrections to get upper-level attention , which enables the attention map to focus more accurately on key anatomical structures, including the border of the corpus callosum or the edge of the lesion; In the Fusion Layer, the upper attention map Downsampling and attention with layer i Align and perform channel splicing to fuse spatial details and structural semantic information at different scales to obtain the refined current layer attention map , which combines high-level semantic structure with low-level texture information to form a refined attention map with fine-grained structural information, providing accurate spatial guidance for subsequent registration; 、 , refined attention feature map , and the decoder output of layer i+1 , together constitute the input of the Decoder Block and fuse the four together in the channel dimension, and finally output the fused feature map , used to estimate the deformation field and complete image registration.

2. The shared weighted dual-stream attention registration method based on large kernel convolution (LKA) according to claim 1 is characterized by: In each large kernel convolution LKA Block, after downsampling, the Down operation is used to extract the local structural features of the image, including gyri, sulci, and cortical boundaries, through depthwise separable convolution DW-Conv and depthwise separable dilated convolution DW-D-Conv. 1×1 convolution is used to adjust the number of channels, and the PReLU activation function is introduced to enhance the nonlinear representation ability to improve the adaptability of typical medical image problems, including blurred boundaries and intensity changes. Among them, the additional depthwise separable convolution DC-Conv is used to enhance the perception of detailed areas of the ventricles and hippocampus.

3. The shared weighted dual-stream attention registration method based on large kernel convolution (LKA) according to claim 2 is characterized by: In the LKA Block, the depthwise separable dilated convolution DW-D-Conv integrates the information of the depthwise separable convolution DW-Conv to obtain a large spatial receptive field, and finally completes the cross-channel feature integration through 1×1 convolution; Decompose a large kernel convolution of size K*K into a convolution with a dilation rate of d Depthwise separable dilated convolution DW-D-Conv, depthwise separable convolution DW-Conv with a kernel size of 2d-1, and a 1×1 convolution; ； ； represents the input feature map, represents 1×1 convolution, represents depth-wise separable dilated convolution, Represents depth-separable convolution, Attention represents the attention after convolution extraction, represents element-wise multiplication, represents the output feature map; After obtaining the long-term dependency information, the importance of each point is estimated and the corresponding attention map is generated. The number of parameters of the decomposed large kernel convolution is and the amount of computation The calculation is as follows: ； ； Where C represents the number of input channels; H and W represent the size of the input feature map, which are height and width respectively; d represents the expansion rate, and K represents the convolution kernel size.

4. The shared weighted dual-stream attention registration method based on large kernel convolution (LKA) according to claim 1 is characterized in that: In CARM Block, enter the moving feature With fixed features , respectively, through the parameter-independent 1×1×1 convolutional layer to integrate cross-channel information, and through the adaptive gating mechanism of the GELU activation function to improve feature discrimination to obtain feature enhancement. and , fixed features Further expand the receptive field through DC Conv to capture a wide range of context information and obtain features containing global context information , and then the moving feature is realized by Hadamard product Features containing global context information Feature matching to obtain the cross attention of the current layer moving image and fixed image , as follows: ； ； ； ； in Represents the Hadamard product. The cross-scale attention optimization strategy downsamples the attention information of the high-resolution layer through the Fusion Layer and passes it to the current layer. The attention map of the current layer is upsampled through the Correct Layer, and the deviation in the attention map of the upper layer is corrected using the global information of the current layer's attention. The Fusion Layer and Connect Layer are specifically implemented as follows: ； ； represents two independent convolutional activation layers, represents the attention map of the previous layer, 、 It represents the attention map after refinement and adjustment, Cat represents the channel splicing operation, Down represents downsampling, and UP represents upsampling.

5. A shared-weight dual-stream attention registration system based on large kernel convolution LKA, characterized by: The system realizes image registration based on the method according to any one of claims 1 to 4.

Images