Image registration method and system combining coordinate perception and frequency-decoupled attention

Abstract

The present application belongs to the technical field of medical image registration, and provides an image registration method and system combining coordinate perception and frequency decoupled attention. The fixed image and the moving image and the three-dimensional coordinate grid are spliced to form a multi-channel composite input tensor. After preliminary fusion of the composite input tensor, local features are extracted. After projection and channel adjustment of the extracted local features, long-distance dependence and fine-grained anatomical details are captured through parallel multi-frequency branches. Spatial-channel hybrid attention modules are used for feature interaction of different branches. Deep semantic features are obtained through multi-scale fusion. The extracted deep semantic features are input into an implicit neural representation module to map the velocity field. The scaling and squaring algorithm is used for numerical integration of the velocity field to generate a differential homeomorphism deformation field, and the moving image is warped to obtain the registration result. The present application significantly improves the ability to describe fine-grained anatomical changes and registration accuracy.

Description

Technical Field

[0001] This invention belongs to the field of medical image registration technology, specifically relating to an image registration method and system that combines coordinate perception and frequency decoupled attention. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Deformable medical image registration aims to construct spatial mapping relationships between medical images from different time points, imaging devices, or imaging modalities, achieving pixel-level or voxel-level precise alignment. This technology plays a crucial role in numerous applications such as medical image analysis, clinical diagnosis, computer-aided therapy, and tumor radiotherapy. Its core objective is to capture the potential correspondences between different images through non-rigid transformation models, thereby establishing semantically consistent spatial correspondences between images. Traditional registration methods typically rely on predefined geometric models, interpolation strategies, and similarity metrics to find the optimal transformation between fixed and moving images through numerical optimization. However, these methods are highly sensitive to initial alignment, parameter selection, and image quality, easily getting trapped in local optima, and have low computational efficiency, especially performing poorly in high-resolution 3D images or large-scale datasets. Furthermore, the deformation complexity of different organs and the differences in intensity distribution among multimodalities further increase the modeling difficulty of traditional methods.

[0004] In recent years, with the widespread application of deep learning in computer vision, the research focus of deformable medical image registration has gradually shifted from traditional registration methods that rely on numerical optimization and target single image pairs to deep learning methods based on unsupervised neural networks. These methods significantly improve registration efficiency by training prediction models on a large number of image pairs to quickly estimate the spatial transformation between new image pairs. However, due to the locality of convolution operations, the receptive field of the model is limited, making it difficult for the network to capture the dependencies between long-distance features in the image. This limitation may lead to a decrease in registration accuracy in image registration tasks that require establishing global consistency or large-scale deformation.

[0005] Furthermore, for image pairs with complex deformations or subtle local differences, unsupervised deep learning-based image registration methods often struggle to accurately model their deformation relationships or effectively capture local correspondences. Additionally, unsupervised deep learning-based image registration methods have difficulty modeling the complex nonlinear correspondences between all image pairs, thus affecting the overall registration accuracy and the model's generalization ability to some extent.

[0006] Some studies have proposed deformation image registration methods based on implicit neural representations (also known as neural fields), providing new ideas for medical image registration tasks. Deformation image registration methods based on implicit neural representations are more suitable for modeling deformation fields with high requirements for accuracy, continuity, and smoothness. However, existing medical image registration methods based on implicit neural representations mostly use multilayer perceptrons (MLPs) with coordinates as input as the parameterization model of the deformation field, resulting in significant limitations in their expressive power in complex registration scenarios. Coordinate-based MLP architectures struggle to effectively extract high-level semantic features such as image texture and anatomical structure. Lacking direct guidance from visual information, the network relies solely on the mapping relationship from coordinates to displacement for optimization. When faced with complex and highly nonlinear deformation patterns, this modeling approach often gets trapped in local minima in low-contrast or texture-repetitive regions, leading to decreased registration accuracy.

[0007] Fully connected structures lack the translational equivariance of convolutional neural networks and the advantages of Transformers in long-range dependency modeling, resulting in a lack of visual inductive bias. This not only increases the risk of getting trapped in local optima during optimization but also introduces significant spectral bias in the frequency domain. In other words, a single MLP network struggles to simultaneously consider different frequency characteristics in multi-scale signal modeling, tending to prioritize fitting low-frequency global deformation trends while neglecting effective modeling of high-frequency details. This leads to an overly smooth deformation field in local regions, failing to accurately depict fine anatomical structural changes and limiting further improvements in registration accuracy. Furthermore, the computational complexity of these models increases quadratically with spatial resolution, easily causing a sharp increase in memory usage and computational overhead in 3D medical image registration. A trade-off between resolution and computational efficiency is usually required, thus limiting the effective modeling of global dependencies. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes an image registration method and system that combines coordinate perception with frequency decoupled attention. This invention can achieve deep fusion of multi-level features while significantly reducing computational complexity, thereby significantly improving the ability to characterize fine-grained anatomical changes and registration accuracy.

[0009] According to some embodiments, the present invention adopts the following technical solution: An image registration method combining coordinate awareness and frequency-decoupled attention includes the following steps: Acquire stationary and moving images; Construct a three-dimensional coordinate grid consistent with the image spatial resolution; The fixed image, the moving image, and the three-dimensional coordinate grid are stitched together to form a multi-channel composite input tensor; After initial fusion of the composite input tensors, local features are extracted and downsampled. After projecting and channel adjusting the extracted local features, long-distance dependencies and fine-grained anatomical details are captured through parallel multi-frequency branches. The spatial-channel hybrid attention module is used to perform feature interaction between different branches, and deep semantic features are obtained through multi-scale fusion. The extracted deep semantic features are input into the implicit neural representation module and mapped to obtain the velocity field; The velocity field is numerically integrated using scaling and squaring algorithms to generate a differential homeomorphic deformation field. This differential homeomorphic deformation field is then used to distort the moving image through a spatial transformation network to obtain the registration result.

[0010] As an alternative implementation, the process of extracting local features and downsampling includes: extracting local features and downsampling them using a feature extraction backbone network composed of multiple stacked RepVGG blocks, wherein each RepVGG block is a multi-branch parallel structure.

[0011] As a further defined implementation, the process of each RepVGG block being a multi-branch parallel structure is as follows: For the first... Each RepVGG block input feature tensor It contains three parallel computation paths: the first one is... Convolutional branches: used to capture contextual information of local anatomical structures; the second one is... Convolutional branch: responsible for feature fusion and projection between channels, increasing non-linear mapping capability; the third is identity mapping branch: exists only when the input and output dimensions are matched, constructing residual connections to preserve the original coordinates and image features. The outputs of all branches are batch normalized, then summed element by element, and activated by the GELU function.

[0012] As an alternative implementation, the process of capturing long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches after projecting and channel adjustment of the extracted local features includes: constructing a three-branch Transformer model, which introduces a frequency decoupling strategy, and distributes the input features to three parallel branches after linear projection, respectively modeling features for different frequency bands; The low-frequency feature branch is used to capture the overall structure and global deformation trend in the image, and the input features are dimension-mapped and downsampled by max pooling. The mid-frequency feature branch is used to model the contextual information and smooth structural transitions at the regional scale. The input features are linearly projected and then dimension-mapped, followed by downsampling using average pooling. The high-frequency feature branch does not perform any downsampling operation and directly models the features at the original feature resolution extracted from the RepVGG block.

[0013] As an alternative implementation, each branch introduces a spatial-channel hybrid attention module, which uses its spatial-channel hybrid attention mechanism to model the feature dependencies within the corresponding frequency band. The feature maps of the low-frequency feature branch and the mid-frequency feature branch are restored to the original spatial resolution through trilinear interpolation. The outputs of the three branches are stitched together in the channel dimension and then fused together through the output linear projection layer to obtain the final multi-frequency feature representation.

[0014] As an alternative implementation, the spatial-channel hybrid attention module includes a spatial branch convolutional additive spatial mixer and a channel branch multi-convolutional head attention module. The spatial branch convolutional additive spatial mixer is used to capture local spatial enhancement features, and the channel branch multi-convolutional head attention module is used to model and obtain global channel features. The two are added together and fused to obtain the final output.

[0015] As a further defined implementation, the spatial branch convolutional additive spatial mixer adopts a 3D depthwise separable convolutional structure. For the input feature map, a channel-wise depthwise convolutional layer independently models the local spatial context of each channel, enhancing spatial awareness without introducing inter-channel coupling. Instance Normalization and ReLU activation functions are introduced to improve the stability of feature distribution and enhance nonlinear expressive power. The channel dimension is compressed by pointwise convolution, and multi-channel features are mapped to single-channel spatial attention maps. The spatial attention weights are generated by activating the Sigmoid function. Based on the spatial attention weights, the spatial attention maps are weighted on the input features in the form of element-wise multiplication.

[0016] As an alternative implementation, the channel branch multi-convolutional head attention module adopts a self-attention mechanism built on the channel dimension. It implicitly encodes global context information by learning the cross-channel covariance relationship. Specifically, the input features are generated into query, key, and value through pointwise convolution, expanding the spatial dimension and rearranging the query, key, and value into a representation based on the channel dimension. The query and key are normalized on the feature dimension. The channel covariance attention map is obtained by calculating the transpose dot product of the normalized query and key. The final global feature representation is obtained by multiplying the attention map with the value.

[0017] As an alternative implementation, the feature extraction backbone network, the three-branch Transformer model, and the spatial-channel hybrid attention module constitute a registration network model. The registration network model is optimized and trained using loss functions, including a similarity loss function and a regularization loss function.

[0018] An image registration system combining coordinate awareness and frequency-decoupled attention includes: The image acquisition module is configured to acquire both stationary and moving images; The coordinate construction module is configured to build a 3D coordinate grid consistent with the image spatial resolution. The stitching module is configured to stitch together fixed and moving images and a 3D coordinate grid to form a multi-channel composite input tensor. The local feature extraction module is configured to extract local features and downsample them after performing preliminary fusion on the composite input tensor; The feature fusion module is configured to project and adjust the channels of the extracted local features, capture long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches, use the spatial-channel hybrid attention module to perform feature interaction between different branches, and obtain deep semantic features through multi-scale fusion. The implicit representation module is configured to input the extracted deep semantic features into the implicit neural representation module and map them to obtain the velocity field; The registration module is configured to numerically integrate the velocity field using a scaling and squaring algorithm to generate a differential homeomorphic deformation field. The moving image is then distorted using the differential homeomorphic deformation field through a spatial transformation network to obtain the registration result.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a frequency-decoupled hybrid Transformer implicit medical image registration method. To address the limitation of traditional coordinate-based INR methods in expressing complex deformation scenarios, this invention combines implicit neural representation with a frequency-aware hybrid Transformer structure, achieving high-precision differential homeomorphic registration within an instance-level optimization framework. This effectively alleviates the problem of insufficient deformation modeling capabilities of a single MLP.

[0020] This invention designs a coordinate-aware feature extraction backbone network and introduces visual inductive bias into implicit registration modeling. Unlike traditional INR methods that rely solely on coordinate input, this paper explicitly fuses a 3D coordinate grid with the image pair at the input end. By introducing a multi-branch RepVGG block structure, it injects a visual inductive bias with local translational equivariance while preserving absolute positional information, thereby enhancing the shallow semantic modeling capability of image texture and anatomical structure and improving registration robustness and accuracy.

[0021] This invention addresses the issue of low-frequency bias in implicit neural representations during multi-scale signal modeling. It mitigates the spectral bias of INR by using three-branch frequency decoupling: the low-frequency branch models the global deformation trend, the mid-frequency branch maintains smooth contextual information, and the high-frequency branch focuses on local edges and texture details. This complementary modeling at the frequency domain level ensures accurate alignment of complex anatomical structures.

[0022] This invention proposes a spatial-channel hybrid attention module. Addressing the issues of excessive computational complexity and unstable local detail modeling in existing Transformers for high-resolution 3D registration tasks, this invention achieves synergistic optimization of local anatomical detail modeling and global channel-dependent capture through a parallel structure of a convolutional additive spatial mixer and multi-convolutional head channel attention, significantly reducing complexity. This enables the model to efficiently process high-resolution 3D volumetric data while enhancing both global and local modeling capabilities.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 A schematic diagram of a method flow according to one embodiment; Figure 2 A schematic diagram of a RepVGG block in one embodiment; Figure 3 This is a schematic diagram of a registration network model according to one embodiment; Figure 4 This is a schematic diagram of a parallel spatial-channel hybrid attention module in one embodiment; Figure 5 This is a Dice score box plot of the proposed HiMiLo-RegNet method and the baseline method's anatomy on IXI data, representing one embodiment. Figure 6 This is a visualization of the proposed HiMiLo-RegNet method and baseline methods registered on the same image pairs on the IXI dataset, including stationary and moving images with their segmentation labels and difference maps, as well as distorted images of various methods with their segmentation labels and estimated deformations (displayed in grid and RGB), and difference maps between distorted and stationary images. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0030] Example 1 An image registration method combining coordinate perception and frequency-decoupled attention includes the following steps: Acquire stationary and moving images; Construct a three-dimensional coordinate grid consistent with the image spatial resolution; The fixed image, the moving image, and the three-dimensional coordinate grid are stitched together to form a multi-channel composite input tensor; After initial fusion of the composite input tensors, local features are extracted and downsampled. After projecting and channel adjusting the extracted local features, long-distance dependencies and fine-grained anatomical details are captured through parallel multi-frequency branches. The spatial-channel hybrid attention module is used to perform feature interaction between different branches, and deep semantic features are obtained through multi-scale fusion. The extracted deep semantic features are input into the implicit neural representation module and mapped to obtain the velocity field; The velocity field is numerically integrated using scaling and squaring algorithms to generate a differential homeomorphic deformation field. This differential homeomorphic deformation field is then used to distort the moving image through a spatial transformation network to obtain the registration result.

[0031] The following is a detailed description.

[0032] Unsupervised deformation image registration based on deep learning refers to the process of learning the optimal spatial transformation relationship between two images through a deep neural network without the supervision of registration labels. It can be represented as follows: (1) in, and These represent stationary and moving images, respectively. Represents the deformation field. Mapping arrive . This indicates that a distorted image is generated by twisting and moving the image through a deformation field. This represents the optimal deformation field. Loss function. It measures the similarity between a fixed image and a distorted image. Hyperparameters used to constrain the smoothness of the deformation field It plays a moderating role in maintaining the balance between image matching accuracy and deformation field smoothness.

[0033] Existing deformation image registration methods based on implicit neural representations typically treat the deformation field as a continuous function with coordinates as independent variables. Furthermore, the deformation field... Parameterized into a set of learnable parameters neural function with weights Defined continuous mapping: , can be represented as follows: (2) in , which are continuous coordinates in space. The characteristic of deformation image registration methods based on implicit neural representations is that neural functions... The input does not contain image pixel values; image information is only used to indirectly adjust the network parameters through the backpropagation gradient of the loss function. However, existing deformation image registration methods based on implicit neural representations typically employ simple coordinate-based multilayer perceptrons (MLPs) as parameterization models. While these methods have advantages in continuous representation, they still have significant limitations. On the one hand, simple MLPs struggle to effectively extract high-level semantic features such as image texture and anatomical structures, relying solely on coordinate-to-deformation mapping, making it difficult to characterize complex and highly nonlinear deformation patterns. On the other hand, a single MLP cannot simultaneously model multi-scale frequency features, tending to favor low-frequency global deformation while neglecting high-frequency details, resulting in overly smooth local deformation fields that limit the registration accuracy of fine anatomical structures.

[0034] To address this issue, the method in this embodiment primarily utilizes a novel registration network model (referred to as HiMiLo-RegNet in this embodiment). On one hand, the registration network model introduces a coordinate-aware RepVGG backbone to inject local visual inductive bias, fundamentally alleviating the problem of missing feature semantics. On the other hand, the HiMiLoTransformer, which decouples frequencies, explicitly separates high- and low-frequency information, effectively mitigating the modeling limitations caused by spectral bias in multi-scale feature spaces, thereby achieving high-precision alignment for complex deformations.

[0035] This embodiment proposes a deep parameterization form that uses a hybrid visual backbone as an implicit function. This allows the network to not only utilize coordinate priors but also fully mine the semantic features of the image itself, thereby achieving more robust convergence in unsupervised optimization. Specifically, as Figure 1 As shown, a 3D coordinate grid with the same spatial resolution as the input image is constructed, with dimensions of [missing information]. Subsequently, and This input tensor is concatenated with the coordinate grid to form a composite input tensor with 5 channels. This design enables the network to possess absolute position awareness from the initial stage, allowing it to distinguish deformation behavior at different spatial locations and thus improve its ability to model global deformation. The composite input tensor is then initially fused through a Merging layer, followed by a feature extraction backbone consisting of four stacked RepVGG blocks. This stage primarily extracts local features and performs downsampling. After processing by the RepVGG module, the spatial resolution of the feature map is reduced to a fraction of the original image. (Right now, The number of channels is expanded to 128. Next, a convolutional layer with a kernel size of 1 and a stride of 1 further projects and adjusts the channels of the features, then inputs them into a cascaded HiMiLo TransformerBlocks module. This module captures long-range dependencies and fine-grained anatomical details through parallel multi-frequency branches (high-frequency, mid-frequency, low-frequency), and utilizes a parallel spatial-channel mixture (CASM-MCA) attention module for feature interaction. Finally, deep semantic features are output through multi-scale fusion. Finally, the extracted deep semantic features are input into an Implicit Neural Representation (INR) module, mapping them to obtain a low-resolution velocity field. Subsequently, through a reshape operation and combined with a 2x upsampling, it is converted to a 3-channel, 128-resolution model. The steady-state velocity field is obtained. To ensure the topological preservation of the generated deformation (i.e., smooth and reversible), a scaling and squaring algorithm is used to numerically integrate the velocity field, generating the final differential homeomorphic deformation field. This deformation field is used to distort the moving image through a spatial transformation network to obtain the registration result. Finally, the parameters of the entire registration network are updated by backpropagation using a calculated loss function.

[0036] The following sections will detail the coordinate-aware RepVGG backbone, the HiMiLo Transformer module, the spatial-channel hybrid attention module, INR, and the scaling and squaring algorithms. Finally, the loss function of the proposed network will be introduced.

[0037] First is the coordinate-aware RepVGG backbone.

[0038] To address the shortcomings of the standard Vision Transformer in capturing local inductive biases and to effectively map inputs containing spatial location information to a high-dimensional feature space, the coordinate-aware RepVGG backbone was introduced. For example... Figure 2 As shown, to enrich the gradient flow and promote feature reuse, each RepVGG block always maintains a multi-branch parallel structure. Specifically, for the Each RepVGG block input feature tensor This module contains three parallel computation paths: 1, Convolutional branches: used to capture contextual information about local anatomical structures. 2. Convolutional branch: Responsible for feature fusion and projection between channels, increasing non-linear mapping capability. Identity mapping branch: Exists only when the input and output dimensions match, constructing residual connections to preserve the original coordinates and image features, effectively alleviating the gradient vanishing problem in deep network optimization. The outputs of all branches are batch normalized and then element-wise summed, followed by the GELU activation function. This process can be formally expressed as: (3) in, This represents the convolution operation. and The first RepVGG block convolution kernel parameters. Compared to single-path convolution, the multi-branch structure constructs a smoother loss surface, making it easier for the optimizer to find a better parameter solution. This generates a more accurate deformation field. In the designed network, four such RepVGG blocks are stacked to progressively downsample the 5-channel composite input (image + coordinates) and project it onto a 128-dimensional feature space, providing high-dimensional features rich in location information for the subsequent HiMiLo Transformer.

[0039] HiMiLo Transformer module Although introducing a coordinate-aware backbone network can extract certain local features, in traditional INR, a single MLP has to undertake the dual tasks of frequency decomposition and feature mapping. During optimization, a spectral bias problem is still prevalent, meaning the model is more likely to fit low-frequency components (overall structure) while its ability to model high-frequency information (fine textures and edges) is relatively insufficient. To alleviate this problem at the network structure level and enhance local fine-grained alignment capabilities while maintaining global modeling capabilities, this embodiment designs the HiMiLo Transformer module.

[0040] like Figure 3As shown, this module differs from the traditional Transformer approach of uniformly modeling all tokens. Instead, it introduces a frequency decoupling strategy, distributing the input features, after linear projection, to three parallel branches, each modeling features for a different frequency band. Let the total dimension of the input features be... ,in and These represent the proportions of feature dimensions allocated to the low-frequency branch and the mid-frequency branch, respectively, and the remaining dimensions... Assign to high-frequency branches.

[0041] The low-frequency feature branch is primarily used to capture the overall structure and global deformation trends in an image. This branch first maps the feature dimensions to... Subsequently, downsampling is performed using max pooling. Max pooling can preserve the most significant structural information in local regions, while effectively expanding the receptive field of subsequent attention modules by reducing spatial resolution, enabling them to model long-distance anatomical correspondences with lower computational overhead. The mid-frequency feature branch focuses on modeling contextual information at the region scale and smooth structural transitions. After linear projection, the features are mapped to the following dimensions: Subsequently, average pooling is used for downsampling. Compared to max pooling, average pooling has low-pass filtering properties, which helps suppress high-frequency noise and maintain the continuity of background tissue, thus providing a stable mesoscale constraint for deformation estimation. The high-frequency feature branch does not undergo any downsampling operation and directly models the features at the original feature resolution extracted by RepVGG, with a dimension of [missing value]. By fully preserving spatial resolution, this branch enhances the network's attention to fine-grained anatomical boundaries and microstructures, mitigating to some extent the problem of insufficient modeling of high-frequency information in implicit neural representations during optimization, thereby preventing the loss of key details in the generated deformation field. In this embodiment, the following is set... and The values ​​are 0.2 and 0.3 respectively.

[0042] Furthermore, a CASM-MCA module is introduced within each frequency branch to model feature dependencies within the corresponding frequency band using its spatial-channel hybrid attention mechanism. Subsequently, the feature maps of the low-frequency and mid-frequency branches are interpolated using trilinear interpolation. The original spatial resolution is restored. Finally, the outputs of the three branches are concatenated along the channel dimension and fused together using an output linear projection layer to obtain the final multi-frequency feature representation.

[0043] (4) Through the above-mentioned hierarchical frequency decoupling design, the HiMiLo Transformer module alleviates the problem of insufficient high-frequency detail modeling in the implicit neural representation optimization process to a certain extent. At the same time, it significantly improves the robustness of registration through multi-scale parallel processing, and realizes accurate modeling of the full spectrum information of medical images.

[0044] Spatial-Channel Hybrid Attention (CASM-MCA) Module Traditional Vision Transformers model all tokens uniformly, leading to high computational redundancy and a tendency to favor global low-frequency contours, thus weakening attention to local anatomical details. To simultaneously model local spatial structure and global channel dependency within the HiMiLoTransformer module while reducing computational overhead, the CASM-MCA module is proposed, such as... Figure 4 As shown, this module consists of two parts: a spatial branch convolutional additive spatial mixer (CASM) and a channel branch multi-convolutional head attention (MCA), and adopts an additive fusion strategy to achieve efficient information integration.

[0045] Convolutional additive spatial mixer: The CASM branch aims to capture local spatial structure information, such as edges, textures, and anatomical variations, in a lightweight manner. Specifically, this branch employs a 3D depthwise separable convolutional structure, given an input feature map... First, through a channel-by-channel Deep convolution independently models the local spatial context for each channel, thereby enhancing spatial awareness without introducing inter-channel coupling. Subsequently, instance normalization and ReLU activation functions are introduced to improve feature distribution stability and enhance non-linear expressive power. Then, through... Pointwise convolution compresses the channel dimension, mapping multi-channel features to a single-channel spatial attention map, which is then activated by the sigmoid function to generate spatial attention weights. Finally, this spatial attention map weights the input features using element-wise multiplication, effectively enhancing the model's ability to perceive locally deformed regions.

[0046] (5) in, and These represent depthwise convolution and pointwise convolution, respectively. Represents the ReLU activation function. This indicates element-wise multiplication.

[0047] Multi-convolutional head attention (MCA) CASM focuses on capturing local spatial structure information, while the MCA branch aims to capture global long-range dependencies. Considering standard self-attention in 3D voxel space (where... The extremely high computational complexity caused by this This embodiment chooses to construct a self-attention mechanism along the channel dimension, implicitly encoding global contextual information by learning cross-channel cross-covariance relationships, effectively improving computational efficiency while ensuring modeling capability. Specifically, input features... pass Pointwise convolution generates a query (i.e., a query). ), key ( ) and value (i.e., value) ),in, , It's the number of heads. Then, the spatial dimensions are expanded, making... The rearrangement is presented in a channel-dimension-centric format, and... and Perform on the feature dimension Normalization was applied to reduce the impact of differences in response amplitudes across different channels on the stability of the attention distribution. This was achieved through... and The channel covariance attention map is obtained by calculating the transpose dot product. : (6) in, and It is the normalized query and key matrix. These are learnable scaling parameters. The final global feature representation is achieved through the attention map. and Multiplying them together gives: (7) Since the MCA module performs attention calculations only on the channel dimension, its main computational overhead is linearly related to the number of spatial tokens, which can be expressed as: This effectively avoids the computational bottleneck of traditional self-attention mechanisms that increases with the square of spatial resolution.

[0048] Finally, the local spatial enhancement features output from the CASM branch are added to the global channel features modeled by the MCA branch and fused to form the final output of the module. This parallel additive structure allows gradients to propagate directly between the two branches, which is beneficial for stable training and efficient optimization. The fusion process can be represented as follows: (8) Features after fusion It also encodes spatial attention and channel attention information, providing a robust feature foundation for the subsequent generation of dense deformation fields.

[0049] INR and differential homeomorphic registration method For the IRN, this embodiment employs a 4-layer multilayer perceptron, where each hidden layer has 128 hidden units. Specifically, the IRN is designed as follows: (9) in, They represent the first Layer weights, biases, and activation functions, p i Let i be a spatial point in the grid, where i is 0 and L is a space. Activation function Gaussian Error Linear Units (GELU) are used.

[0050] In this embodiment, a differential homeomorphic deformation model is used to estimate the deformation field. This is because differential homeomorphic registration preserves the smoothness, reversibility, and topological preservation of the transformation relationship between the two images. In differential homeomorphic image registration, the deformation field... Specifically, it can be expressed as: (10) in, (11) Here, in this embodiment, the deformation field By analyzing the stationary velocity field in ordinary differential equation (11) exist Integrating the above, we obtain equation (10). This represents the identity transform. Furthermore, in this embodiment, S=1, therefore the final deformation field is... .

[0051] In differential homeomorphic image registration, the stationary velocity field Elements considered as Lie algebras generate time-1 deformable fields via exponential mappings. The deformation field belongs to the corresponding Lie group, thus forming a parameterized differential homeomorphic subgroup, achieving continuous and invertible deformation modeling. To obtain the deformation field with time 1... The scaling and squaring method

[39] was used to analyze the stationary velocity field. Integrate over time t. Specifically, give the initial deformation field. Where T represents the total time step used in the proposed method, and in this embodiment, T=7 is set. Therefore, through recursive calculation... The final deformation field can be obtained. .

[0052] The loss function used in the method proposed in this embodiment consists of a similarity loss function and a regularization loss function. According to formula (1), the complete loss function in this paper is expressed by the following formula: (12) Among them, the similarity loss function It measures the similarity between distorted and fixed images. Regularized loss function. Constraining the spatial gradient of the deformation field makes the estimated deformation field smooth. These are the weights of the smooth loss function.

[0053] The similarity loss function in this embodiment is composed of the local normalized cross-correlation metric (LNCC), and its mathematical expression is as follows: (13) in, Represents the image domain. and respectively with One of the centers The expected value of the size window, where . It represents in The element intensity values ​​within the large and small windows, in the experiment The value is set to 9. f It is a fixed image, I m Since the images are moving, the LNCC value ranges from [-1, 1]. -1 represents no correlation between the two images, while 1 represents a very high correlation. Therefore, the final similarity loss function is: .

[0054] Regularized loss function The goal is to avoid generating folded deformation fields, typically by applying local smoothing constraints to the spatial gradient to smooth the estimated deformation field. Therefore, the regularized loss function... The definition is as follows: (14) in, Represents the displacement field At point gradient at, It represents Norm. In this embodiment, the spatial gradient is approximated by the difference between adjacent voxels.

[0055] In this embodiment, a comparative analysis is performed. The OASIS dataset and the IXI dataset are used as experimental datasets. The OASIS dataset includes 414 brain T1-weighted MR images, which are preprocessed using FreeSurfer, including skull dissection, spatial normalization, subcutaneous tissue segmentation, and affine transformation. Finally, all images are cropped to a size of [size missing]. For learning-based methods, to achieve atlas-to-patient registration, similar to the baseline method, the obtained brain MRI atlases were used as stationary images, and 576 brain T1-weighted MR images from the IXI dataset were used as moving images. The IXI dataset was divided into a training set of 403 images, a validation set of 58 images, and a test set of 115 images. Finally, the registration performance was validated using subcortical segmentation maps of 30 provided anatomical structures as ground truth. For registration methods based on implicit neural representations, registration was performed using only the test set of each dataset.

[0056] This embodiment uses the following evaluation metrics to compare the proposed method with baseline methods: 1. Dice similarity coefficient (DSC). DSC calculates the volume overlap between the same anatomical / organ segments in the distorted and fixed images. The DSC value ranges from 0 to 1; a higher score indicates higher registration accuracy. Finally, the average DSC score of all methods for the provided anatomical / organ structures is taken as the final reported score, and the standard deviation of DSC is calculated to evaluate the stability of the method. 2. Hausdorff distance (HD95), which describes the similarity between the same anatomical / organ segments in the distorted and fixed images. HD95 represents the 95th percentile; the smaller the HD95, the greater the similarity between the same anatomical / organ segments in the distorted and fixed images. 3. Jacobian matrix. The Jacobian matrix quantifies the regularity of the deformation field. This embodiment reports the percentage of non-positive values ​​in the determinant of the Jacobian matrix in the deformation field (e.g., ...). ) and the standard deviation of the percentage of negative Jacobian determinant values. The Jacobian determinant can represent the properties of the local deformation field. When the Jacobian determinant is positive, it can guarantee the invertibility and topological preservation of the local deformation field. 4. Wilcoxon Rank-Sum Test. The Wilcoxon Rank-Sum Test is a non-parametric statistical method commonly used to test whether there is a significant difference between the distributions of two groups of sample data. This method calculates the rank sum by jointly ranking the two groups of samples and infers whether there is a statistically significant difference between the two groups. In this embodiment, the Wilcoxon Rank-Sum Test is used to compare the statistical model performance of the proposed method with the baseline method. Generally, a p-value less than 0.05 indicates that there is a significant difference between the two sample data. 5. GPU utilization. GPU utilization represents the amount of GPU used by the model, in MB.

[0057] The proposed method was systematically compared with currently recognized representative registration methods, covering 2 traditional methods, 6 deep learning-based unsupervised deformation image registration methods, and 2 neural field-based registration methods. The hyperparameter settings for the traditional methods were referenced from the original method, while the settings for the other deep learning and neural field-based methods were strictly replicated according to the settings provided in their respective papers. Furthermore, the training cycle for the deep learning-based unsupervised deformation image registration method was set to 500. Specific parameter settings are explained below: SyN: Implement SyN in the publicly available Python medical imaging library ANTsPy, where the mean squared error is used as the optimization metric, the number of iterations is set to (160, 80, 40), and the transformation type is set to 'SyNOnly'.

[0058] NiftyReg: Implements a fast free deformation algorithm on the open-source medical image registration tool NiftyReg. Each image pair is converted in the "reg_f3d" executable file, using mean squared error as the optimization metric, and setting the weight of the bending energy penalty term to 0.0002.

[0059] VoxelMorph: Apply the official VoxelMorph-1 model.

[0060] ViT-V-Net: Based on the official implementation details, some minor adjustments have been made, including setting the batch size to 1 and using NCC as the similarity metric function.

[0061] TransMorph: Same implementation details as provided by the official documentation.

[0062] TransMatch: Same implementation details as provided in the official documentation.

[0063] RDP: Due to GPU memory limitations, the same approach as in UTSRMorph is used, setting the number of RDP channels to 12 on the OASIS and IXI datasets, and to 10 on the Abdomen dataset. Learning rate and regularization hyperparameters are also used. Set them to 0.0001 and 1 respectively.

[0064] UTSRMorph: Same implementation details as provided by the official documentation.

[0065] IDIR: Optimize for 2500 epochs per pair of images, with a learning rate of 0.0001, and employing bending energy penalty and Jacobi regularization, with hyperparameters set to 10 and 0.05, respectively.

[0066] NODEO: Optimizes for 600 epochs per image pair. The implementation details on the OASIS dataset are the same as those provided in the official documentation. On the IXI dataset, the learning rate and smoothing regularization hyperparameter are set to 0.001 and 0.5, respectively.

[0067] The proposed method was implemented in an environment with an NVIDIA GeForce RTX 4090 GPU, Ubuntu 20.04, PyTorch 2.0.1, and Python 3.8. The Adam optimizer was used to update network parameters with a batch size of 1. On the OASIS, IXI, and LPBA datasets, the epoch was set to 600 and the learning rate to 0.0005, while LNCC and [other parameters were assigned]. As a loss function, the regularization hyperparameter Set all to 1.

[0068] On the IXI dataset, as shown in Table 1, the proposed HiMiLo-RegNet outperforms all baseline methods in DSC, its standard deviation, and HD. Specifically, compared to traditional methods, deep learning-based unsupervised image registration methods, and neural field-based registration methods, HiMiLo-RegNet achieves average improvements of 17.01%, 2.76%, and 2.38% in DSC exponent, respectively. This embodiment uses the Wilcoxon Rank-Sum Test to compare the DSC performance of the proposed method with baseline methods for statistical performance comparison. The p-values ​​for TGINRMorph are all p < 0.05 compared to the baseline methods. Figure 5As shown, the proposed HiMiLo-RegNet achieves the best DSC for most segmentation structures. It can be observed that the deformation field obtained by traditional methods exhibits significant local over-distortion and loss of detail, with uneven mesh deformation leading to limited overall registration accuracy. Convolutional learning methods improve local registration accuracy to some extent, but the deformation field still exhibits strong irregular curvature, insufficient smoothness, and inadequate boundary preservation for some small-volume structures. In contrast, the deformation field obtained by transformer learning methods is smoother, but still has limitations in preserving boundaries and fine structures. IDIR and NODEO, due to their continuous representation characteristics, produce relatively smooth results with less folding compared to traditional and learning-based methods. However, they sometimes sacrifice some local details at the boundaries. The HiMiLo-RegNet method proposed in this embodiment demonstrates superior characteristics in both deformation field visualization and registration results: its deformation field is generally smooth with almost no folding, its RGB representation is continuous and can simultaneously reflect local differences; simultaneously, in the segmentation results, it can better preserve the internal structure and boundary information of complex brain regions and small-volume organs.

[0069] like Figure 6 As shown, apart from the traditional method, the difference slice images of the other methods appear visually similar, indicating that each method can achieve high-quality registration results. It is important to note that the difference image primarily reflects the intensity difference between image gray levels, rather than the degree of alignment of structural positions. When the registration error is small and the intensity difference is normalized and compressed, the difference image is not sensitive to subtle changes in local deformation, thus the visual difference is limited. Although the visual changes are not significant, the improvement in quantitative indicators still shows that HiMiLo-RegNet outperforms baseline methods in terms of structural alignment accuracy.

[0070] Table 1 shows the results of different methods for the IXI dataset.

[0071]

[0072] In summary, HiMiLo-RegNet overcomes the spectral bias of implicit neural representations (INRs) while preserving absolute positional information by fusing the coordinate-aware RepVGG backbone, the HiMiLo Transformer module, and the spatial-channel hybrid attention (CASM-MCA) module. This significantly enhances the model's ability to model multi-scale details and local anatomical structures while effectively reducing computational complexity. Experimental results show that HiMiLo-RegNet achieves state-of-the-art registration performance on multiple datasets.

[0073] Example 2 An image registration system combining coordinate awareness and frequency-decoupled attention includes: The image acquisition module is configured to acquire both stationary and moving images; The coordinate construction module is configured to build a 3D coordinate grid consistent with the image spatial resolution. The stitching module is configured to stitch together fixed and moving images and a 3D coordinate grid to form a multi-channel composite input tensor. The local feature extraction module is configured to extract local features and downsample them after performing preliminary fusion on the composite input tensor; The feature fusion module is configured to project and adjust the channels of the extracted local features, capture long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches, use the spatial-channel hybrid attention module to perform feature interaction between different branches, and obtain deep semantic features through multi-scale fusion. The implicit representation module is configured to input the extracted deep semantic features into the implicit neural representation module and map them to obtain the velocity field; The registration module is configured to numerically integrate the velocity field using a scaling and squaring algorithm to generate a differential homeomorphic deformation field. The moving image is then distorted using the differential homeomorphic deformation field through a spatial transformation network to obtain the registration result.

[0074] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0075] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0077] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image registration method combining coordinate perception and frequency-decoupled attention, characterized in that, Includes the following steps: Acquire stationary and moving images; Construct a three-dimensional coordinate grid consistent with the image spatial resolution; The fixed image, the moving image, and the three-dimensional coordinate grid are stitched together to form a multi-channel composite input tensor; After initial fusion of the composite input tensors, local features are extracted and downsampled. After projecting and channel adjusting the extracted local features, long-distance dependencies and fine-grained anatomical details are captured through parallel multi-frequency branches. The spatial-channel hybrid attention module is used to perform feature interaction between different branches, and deep semantic features are obtained through multi-scale fusion. The extracted deep semantic features are input into the implicit neural representation module and mapped to obtain the velocity field; The velocity field is numerically integrated using scaling and squaring algorithms to generate a differential homeomorphic deformation field. The moving image is then distorted using the differential homeomorphic deformation field through a spatial transformation network to obtain the registration result. After projecting and adjusting the channels of the extracted local features, the process of capturing long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches includes: constructing a three-branch Transformer model, which introduces a frequency decoupling strategy, and distributes the input features to three parallel branches after linear projection, respectively modeling features in different frequency bands; The low-frequency feature branch is used to capture the overall structure and global deformation trend in the image, and the input features are dimension-mapped and downsampled by max pooling. The mid-frequency feature branch is used to model the contextual information and smooth structural transitions at the regional scale. The input features are linearly projected and then dimension-mapped, followed by downsampling using average pooling. The high-frequency feature branch does not perform any downsampling operation and directly models the features at the original feature resolution extracted from the RepVGG block.

2. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 1, characterized in that, The process of extracting local features and downsampling includes: extracting local features and downsampling them using a feature extraction backbone network composed of multiple stacked RepVGG blocks, where each RepVGG block is a multi-branch parallel structure.

3. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 2, characterized in that, The process of each RepVGG block being a multi-branch parallel structure is as follows: For the first... Each RepVGG block input feature tensor It contains three parallel computation paths: the first one is... Convolutional branches: used to capture contextual information of local anatomical structures; the second one is... Convolutional branch: responsible for feature fusion and projection between channels, increasing non-linear mapping capability; the third is identity mapping branch: exists only when the input and output dimensions are matched, constructing residual connections to preserve the original coordinates and image features. The outputs of all branches are batch normalized, then summed element by element, and activated by the GELU function.

4. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 1, characterized in that, Each branch introduces a spatial-channel hybrid attention module, which uses its spatial-channel hybrid attention mechanism to model the feature dependencies within the corresponding frequency band. The feature maps of the low-frequency feature branch and the mid-frequency feature branch are restored to the original spatial resolution through trilinear interpolation. The outputs of the three branches are stitched together in the channel dimension and then fused together through the output linear projection layer to obtain the final multi-frequency feature representation.

5. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 1, characterized in that, The spatial-channel hybrid attention module includes a spatial branch convolutional additive spatial mixer and a channel branch multi-convolutional head attention module. The spatial branch convolutional additive spatial mixer is used to capture local spatial enhancement features, and the channel branch multi-convolutional head attention module is used to model and obtain global channel features. The two are added together and fused to obtain the final output.

6. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 5, characterized in that, The spatial branch convolutional additive spatial mixer adopts a 3D depthwise separable convolutional structure. For the input feature map, a channel-wise deep convolutional layer independently models the local spatial context of each channel, enhancing spatial awareness without introducing inter-channel coupling. Instance Normalization and ReLU activation functions are introduced to improve the stability of feature distribution and enhance nonlinear expressive power. The channel dimension is compressed by pointwise convolution, and multi-channel features are mapped to single-channel spatial attention maps. The spatial attention weights are generated by activating the Sigmoid function. Based on the spatial attention weights, the spatial attention maps are weighted on the input features in the form of element-wise multiplication.

7. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 1, characterized in that, The channel-branch multi-convolutional head attention module employs a self-attention mechanism built on the channel dimension. It implicitly encodes global context information by learning cross-channel covariance relationships. Specifically, input features are generated into queries, keys, and values ​​through pointwise convolution, expanding the spatial dimension and rearranging the queries, keys, and values ​​into a channel-centric representation. The queries and keys are then normalized on the feature dimension. The channel covariance attention map is obtained by calculating the transpose dot product of the normalized queries and keys. The final global feature representation is obtained by multiplying the attention map with the values.

8. The image registration method combining coordinate perception and frequency decoupled attention as described in claim 4, characterized in that, The feature extraction backbone network, the three-branch Transformer model, and the spatial-channel hybrid attention module constitute the registration network model. The registration network model is optimized and trained using loss functions, including similarity loss function and regularization loss function.

9. An image registration system combining coordinate perception and frequency-decoupled attention, characterized in that, include: The image acquisition module is configured to acquire both stationary and moving images; The coordinate construction module is configured to build a 3D coordinate grid consistent with the image spatial resolution. The stitching module is configured to stitch together fixed and moving images and a 3D coordinate grid to form a multi-channel composite input tensor. The local feature extraction module is configured to extract local features and downsample them after performing preliminary fusion on the composite input tensor; The feature fusion module is configured to project and adjust the channels of the extracted local features, capture long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches, use the spatial-channel hybrid attention module to perform feature interaction between different branches, and obtain deep semantic features through multi-scale fusion. The implicit representation module is configured to input the extracted deep semantic features into the implicit neural representation module and map them to obtain the velocity field; The registration module is configured to numerically integrate the velocity field using a scaling and squaring algorithm to generate a differential homeomorphic deformation field, and then use the differential homeomorphic deformation field to distort the moving image through a spatial transformation network to obtain the registration result. After projecting and adjusting the channels of the extracted local features, the process of capturing long-distance dependencies and fine-grained anatomical details through parallel multi-frequency branches includes: constructing a three-branch Transformer model, which introduces a frequency decoupling strategy, and distributes the input features to three parallel branches after linear projection, respectively modeling features in different frequency bands; The low-frequency feature branch is used to capture the overall structure and global deformation trend in the image, and the input features are dimension-mapped and downsampled by max pooling. The mid-frequency feature branch is used to model the contextual information and smooth structural transitions at the regional scale. The input features are linearly projected and then dimension-mapped, followed by downsampling using average pooling. The high-frequency feature branch does not perform any downsampling operation and directly models the features at the original feature resolution extracted from the RepVGG block.

Images