A method and system for MRI image registration

Abstract

The application provides an MRI image registration method, which comprises the following steps: feature extraction is performed on a fixed image and a floating image to be registered to obtain feature maps of multiple scales; an initial deformation field is predicted by an attention mechanism at the lowest resolution scale, and the floating image feature map is coarsely aligned; the deformation field is progressively up-sampled from a low resolution to a high resolution; after each up-sampling, the fixed image feature map and the floating image feature map transformed by the current deformation field are combined, a residual deformation field is predicted by using similarity matching and edge structure double clues fusion, and the residual deformation field is superimposed with the current deformation field to obtain a cumulative deformation field as the input of the next up-sampling; and the deformation field obtained at the highest resolution is applied to the floating image to complete the image registration. The application solves the problems in the prior art, such as difficulty in balancing global deformation capture and calculation efficiency, error accumulation in multi-scale deformation estimation, insufficient local detail alignment, and poor robustness of feature extraction during image registration.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an MRI image registration method and system. Background Technology

[0002] Medical image registration is a fundamental technique in the field of medical image processing. It aims to transform images acquired at different times, modalities, or from different individuals into a unified spatial coordinate system, aligning identical anatomical structures in space. Registered images can fuse multi-source information, providing crucial support for disease diagnosis, surgical navigation, treatment planning, and population statistical analysis. Magnetic Resonance Imaging (MRI) of the brain, due to its excellent soft tissue contrast and non-invasiveness, plays a vital role in neuroscience research and the clinical diagnosis and treatment of brain diseases. Accurate registration of brain MRI images is a prerequisite for constructing brain atlases, analyzing changes in brain structure, and locating lesions.

[0003] Traditional image registration methods are mainly divided into two categories: feature-based and grayscale-based. Feature-based methods rely on the extraction and matching of geometric features such as key points and edges in the image. Their accuracy is limited by the accuracy of feature extraction and is prone to large errors when the anatomical structure is complex or the image quality is poor. Grayscale-based methods directly utilize pixel grayscale information and solve for the deformation field by iteratively optimizing the similarity metric. Although this avoids the shortcomings of feature extraction, the iterative process is computationally intensive and the registration speed is slow, making it difficult to meet the real-time requirements of clinical practice. In recent years, deep learning methods, especially Convolutional Neural Networks (CNNs), have been introduced into image registration tasks. By directly predicting the deformation field through an encoder-decoder structure, they have significantly improved the registration speed. Building on this, to further enhance global modeling capabilities, researchers have begun to apply Transformers (a deep learning structure based on self-attention mechanisms that excels at modeling global information) to registration networks, using self-attention mechanisms to capture long-range spatial dependencies.

[0004] Despite the significant advancements in speed and accuracy achieved by the aforementioned deep learning registration methods, several unresolved technical challenges remain. Existing medical image registration methods struggle to achieve a good balance between global deformation capture and computational efficiency. Methods relying solely on CNNs have limited receptive fields, while methods incorporating Transformers incur excessive parameter counts and computational overhead. Furthermore, during multi-scale progressive deformation estimation, estimation errors from lower resolution stages tend to propagate and accumulate to higher resolutions. Additionally, the high-resolution stage lacks sufficient alignment capability for local anatomical details and edge structures, leading to blurred boundaries or local misalignment in the registration results. Moreover, the robustness of existing feature extraction networks to imaging noise and local blurring needs improvement, impacting the stability of deformation field prediction. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide an MRI image registration method and system, which aims to solve the problems in the prior art where it is difficult to balance global deformation capture and computational efficiency, error accumulation in multi-scale deformation estimation and insufficient alignment of local details, and poor robustness of feature extraction.

[0006] An MRI image registration method according to an embodiment of the present invention includes: Obtain a fixed image and a floating image to be registered, and perform feature extraction on the fixed image and the floating image to determine feature maps at multiple different resolution scales; The attention mechanism is used to process the fixed image feature map and the floating image feature map at the lowest resolution scale to predict the initial deformation field. The initial deformation field is then used to transform and coarsely align the floating image feature map at the current scale to obtain the floating feature map. The initial deformation field is upsampled stepwise from the lowest resolution scale to the highest resolution scale. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. Then, the residual deformation field is predicted by fusing the fixed image feature map at the current scale and the deformed floating feature map through dual-clue fusion of similarity matching cues and edge structure cues. The current upsampled deformation field is superimposed with the residual deformation field to obtain the cumulative deformation field at the current scale. This cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained. The highest resolution deformation field is applied to the floating image to generate a registered image.

[0007] In addition, the MRI image registration method according to the above embodiments of the present invention may also have the following additional technical features: Furthermore, the step of extracting features from the fixed image and the floating image to determine feature maps at multiple different resolution scales includes: The fixed image and the floating image are respectively input into a dual-stream shared hierarchical feature encoder, so that the input image is downsampled step by step by a residual block and multiple residual attention blocks cascaded in the hierarchical feature encoder, thereby obtaining multiple feature maps of different resolution scales output by the residual block and the residual attention block; In this process, the feature map size is reduced between adjacent blocks using average pooling; In the residual block, local structural features are extracted through two layers of 3D convolution, instance normalization, GELU activation function, and residual connections; In the residual attention block, the operation of the residual block is executed first, and then the channel response is adaptively recalibrated through the preset channel attention module.

[0008] Furthermore, the steps for processing the fixed and floating image feature maps at the lowest resolution scale based on the attention mechanism to predict the initial deformation field include: The fixed image feature map and the floating image feature map at the lowest resolution scale are concatenated along the channel dimension and then input into the window channel self-attention module; In the window channel self-attention module, the current input feature map is normalized by layer, and then the window self-attention operation and the shift window self-attention operation are executed alternately. At the same time as each self-attention operation is executed, the channel attention operation is executed on the same normalized feature map. The result of the self-attention operation and the result of the channel attention operation are added element by element to obtain the intermediate feature map. The intermediate feature map obtained in the previous round is used as the new current input feature map. The above steps for determining the intermediate feature map are repeated at least once, and the intermediate feature map obtained in the last round is used as the initial deformation field. The window channel self-attention module does not include position encoding and adopts a single-head self-attention mechanism.

[0009] Furthermore, the steps for predicting the residual deformation field through dual-clue fusion of similarity matching cues and edge structure cues include: The local window similarity between the fixed image feature map at the current scale and the transformed floating image feature map is calculated through the similarity matching branch, and the first sub-deformation field is generated based on the similarity weighted predefined relative displacement grid. By using edge deformation branching and edge convolution to extract structural boundary features of the fixed image feature map at the current scale and the transformed floating image feature map, the second sub-deformation field is generated through regression. The first sub-deformation field and the second sub-deformation field are concatenated and fused in the channel dimension to generate a velocity field. The velocity field is scaled and integraled squared to output a residual deformation field that satisfies the homeomorphic constraint.

[0010] Furthermore, the step of generating the first sub-deformation field through the similarity matching branch includes: For each voxel location, the dot product similarity is calculated between the feature vectors of the fixed image feature map and the transformed floating image feature map within a preset local window; The matching weight matrix is ​​obtained by applying a flexible maximum function along the window dimension; The first sub-deformation field is obtained by multiplying the matching weight matrix with the predefined relative displacement grid.

[0011] Furthermore, the step of generating the second sub-deformation field through the edge deformation branch includes: Edge enhancement features are extracted from both the fixed image feature map and the transformed floating image feature map using edge convolution. The second sub-deformation field is obtained by concatenating the two sets of edge enhancement features in the channel dimension and then performing three-dimensional convolutional regression.

[0012] Furthermore, the method employs an unsupervised loss function for training, which includes an image similarity metric and a deformation field smoothing regularization term.

[0013] Another object of the present invention is to provide an MRI image registration system for implementing the above-described MRI image registration method, the system comprising: The feature extraction module is used to acquire the fixed image and the floating image to be registered, and to perform feature extraction on the fixed image and the floating image to determine feature maps at multiple different resolution scales; The coarse alignment module is used to process the fixed image feature map and the floating image feature map at the lowest resolution scale based on the attention mechanism to predict the initial deformation field, and to transform and coarsely align the floating image feature map at the current scale using the initial deformation field to obtain the floating feature map. The step-by-step processing module is used to upsample the initial deformation field from the lowest resolution scale to the highest resolution scale step by step. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. Then, the fixed image feature map at the current scale and the deformed floating feature map are used to predict the residual deformation field by fusing similarity matching cues and edge structure cues. The current upsampled deformation field and the residual deformation field are superimposed to obtain the cumulative deformation field at the current scale. The cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained. The registration module is used to apply the highest resolution deformation field to the floating image to generate a registered image.

[0014] Another objective of this invention is to provide a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the MRI image registration method described above.

[0015] Another object of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the MRI image registration method described above.

[0016] This invention extracts multi-scale features by designing a lightweight hierarchical feature encoder. At the lowest resolution scale, a window channel self-attention module captures global deformation at low computational cost. Subsequently, it upsamples progressively to higher resolution scales, predicting residual deformation at each level through similarity matching and edge structure dual-cue fusion. The upsampled deformation is then superimposed with the residual deformation to achieve progressive error compensation. This scheme uses the window self-attention mechanism only at the low-resolution stage, eliminating the need to deploy a highly complex Transformer module on the full-resolution feature map, significantly reducing the number of parameters and computational overhead. Simultaneously, dual-cue fusion explicitly strengthens local matching and edge constraints in the fine registration stage, effectively suppressing multi-scale error accumulation and maintaining the clarity of anatomical boundaries without relying on large amounts of labeled data or repeated iterative optimization. Furthermore, the residual connections and channel attention mechanisms in the hierarchical feature encoder enhance the network's robustness to noise and local blurring, eliminating the need for additional preprocessing for different imaging qualities. Therefore, this invention solves the problems in the prior art where it is difficult to balance global deformation capture and computational efficiency during image registration, error accumulation and insufficient local detail alignment in multi-scale deformation estimation, and poor robustness of feature extraction. Attached Figure Description

[0017] Figure 1 This is a flowchart of the MRI image registration method in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the MRI image registration system in the second embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device in the third embodiment of the present invention; Figure 4 This is a structural diagram of a hierarchical feature encoder according to an embodiment of the present invention; Figure 5 This is a structural diagram of the window and channel self-attention module in one embodiment of the present invention, (a) is the window and channel self-attention module, and (b) is the channel attention module; Figure 6 This is a structural diagram of a dual-cable deformation fusion module according to an embodiment of the present invention; Figure 7 This is a network structure diagram of an MRI image registration method according to an embodiment of the present invention; Figure 8 The images show the registration results of the method in one embodiment of the present invention on the LPBA40 and Mindboggle datasets and the difference images before and after registration. (a)–(c) are the moving image, the fixed image, and the registration result after deformation (Warped), respectively. (d) and (e) are the difference heatmaps before registration (Moving–Fixed) and after registration (Warped–Fixed), respectively. Figure 9 The images show the visualization results of different methods in one embodiment of the present invention on the Mindboggle dataset, where (a) is a fixed image; (b) is a fixed image; (c) is the registration result of SyN; (d) is the registration result of VoxelMorph; (e) is the registration result of TransMorph; (f) is the registration result of Im2Grid; (g) is the registration result of PR++; (h) is the registration result of PIVIT; and (i) is the registration result of the algorithm in this embodiment. Figure 10 This is a TRE box plot of the maximum outgoing to maximum incoming phase of the method in one embodiment of the present invention on the Creatis dataset; Figure 11 This is a visual comparison of PIVIT in one embodiment of the present invention and the algorithm in this embodiment. a–c are moving images, d–f are fixed images; g–i are the registration results of the comparison method PIVIT, and j–l are the registration results of this method. Figure 12 The following is a comparison of the displacement field deformation mesh visualization results of different registration algorithms in one embodiment of the present invention on the Mindboggle dataset. (a) is the deformation mesh of VoxelMorph; (b) is the deformation mesh of TransMorph; (c) is the deformation mesh of PR++; (d) is the deformation mesh of Im2Grid; (e) is the deformation mesh of RDN; (f) is the deformation mesh of PIVIT; (g) is the deformation mesh of the algorithm in this embodiment. The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] Unless otherwise defined, 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. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] Example 1 Please see Figure 1 The figure shows an MRI image registration method in the first embodiment of the present invention, which specifically includes S01-S04.

[0021] S01, acquire the fixed image and floating image to be registered, and perform feature extraction on the fixed image and floating image to determine feature maps at multiple different resolution scales.

[0022] Specifically, the fixed image and the floating image are respectively input into the hierarchical feature encoder shared by the two streams, so that the input image is downsampled step by step by a residual block and multiple residual attention blocks cascaded in the hierarchical feature encoder, thereby obtaining multiple feature maps of different resolution scales output by the residual block and the residual attention block. In this process, the feature map size is reduced between adjacent blocks using average pooling. Within the residual blocks, local structural features are extracted through two layers of 3D convolution, instance normalization, the GELU (Gaussian Error Linear Unit) activation function, and residual connections. In the residual attention blocks, the residual block operations are performed first, followed by adaptive recalibration of the channel responses using a pre-defined channel attention module. More specifically, residual connections alleviate the gradient decay problem in deep network training, preserving key geometric details such as texture and edges. Simultaneously, a channel attention mechanism is introduced to adaptively weight feature channels, highlighting anatomical structural features more critical to the registration task (such as ventricular edges and cortical sulci), while suppressing noise and redundant information. Thus, even with local noise or blurring in the image, the encoder can still provide robust and discriminative feature representations.

[0023] In specific implementation, such as Figure 4 As shown, this method designs a hierarchical feature encoder for fixed images. With floating images Multi-scale feature extraction is performed. The encoder employs a dual-stream shared architecture, consisting of a cascaded residual block (Rc block) and three residual attention blocks (RcRa blocks). Inter-layer downsampling is achieved through average pooling (AvgPool), forming a feature pyramid from high-resolution details to low-resolution semantics. Given the input volume data, the encoder sequentially outputs feature pairs at four scales, representing the features of the fixed image. Features of floating images The output feature sizes are respectively the input and output feature sizes. The number of channels is denoted as C.

[0024] This design aims to compensate for the loss of detail caused by the insufficient modeling of long-range dependencies and weak cross-scale consistency in pure CNN encoders. Specifically, the Rc block (Residual Convolution Block) employs two layers. Convolutional processing, combined with Instance Normalization and GELU activation, models the local structure. Residual connections provide direct information pathways, mitigating gradient decay and network degradation in deep feature extraction and helping to preserve geometric details such as edges and textures during multi-scale extraction. The Attention Convolution Block (AC block) introduces a Squeeze-and-Excitation (SE) mechanism on top of the convolutional features, adaptively recalibrating channel responses to highlight anatomical features more critical for registration and suppress redundant information. Through this design, the encoder can obtain more discriminative multi-scale features at a lower computational cost, providing stable input for subsequent coarse-to-fine deformation estimation.

[0025] S02, based on the attention mechanism, the fixed image feature map and the floating image feature map at the lowest resolution scale are processed to predict the initial deformation field, and the floating image feature map at the current scale is transformed and coarsely aligned using the initial deformation field to obtain the floating feature map.

[0026] Specifically, the fixed image feature map and the floating image feature map at the lowest resolution scale are concatenated along the channel dimension and then input into the window channel self-attention module. In the window channel self-attention module, the current input feature map is layer-normalized, and then window self-attention operations and shift window self-attention operations are executed alternately. Simultaneously with each self-attention operation, a channel attention operation is performed on the same normalized feature map. The results of the self-attention operation and the channel attention operation are added element-wise to obtain an intermediate feature map. The intermediate feature map obtained in the previous round is used as the new current input feature map, and the above steps for determining the intermediate feature map are repeated at least once. The intermediate feature map obtained in the last round is used as the initial deformation field. The window channel self-attention module does not include position encoding and employs a single-head self-attention mechanism. More specifically, the window self-attention mechanism is used to efficiently model the spatial dependencies between distant voxels in the low-resolution feature space, while channel attention suppresses redundant channel responses to enhance the stability of deformation regression. By deploying a lightweight Transformer module only at a low-resolution scale, expensive global computations on high-resolution feature maps are avoided, significantly reducing memory usage and inference latency. Furthermore, the generated initial deformation field effectively compensates for the overall offset and rotation of the floating image, providing a good initial state for subsequent fine alignment at high-resolution scales and preventing error accumulation due to excessive initial deviations. This allows for the capture of large-scale global deformations in medical images with relatively low computational overhead.

[0027] In practical implementation, because 3D medical image registration requires the network to establish a reliable correspondence between features of fixed and deformed images to regress the deformation field, this necessitates capturing long-distance spatial dependencies across regions to adapt to large deformations, while also strictly controlling computational load and memory usage under the high dimensionality and high resolution conditions of 3D volumetric data. Although global self-attention can directly model long-range relationships between arbitrary tokens (feature units, usually referring to representations obtained from image patches or local feature mappings), its computational and storage complexity increases quadratically with the number of tokens, making it too costly in dense 3D prediction scenarios. Furthermore, relying solely on spatial attention easily overlooks differences in channel importance, leading to the accumulation of redundant responses and affecting the stability of deformation regression.

[0028] Based on this, this method deploys the Window Channel-Self Attention (WCSA) module in the low-resolution registration stage, such as... Figure 5As shown, this module deforms the corresponding voxel matching relationships between feature maps into the spatial neighborhood, thus eliminating the need to use Transformer on large-scale feature maps at high scales. Furthermore, positional encoding is removed, only single-head self-attention is used, and the number of channels is reduced. These operations speed up the process and significantly reduce the number of model parameters. Simultaneously, WCSA employs a cascade of Window Self-Attention (W-SA) and Shifted Window Self-Attention (SW-SA) for spatial dependency modeling: W-SA captures fine-grained matching relationships within a local window, while SW-SA establishes cross-window connections through a shifting strategy to facilitate long-distance information propagation. At the same time, the module introduces Channel Attention (CA) in parallel at each attention update stage to adaptively recalibrate channel responses, suppressing redundant channels and strengthening key structural representations related to deformation estimation, thereby improving the discriminativity and robustness of feature representation while maintaining efficiency.

[0029] The CA module compresses the global spatial information of the channel dimension into channel statistics and generates gating coefficients to highlight key channels and suppress redundant responses, such as... Figure 5 As shown in (b). Given input features CA can be expressed as:

[0030]

[0031] in Indicates the input feature map, This represents the channel gating weights generated by the channel attention module. This indicates the output of the channel attention module. This is global pooling, used to aggregate global context information across spatial dimensions; It is sigmoid. Describe element-wise multiplication. Corresponding to a lightweight mapping consisting of several convolutions and GELU, this residual gating form can achieve effective channel recalibration while maintaining stable training. Meanwhile, let the... The fixed and deformation characteristics at each scale are respectively and The input is The calculation process of WCSA can be expressed as follows:

[0032]

[0033]

[0034]

[0035]

[0036] in, Indicates the first Fixed image features at each scale and deformed image features The concatenated input features and This represents the intermediate features after attention update. This represents the output features after further mapping by a multilayer perceptron. Layer Normalization is represented by W-SA, which stands for Window Self-Attention and is used to model spatial dependencies within a local window. This refers to shifted window self-attention, used to establish cross-window information interaction; Channel attention is used for adaptive recalibration of channel responses. These are the weight coefficients of the channel attention branch, used to adjust their impact on feature optimization and representation enhancement; This refers to a multi-layer perceptron, used for nonlinear feature transformation; Indicates the first Deformation field predicted at each scale; express 3D convolution is used to map features to the deformation field parameter space. The in-window self-attention can be represented as:

[0037] in Obtained through a linear mapping, the values ​​are query, key, and value, respectively. The dimensions are query and key. For relative position offset; For patch size; This represents the number of patches in the 3D feature. The relative position range is... ,in Then the bias matrix Perform parameter optimization, matrix The value is determined by Obtained by sampling.

[0038] S03, the initial deformation field is upsampled stepwise from the lowest resolution scale to the highest resolution scale. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. Then, the residual deformation field is predicted by fusing the fixed image feature map at the current scale and the deformed floating feature map through dual-clue fusion of similarity matching cues and edge structure cues. The current upsampled deformation field is superimposed with the residual deformation field to obtain the cumulative deformation field at the current scale. This cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained.

[0039] In practice, the low-resolution initial deformation field is progressively upsampled as the initial estimate for the current scale. Then, using fixed image features at the current scale and floating image features transformed from the initial estimate, a dual-cue fusion module predicts the missing residual deformation field at the current scale. The upsampled deformation and residual deformation are then superimposed to obtain the cumulative deformation field, which is then passed to the next higher resolution. This decomposes the global deformation estimation into a low-resolution coarse estimate and a high-resolution detail compensation, effectively cutting off the error propagation path and significantly improving the alignment accuracy of fine anatomical structures and edges such as sulci and gyri, achieving precise characterization of local details. This solves the problems of progressive error accumulation in multi-scale deformation estimation and insufficient alignment of local details and blurred edge structures at the high-resolution stage.

[0040] Specifically, the similarity matching branch calculates the local window similarity between the fixed image feature map at the current scale and the transformed floating image feature map, and generates the first sub-deformation field based on a similarity-weighted predefined relative displacement grid. The edge deformation branch uses edge convolution to extract the structural boundary features of the fixed image feature map at the current scale and the transformed floating image feature map, and regresses to generate the second sub-deformation field. The first and second sub-deformation fields are concatenated along the channel dimension and then convolved to generate a velocity field. The velocity field is scaled and integrally squared to output a residual deformation field that satisfies the homeomorphism constraint. In practice, by introducing two complementary deformation cues—local similarity matching and edge structure preservation—the residual deformation field at the current scale is accurately predicted while ensuring the physical rationality of the deformation. This dual-cue fusion mechanism significantly enhances the modeling ability for complex local deformations. The similarity branch improves the matching stability in sparse texture regions, the edge branch effectively combats boundary blurring caused by multi-level deformation accumulation, and the homeomorphism constraint ensures the smoothness, reversibility, and fold-free properties of the deformation field.

[0041] Furthermore, the steps for generating the first sub-deformation field through the similarity matching branch include: for each voxel position, calculating the dot product similarity between the feature vectors of the fixed image feature map and the transformed floating image feature map within a preset local window; applying the flexible maximum function in the window dimension to obtain the matching weight matrix; and multiplying the matching weight matrix with a predefined relative displacement grid to obtain the first sub-deformation field.

[0042] Furthermore, the steps for generating the second sub-deformation field through edge deformation branches include: extracting edge enhancement features from the fixed image feature map and the transformed floating image feature map using edge convolution; concatenating the two sets of edge enhancement features in the channel dimension and then performing three-dimensional convolution regression to obtain the second sub-deformation field.

[0043] More specifically, the similarity matching branch explicitly calculates the local correlation between feature vectors, providing direct matching constraints for fine-grained displacements and solving the problem of pure regression networks being insensitive to local details. Furthermore, the edge deformation branch utilizes edge convolution to enhance the information of organ contours and structural boundaries, effectively preventing the blurring or degradation of anatomical boundaries during registration. In addition, scaling and square integration of the velocity field generates a residual deformation field that satisfies homeomorphic constraints, ensuring the topological preservation of the deformation field. This allows the progressively stacked residual deformation to effectively suppress error accumulation in multi-scale estimation. The resulting highest-resolution deformation field achieves accurate alignment of the overall contour while maintaining the clarity and sharpness of minute structures such as sulci and gyri. Thus, a coarse-to-fine pyramid strategy compensates for subtle local deformations missed in low-resolution estimation.

[0044] In practical implementation, the highest resolution deformation field is determined through a dual-clue deformation fusion module. Since this module is geared towards the high-resolution stage of pyramid decoding, it aims to alleviate the problems of insufficient learning of local deformation details and easy degradation of edge structures. Its core idea is to introduce two types of complementary cues: on the one hand, through a similarity matching branch on fixed features... With movement characteristics Explicitly establishing local correspondences between the two branches provides more direct matching constraints for fine-grained displacement estimation. On the other hand, the EdgeConv branch enhances the representation of structural boundaries and local geometric details through edge deformation, improving the alignment capability with contours and interface regions. The outputs of the two branches are concatenated along the channel dimension, then fused through convolution to generate a velocity field. This is further processed by a differential homeomorphic transformation (Diff) layer to generate a smooth and approximately reversible deformation field, thereby improving local registration accuracy and structural consistency at high resolution while ensuring end-to-end training stability. Figure 6 As shown in the figure. Similarity matching in the figure is performed at the [number]th [time]. Multi-scale features of fixed and moving images at various scales , Local window similarity is used to estimate the sub-deformation field. Specifically, for each voxel location, fixed and floating features are used... The candidate features within the local window are subjected to dot product similarity calculation, and Softmax is applied along the window dimension to obtain the matching weight matrix:

[0045] in This indicates that mobile features are locally unfolded to construct a candidate set. Subsequently, the matching weights are compared with a predefined relative displacement grid. Multiplying them together yields the sub-deformation field at the corresponding scale:

[0046] This process is equivalent to performing differentiable soft matching within a local search window, achieving dense displacement regression through weighted expectation of candidate displacements, thereby explicitly establishing cross-image local correspondences and improving the stability of deformation estimation. After obtaining the fused features, the module... Convolutional Regression Three-Channel Velocity Field To ensure the geometric rationality of the deformation, Diff integrates the velocity field into a deformation field that satisfies the homeomorphic constraint:

[0047] in , The target deformation is determined. This embodiment employs scaling-and-squaring to numerically integrate the above equations: first, the velocity field is scaled to... To obtain the corresponding small deformation, the deformation self-composite is repeatedly performed to achieve square iteration, and finally the deformation field at this scale is obtained. The above design enables the deformation field to have better smoothness and reversibility under the condition of end-to-end differentiability, thereby enhancing the stability and structural fidelity of high-resolution stage registration.

[0048] S04 applies the highest resolution deformation field to the floating image to generate a registered image.

[0049] Furthermore, the method employs an unsupervised loss function for training, which includes an image similarity metric and a deformation field smoothing regularization term. Specifically, using an end-to-end unsupervised training strategy, the network optimizes its parameters by maximizing image similarity and constraining deformation field smoothness without requiring manual annotation of the gold standard deformation field. This ensures that the output registered image maintains a high degree of consistency with the fixed image in terms of grayscale distribution and anatomical structure. Simultaneously, due to the regularization term, the deformation field maintains physiological interpretability and smoothness, avoiding non-physical deformation phenomena such as mesh folding.

[0050] In practical implementation, to train the registration network proposed in this method, this scheme adopts a composite loss function for unsupervised learning, which consists of an image similarity metric and a deformation field regularization term, and its form is as follows:

[0051] in, and These represent fixed and floating images, respectively. To predict the deformation field, This represents a floating image after deformation. This represents the regularization loss weight. This represents the similarity metric. This represents the regularization term.

[0052] In this method, the loss function on the brain dataset is set to locally normalized cross-correlation, which is defined as follows:

[0053] in, Represents the entire volume domain, voxel Therefore, the body element Centered on, size is voxels in the local neighborhood of . , and These represent the average voxel values ​​within the local neighborhood.

[0054] In the generalization experiment on the lung dataset, the similarity metric used was Normalized Cross-Correlation (NCC), and the regularization metric used was NCC to measure the global correlation between stationary and deformed images, defined as follows:

[0055] in, Represents the entire volume domain. Voxel position, and They represent in The average voxel value on the surface. When using NCC as the loss function, to achieve consistent optimization that maximizes cross-correlation, the range of NCC values ​​is [value missing]. The closer the value is to 1, the more similar the two images are. To use this value to minimize the optimization objective, its negative value is usually taken as the similarity loss during training. This scheme defines:

[0056] Through the aforementioned composite loss, the network simultaneously optimizes the image alignment and the smoothness of the deformation field under unsupervised conditions, thereby obtaining stable and geometrically reasonable registration results.

[0057] As an example, and not a limitation, in some alternative embodiments, the overall structure of this solution is as follows: Figure 7 As shown, this scheme employs a coarse-to-fine pyramid residual deformation estimation strategy, consisting of a hierarchical feature encoder and a coarse-to-fine registration framework. First, the encoder processes the input... This involves extracting multi-scale features, where the scale is downsampled progressively from high resolution to low resolution. At the lowest resolution layer... Above, the Window Channel Self-Attention Module (WCSA) models cross-graph space based on channel saliency to predict the initial deformation field. The network then performs deformation on the floating image features to achieve coarse alignment. Following this, it enters the fine registration stage, where the network employs a dual-clip deformation fusion module (DCDF) to progressively refine the deformation at higher resolution scales. Specifically, in the first... Each scale (corresponding to coarse to fine) First, the deformation field from the previous scale is upsampled to the current resolution to guide the deformation of the floating features. Then, DCDF predicts the residual deformation at this scale. The cumulative deformation is obtained by adding it element-wise to the upsampled deformation:

[0058] in, Indicates upsampling, This indicates element-wise addition, ultimately achieving the desired result at full resolution. Convolution is used to perform channel mapping to obtain dense deformation fields. This coarse-to-fine residual refinement strategy can prioritize capturing large-scale deformations at low resolution and gradually compensate for local details at high resolution, thereby improving registration accuracy and stability.

[0059] As an example, and not a limitation, in some optional embodiments, to evaluate the effectiveness and generalization ability of the proposed algorithm, this embodiment selects three publicly available datasets for experiments, including two brain MRI datasets, LPBA40 and Mindboggle, and one lung CT dataset, Creatis. The brain dataset is used to verify the algorithm's registration performance under the same modality and similar anatomical structures; the lung dataset is used to examine the model's cross-task generalization ability under different anatomical structures and imaging modalities.

[0060] Specifically, all brain scan images were first subjected to intensity normalization, and the FreeSurfer method was used to perform decranialization to remove interference from non-brain tissue. Subsequently, all brain image voxels were uniformly resampled to... The image size was adjusted to 160×192×160 by center cropping to ensure consistency in spatial resolution and input size for different samples, which facilitates network training and fair comparison.

[0061] To further verify the generalization ability of the algorithm, this embodiment introduces the Creatis lung CT dataset for supplementary experiments. This dataset differs significantly from brain MRI in both anatomical structure (lung respiratory motion) and imaging modality (CT), enabling a more comprehensive evaluation of the model's applicability across different registration tasks. The Creatis dataset comes from the Léon Bérard Cancer Center in Lyon, France, and contains 4D-CT lung data from six cases, each consisting of CT images of 10 respiratory phases; the images of the maximum inhalation and maximum exhalation phases are annotated with approximately 10 keypoints, which can be used to evaluate metrics such as keypoint error.

[0062] The preprocessing steps for the Creatis dataset are as follows: First, a region growing segmentation algorithm is used to remove irrelevant information such as the thoracic cavity contour, bones, and the shooting background, retaining only the lung parenchyma region to reduce the interference of non-target structures on registration; then, a threshold is set according to the intensity distribution of the CT images to suppress noise and clean up abnormal voxels; finally, all images are uniformly resampled to a fixed size of 288×240×288 and intensity normalization is performed to ensure that the scale and intensity range of the input data are consistent.

[0063] The experimental environment and training settings for this method are as follows: For the software environment, the brain dataset experiments used the PyTorch deep learning framework, Python 3.9 programming language, and Windows 11 operating system; the lung dataset experiments used the PyTorch deep learning framework, Python 3.10 programming language, and Ubuntu 22.04 LTS operating system. For the hardware environment, the brain experiments ran on an NVIDIA GeForce RTX 3090 graphics card with an AMD Ryzen 5 5600 CPU; the lung experiments ran on an NVIDIA A100 PCIe 40GB graphics card with an Intel Xeon Processor (Skylake, IBRS, 10 vCPU). The Adam optimizer was used during training, with a batch size of 1. For hyperparameter settings, the learning rate was set to 0.0001, and the regularization loss term weights were... Set to 0.5; the training epochs for both the LPBA40 and Mindboggle datasets are 30, while the training epochs for the Creatis dataset are 100.

[0064] To evaluate the effectiveness of the proposed LWCDF-Net in brain MRI image registration, experiments were conducted on the LPBA40, Mindboggle, and Creati datasets, and comparative analyses were performed with several representative registration methods. The comparison methods included the traditional nonlinear registration algorithm SyN, as well as deep learning-based registration algorithms VoxelMorph, TransMorph, PR++, Im2Grid, RDN, ModeT, and PIVIT. To ensure fairness, all methods were trained and evaluated using the same training and test sets. Tables 1 and 2 show the registration results of each method on the LPBA40 and Mindboggle datasets, respectively, with the best results marked by an underline.

[0065] Table 4.1 Quantitative analysis of different algorithms on the LPBA40 dataset

[0066] Table 2. Quantitative analysis of different algorithms on the Mindboggle dataset

[0067] As shown in Tables 1 and 2, our method achieved the best registration performance on both the LPBA40 and Mindboggle brain datasets. On LPBA40, our method achieved a Dice similarity coefficient (Dice) of 71.95±1.6, and average symmetric surface distance (ASSD) and 95th percentile Hausdorff distance (HD95) of 1.337±0.11 and 6.057±0.57, respectively. All three metrics were optimal, indicating that the model can achieve more accurate boundary alignment while improving the overall overlap of anatomical structures. On Mindboggle, this method also achieved the best results: Dice was 59.86±1.6, ASSD was 0.983±0.12, and HD95 was 5.919±0.56. Compared with the better-performing ModeT, Dice was improved by 0.45, and ASSD and HD95 were further reduced. The improvement in Dice indicates more complete overlap of anatomical structures after registration; the decrease in ASSD and HD95 indicates that the average boundary error and extreme error were effectively suppressed, demonstrating more accurate and robust deformation estimation. The consistent improvement in metrics on both datasets validates that LWCDF-Net has good generalization ability and robustness to different brain data distributions and anatomical differences.

[0068] To visually evaluate the effectiveness of the proposed method in brain MRI registration, registration results and subtraction difference maps before and after registration are displayed on the LPBA40 and Mindboggle datasets. These differences are obtained by calculating the absolute difference in pixel values ​​between images and presented as heatmaps. Figure 8 As shown, the upper part is LPBA40, and the lower part is Mindboggle; from top to bottom, they are sagittal, coronal, and transverse sections. Figure 8 In the middle (a)–(c), the images are the moving image, the fixed image, and the registration result after deformation, respectively. Figure 8 (d) and Figure 8 The image in (e) shows the difference heatmaps before (Moving–Fixed) and after (Warped–Fixed). In the heatmap, darker colors indicate smaller differences and better alignment; brighter colors indicate larger differences and the presence of local mismatches.

[0069] From a visual perspective, Figure 8 (c) and Figure 8In (b), the overall contour and key anatomical structures are more consistent, indicating that this method can revert to reasonable deformations and achieve good global and local alignment. Combined with the difference diagram, it can be seen that before registration in LPBA40... Figure 8 In the middle (d) area, there is a large area of ​​highlighted region between the outer contour of the brain and the cortical boundary, indicating a significant mismatch; after registration Figure 8 In the middle (e), the highlight range and intensity are significantly reduced, indicating that the overlap between the deformed image and the fixed image increases and the absolute difference decreases overall. The same trend is also observed on Mindboggle: the difference heatmap after registration becomes darker overall, with only a few edge regions still showing differences, indicating that the model still has stable registration ability under complex structures.

[0070] like Figure 9 As shown, a comparative analysis of the registration visualization results on the Mindboggle dataset was conducted. Figure 9 The first to third rows correspond to the registration results in the sagittal, coronal, and transverse planes, respectively. Pink boxes are used to highlight and enlarge the white matter contour, choroid plexus, and septum pellucidum regions to emphasize the alignment differences between different methods at subtle anatomical structures. Overall, the traditional method SyN tends to exhibit uneven deformation or insufficient detail alignment near complex local structures; boundary misalignment and texture blurring are still observable in magnified areas. Deep learning methods improve overall alignment, but some methods still exhibit oversmoothing at small structures, resulting in impaired local boundary clarity. In contrast, the method proposed in this embodiment generates deformed images that are more consistent with the fixed image in terms of global structure and local details, maintaining high alignment quality even in areas with significant anatomical differences and complex structural boundaries. Especially in the magnified white matter contour, choroid plexus, and septum pellucidum regions, the proposed method shows higher contour fit and structural morphology closer to the fixed image, indicating stronger modeling and recovery capabilities for local anatomical details. This qualitative result is consistent with the aforementioned quantitative evaluation, further validating the advantages of the proposed method in registration accuracy and structural preservation.

[0071] To verify the effectiveness of the improved modules proposed in this embodiment, ablation experiments were conducted on the LPBA40 and Mindboggle brain datasets. The specific settings are as follows: Baseline is the baseline network; Method1 represents the introduction of a hierarchical feature encoder on top of the Baseline; Method2 further adds a WCSA module to Method1; Method3 is the complete network of this embodiment, i.e., adding a DCDF module to Method2. The ablation results are shown in Tables 3 and 4, respectively.

[0072] Table 3 Ablation experimental results on the LPBA40 dataset

[0073] Table 4 Ablation experimental results on the Mindboggle dataset

[0074] As shown in Tables 3 and 4, with the gradual introduction of improved modules, the registration performance of the model on both the LPBA40 and Mindboggle datasets has been steadily improved, demonstrating the effective gains of each module to the overall framework. Compared to the baseline, Method1's Dice improved by 0.38 and 0.39 percentage points, respectively, while ASSD and HD95 decreased simultaneously, indicating that the introduced hierarchical feature encoder can more fully model multi-scale features and enhance fine-grained representation capabilities. Building on this, Method2, which incorporates WCSA, further improved Dice by 0.48 and 1.98 percentage points relative to Method1, indicating that this module helps capture long-range dependencies in low-resolution feature spaces, thereby improving the ability to model global correspondences in large deformation scenarios. Ultimately, Method3, which integrates all modules, improved the Dice score of Method2 by 0.58 and 1.67 percentage points on both datasets, and achieved the best results on the ASSD and HD95 metrics. This indicates that the DCDF fusion module can further enhance local structure alignment and effectively suppress residual errors at the boundaries, thereby achieving a comprehensive improvement in both accuracy and robustness.

[0075] This embodiment uses the publicly available Creatis lung registration dataset for generalization experiments. The experiments compared several classic deep learning registration algorithms. In Table 5, the registration performance data for RPM, D-FCN, and GraphRegNet algorithms are directly cited from their respective literatures, while the results for the PIVIT algorithm are reproduced from this embodiment to ensure the reproducibility and fairness of the comparison. Quantitative comparisons of different methods on this dataset allow for a comprehensive evaluation of the model's generalization ability. Experimental results show that this method achieves the best performance among all compared algorithms, with a mean and standard deviation of 0.98 mm and 0.69 mm for the target registration error (TRE), respectively, representing further improvements compared to RPM, D-FCN, GraphRegNet, and PIVIT. These experimental results demonstrate that this method not only achieves high accuracy in brain MRI registration tasks but also maintains good registration performance and generalization ability in independent tests on the lung dataset.

[0076] Table 5. TRE (mm) of different algorithms on the Creatis dataset

[0077] Figure 10Box plots of the TREs (Transmission Relationships) at approximately 100 expert-annotated points for registration from the maximum exhalation phase to the maximum inhalation phase in six cases from the Creati dataset are presented. It can be observed that the median TREs for each case are generally similar, and the box sizes do not differ significantly, indicating that the model's error distribution exhibits good consistency and stability across different cases. However, a small number of outliers (marked in red) are present in some cases, suggesting that significant registration errors may still occur in areas of large local deformation or structural boundaries. Overall, these results demonstrate that the proposed method is robust to lung data and maintains a low level of registration error at most annotated points.

[0078] Meanwhile, a visual comparison of the registration results is provided on case 1 of the Creatis dataset, such as... Figure 11 As shown. Figure 11 In the image, a–c represents the moving images. Figure 11 In the image, d–f represents a fixed image; Figure 11 In the figure, g–i represents the registration result of the comparison method PIVIT. Figure 11 The values ​​j–l in the table represent the registration results of this method. Each row, from top to bottom, corresponds to axial, coronal, and sagittal slices of the lung image, respectively. The red rectangles and their magnified views are used to highlight structural alignment differences in local detail areas.

[0079] from Figure 11 As can be seen, the deformed image and the fixed image obtained after registration are structurally consistent. Compared with PIVIT, the method in this embodiment shows better consistency in the position matching of small anatomical structures such as blood vessels and airways within the magnified area, better edge continuity, less local blurring and misalignment, stronger ability to depict local deformation, and better stability. Figure 11 It can be seen that the method of this embodiment can achieve a relatively good structural alignment effect in the lung registration task.

[0080] To further evaluate the trade-off between model complexity and deformation field topology preservation of the proposed method, this embodiment statistically analyzes the parameter count of different methods and the proportion of voxels with negative Jacobian determinants in the deformation field on the LPBA40 dataset. This reflects the degree of folding in the deformation field; a lower proportion indicates that the deformation becomes closer to a differential homeomorphism, and the stronger the topology preservation ability. As shown in Table 6, Transformer-based methods typically have a high parameter scale; for example, TransMorph has 46.557M parameters. for The folding ratio of PR++ is significantly higher than that of other methods; the folding ratio of PR++ is also [missing information]. This indicates that non-physical folding is more likely to occur when modeling complex deformations. In contrast, convolutional or hybrid structure methods have a lower overall folding ratio, such as RDN. for ModeT and PIVIT are respectively and This demonstrates better topology constraint performance. The method in this embodiment exhibits strong topology preservation capabilities while maintaining a low parameter count: with only 1.155M parameters, Down to It significantly outperforms methods such as TransMorph, PR++, Im2Grid, RDN, ModeT, and PIVIT.

[0081] Table 6. Comparative Analysis of Lightweight Registration Algorithms and Topology Preservation Capabilities on the LPBA40 Dataset

[0082] To visually compare the smoothness and topology preservation of the deformation field, Figure 12 Tables (a)-(g) illustrate the displacement field deformation mesh on Mindboggle. Compared to some methods that exhibit mesh compression / distortion or even potential intersections in local regions, the mesh deformation of this method is continuous, smooth, and more regular, and is consistent with... The statistical results are consistent, indicating that the deformation is more stable and has better topological consistency. All results show that the method in this embodiment can effectively suppress deformation folding and generate a smoother and topologically consistent deformation field while maintaining a relatively lightweight approach, thereby further improving the interpretability and stability of the deformation in addition to registration accuracy.

[0083] To comprehensively evaluate the performance of the method in this embodiment in terms of real-time performance and computational complexity, this embodiment reproduced and compared various registration algorithms under the same experimental environment, including SyN, VoxelMorph, TransMorph, PR++, Im2Grid, RDN, and PIVIT. Table 7 shows the average registration time of different algorithms on two brain datasets. The results show that the traditional SyN algorithm exhibits superior accuracy in different registration tasks, but its computational efficiency is relatively low, especially in registration tasks with larger datasets, where the average efficiency reaches 202.861s and 226.143s, which is difficult to meet real-time requirements. In contrast, the deep learning-based VoxelMorph can directly generate the final deformation field by relying on a simple end-to-end forward inference structure, with average registration times of 0.173s and 0.171s, respectively, showing the best real-time performance. However, due to the simple network structure, the registration accuracy is poor, and the Dice values ​​are 7.84 and 5.90 percentage points lower than the algorithm in this embodiment, respectively. The PIVIT network primarily processes deformation information at low resolution, offering short registration time but poor accuracy, with Dice values ​​1.44 and 4.04 percentage points lower than the algorithm in this embodiment, respectively. The algorithm in this embodiment uses a lightweight window channel self-attention module to capture long-range spatial dependencies on the baseline structure, maintaining registration accuracy while reducing registration time. The average registration time on the two datasets is 0.275s and 0.327s, both less than 1 second, significantly faster than traditional algorithms and exhibiting good real-time performance in deep learning algorithms. These results demonstrate that the algorithm in this embodiment improves registration accuracy while maintaining a relatively low registration time, exhibiting good real-time performance.

[0084] Table 7. Average registration time of different registration algorithms on two brain datasets.

[0085] In summary, this invention extracts multi-scale features by designing a lightweight hierarchical feature encoder. At the lowest resolution scale, a window channel self-attention module captures global deformation at a low computational cost. Subsequently, it upsamples progressively to higher resolutions, predicting residual deformation at each level through similarity matching and edge structure dual-cue fusion. The upsampled deformation is then superimposed with the residual deformation to achieve progressive error compensation. This scheme uses the window self-attention mechanism only at the low-resolution stage, eliminating the need to deploy a highly complex Transformer module on the full-resolution feature map, significantly reducing the number of parameters and computational overhead. Simultaneously, dual-cue fusion explicitly strengthens local matching and edge constraints in the fine registration stage, effectively suppressing multi-scale error accumulation and maintaining the clarity of anatomical boundaries without relying on large amounts of labeled data or repeated iterative optimization. Furthermore, the residual connections and channel attention mechanisms in the hierarchical feature encoder enhance the network's robustness to noise and local blurring, eliminating the need for additional preprocessing for different imaging qualities. Therefore, this invention solves the problems in the prior art where it is difficult to balance global deformation capture and computational efficiency during image registration, error accumulation and insufficient local detail alignment in multi-scale deformation estimation, and poor robustness of feature extraction.

[0086] Example 2 Please see Figure 2 The diagram shows a structural block diagram of the MRI image registration system proposed in the second embodiment of the present invention. The MRI image registration system 200 includes: a feature extraction module 21, a coarse alignment module 22, a step-by-step processing module 23, and a registration module 24, wherein: Feature extraction module 21 is used to acquire the fixed image and the floating image to be registered, and to perform feature extraction on the fixed image and the floating image to determine feature maps at multiple different resolution scales; The coarse alignment module 22 is used to process the fixed image feature map and the floating image feature map at the lowest resolution scale based on the attention mechanism to predict the initial deformation field, and to transform the floating image feature map at the current scale using the initial deformation field to obtain the coarsely aligned floating feature map at that scale. The step-by-step processing module 23 is used to upsample the initial deformation field from the lowest resolution scale to the highest resolution scale step by step. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. The deformed floating feature map at the lowest resolution scale is the coarsely aligned floating feature map. Then, the fixed image feature map at the current scale and the deformed floating feature map are used to predict the residual deformation field through dual-clue fusion of similarity matching cues and edge structure cues. The current upsampled deformation field is superimposed with the residual deformation field to obtain the cumulative deformation field at the current scale. The cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained. The registration module 24 is used to apply the highest resolution deformation field to the floating image to generate a registered image.

[0087] Example 3 In another aspect, the present invention also proposes an electronic device, please refer to [link to relevant documentation]. Figure 3 The diagram shows an electronic device according to the third embodiment of the present invention, including a memory 20, a processor 10, and a computer program 30 stored in the memory and executable on the processor. When the processor 10 executes the computer program 30, it implements the MRI image registration method as described above.

[0088] In some embodiments, the processor 10 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 20 or process data, such as executing access restriction programs.

[0089] The memory 20 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of an electronic device, such as the hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 20 can include both internal and external storage units of the electronic device. The memory 20 can be used not only to store application software and various types of data of the electronic device, but also to temporarily store data that has been output or will be output.

[0090] It should be pointed out that, Figure 3 The structure shown does not constitute a limitation on the electronic device. In other embodiments, the electronic device may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0091] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the MRI image registration method described above.

[0092] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0093] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0094] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0095] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0096] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. An MRI image registration method, characterized in that, The method includes: Obtain a fixed image and a floating image to be registered, and perform feature extraction on the fixed image and the floating image to determine feature maps at multiple different resolution scales; The attention mechanism is used to process the fixed image feature map and the floating image feature map at the lowest resolution scale to predict the initial deformation field. The initial deformation field is then used to transform and coarsely align the floating image feature map at the current scale to obtain the floating feature map. The initial deformation field is upsampled stepwise from the lowest resolution scale to the highest resolution scale. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. Then, the residual deformation field is predicted by fusing the fixed image feature map at the current scale and the deformed floating feature map through dual-clue fusion of similarity matching cues and edge structure cues. The current upsampled deformation field is superimposed with the residual deformation field to obtain the cumulative deformation field at the current scale. This cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained. The highest resolution deformation field is applied to the floating image to generate a registered image.

2. The MRI image registration method according to claim 1, characterized in that, The steps of extracting features from the fixed image and the floating image to determine feature maps at multiple different resolution scales include: The fixed image and the floating image are respectively input into a dual-stream shared hierarchical feature encoder, so that the input image is downsampled step by step by a residual block and multiple residual attention blocks cascaded in the hierarchical feature encoder, thereby obtaining multiple feature maps of different resolution scales output by the residual block and the residual attention block; In this process, the feature map size is reduced between adjacent blocks using average pooling; In the residual block, local structural features are extracted through two layers of 3D convolution, instance normalization, GELU activation function, and residual connections; In the residual attention block, the operation of the residual block is executed first, and then the channel response is adaptively recalibrated through the preset channel attention module.

3. The MRI image registration method according to claim 1, characterized in that, The steps for processing fixed and floating image feature maps at the lowest resolution scale based on an attention mechanism to predict the initial deformation field include: The fixed image feature map and the floating image feature map at the lowest resolution scale are concatenated along the channel dimension and then input into the window channel self-attention module; In the window channel self-attention module, the current input feature map is normalized by layer, and then the window self-attention operation and the shift window self-attention operation are executed alternately. At the same time as each self-attention operation is executed, the channel attention operation is executed on the same normalized feature map. The result of the self-attention operation and the result of the channel attention operation are added element by element to obtain the intermediate feature map. The intermediate feature map obtained in the previous round is used as the new current input feature map. The above steps for determining the intermediate feature map are repeated at least once, and the intermediate feature map obtained in the last round is used as the initial deformation field. The window channel self-attention module does not include position encoding and adopts a single-head self-attention mechanism.

4. The MRI image registration method according to claim 1, characterized in that, The steps for predicting residual deformation fields using dual-clue fusion of similarity matching cues and edge structure cues include: The local window similarity between the fixed image feature map at the current scale and the transformed floating image feature map is calculated through a similarity matching branch, and the first sub-deformation field is generated based on the similarity weighted by a predefined relative displacement grid. By using edge deformation branching and edge convolution to extract structural boundary features of the fixed image feature map at the current scale and the transformed floating image feature map, the second sub-deformation field is generated through regression. The first sub-deformation field and the second sub-deformation field are concatenated and fused in the channel dimension to generate a velocity field. The velocity field is scaled and integraled squared to output a residual deformation field that satisfies the homeomorphic constraint.

5. The MRI image registration method according to claim 4, characterized in that, The step of generating the first sub-deformation field through the similarity matching branch includes: For each voxel location, the dot product similarity is calculated between the feature vectors of the fixed image feature map and the transformed floating image feature map within a preset local window; The matching weight matrix is ​​obtained by applying a flexible maximum function along the window dimension; The first sub-deformation field is obtained by multiplying the matching weight matrix with the predefined relative displacement grid.

6. The MRI image registration method according to claim 4, characterized in that, The step of generating the second sub-deformation field through the edge deformation branch includes: Edge enhancement features are extracted from both the fixed image feature map and the transformed floating image feature map using edge convolution. The second sub-deformation field is obtained by concatenating the two sets of edge enhancement features in the channel dimension and then performing three-dimensional convolutional regression.

7. The MRI image registration method according to claim 1, characterized in that, The method is trained using an unsupervised loss function, which includes an image similarity metric and a deformation field smoothing regularization term.

8. An MRI image registration system, characterized in that, The system for implementing the MRI image registration method according to any one of claims 1 to 7, the system comprising: The feature extraction module is used to acquire the fixed image and the floating image to be registered, and to perform feature extraction on the fixed image and the floating image to determine feature maps at multiple different resolution scales; The coarse alignment module is used to process the fixed image feature map and the floating image feature map at the lowest resolution scale based on the attention mechanism to predict the initial deformation field, and to transform and coarsely align the floating image feature map at the current scale using the initial deformation field to obtain the floating feature map. The step-by-step processing module is used to upsample the initial deformation field from the lowest resolution scale to the highest resolution scale step by step. After each upsampling, the floating image feature map at the current scale is transformed using the current upsampled deformation field to obtain the deformed floating feature map at the current scale. Then, the fixed image feature map at the current scale and the deformed floating feature map are used to predict the residual deformation field by fusing similarity matching cues and edge structure cues. The current upsampled deformation field and the residual deformation field are superimposed to obtain the cumulative deformation field at the current scale. The cumulative deformation field is used as the input deformation field for the next upsampling stage until the highest resolution deformation field is obtained. The registration module is used to apply the highest resolution deformation field to the floating image to generate a registered image.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the MRI image registration method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the MRI image registration method as described in any one of claims 1-7.

Images