A pathological whole slide image registration method, device and medium

Abstract

The present application relates to a kind of pathological whole section image registration method, equipment and medium, comprising: to fixed image and mobile image down sampling obtains first image, and obtains first tissue area mask;According to first image and first tissue area mask, perform low-resolution rough rigidity registration;After first tissue area mask up sampling, construct whole section graph structure, adopt graph neural network to model deformation field, learn registration problem with deformation field and model as the optimization learning problem of graph neural network, using optimized graph neural network to perform high-resolution non-rigid registration, output image block level discrete node displacement;Interpolation is carried out to the output image block level discrete node displacement, and continuous deformation field is obtained;Using continuous deformation field to carry out spatial transformation to mobile image, obtain the registration result aligned with fixed image.Compared with prior art, the present application improves the stability and biological rationality of pathological whole section image non-rigid registration, and effectively improves the efficiency of medical image processing.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a method, device and medium for registering whole pathological slide images. Background Technology

[0002] Whole-slice images (WSIs) are high-resolution digital pathological images. They digitize pathological tissue sections using whole-slice digital scanning technology and are widely used in tumor pathology diagnosis, tissue structure analysis, and multimodal bioinformatics fusion research. Because tissue sections are inevitably affected by factors such as section deformation, staining differences, and scanning bias during preparation, significant differences often exist in the spatial location and morphological structure between different sections. Whole-slice image registration technology aligns the spatial information of multiple WSIs and is a key foundation for achieving three-dimensional reconstruction of continuous sections and integration of multi-stained section data.

[0003] WSI registration methods can be divided into traditional methods based on display optimization conditions and deep learning-based methods. Traditional registration methods can be roughly divided into the following three categories: (1) Feature-based methods: using SIFT operators to extract features from the image, and then matching key feature points to achieve image registration; (2) Intensity-based methods: relying on the similarity of the pixel intensity distribution of the image for registration; (3) Frequency domain-based methods: using Fourier transform and other methods to convert the image to the frequency domain and align the image in the frequency domain. With the continuous development of machine learning methods, deep learning-based registration methods have gradually become a research hotspot in the field of digital pathology. Compared with traditional registration methods, deep learning-based methods introduce deep learning models such as convolutional neural networks and transformers for image feature extraction and matching, which can better eliminate the differences generated in the experimental processes such as slide preparation, staining, and scanning, and correct complex local non-rigid deformations, and have better generalization ability.

[0004] Among existing deep learning registration strategies, the deformation field learning strategy aims to train and optimize the correspondence between two slices generated by the machine learning model. This method has achieved good results in the registration of medical images such as magnetic resonance MRI and computed tomography CT. However, deformation field registration techniques still have the following limitations in the application of whole-slice image registration: First, unlike MRI and CT images, the true deformation field of pathological sections is unobservable. Therefore, WSI registration typically employs unsupervised learning methods, using locally normalized cross-correlation or mutual information as a measure of image similarity. However, due to the extremely high pixel count in whole-section pathological images, calculating cross-correlation similarity during training slows down the process. Furthermore, pixel intensity-based similarity metrics are easily affected by differences in slide preparation, staining, and image scanning, leading to unstable training.

[0005] Second, existing registration methods using convolutional neural networks (CNN) or U-Net mainly learn appearance features and similarities on regular pixels or patch grids. Their adjacency relationships are implicitly determined by spatial location, ignoring the correlation between physical and biological tissue blocks, making it difficult to guarantee global consistency and the biological rationality of the registration results. Summary of the Invention

[0006] The purpose of this invention is to overcome the defects of the prior art by providing a method, device and medium for registering pathological whole slide images.

[0007] The objective of this invention can be achieved through the following technical solutions: According to a first aspect of the present invention, a method for registering whole pathological slide images is provided, comprising: S1, acquire the fixed image and the moving image to be registered, wherein the fixed image and the moving image are pathological whole slide images; S2, downsample the fixed image and the moving image respectively to obtain the corresponding first image, and obtain the first tissue region mask corresponding to the first image; S3. Based on the first image and the corresponding first tissue region mask, perform low-resolution coarse rigid registration to obtain the rigid transformation matrix at coarse resolution. S4, after mapping the first tissue region mask to the original resolution, a second tissue region mask is obtained, and the sliced ​​tissue region is extracted from the second tissue region mask; S5, multiple sampling points are generated at intervals within the sliced ​​tissue area. Image blocks of a fixed size are extracted from the initial moving image and the fixed image, with the sampling points as the center. The sampling points of each image block are defined as nodes, and the pixel positions of the sampling points are defined as node position information. Then, the rigid transformation matrix is ​​mapped to the original resolution and applied to the node coordinates of the moving image. The displacement vector between two adjacent nodes is defined as the edge feature, and the full slice image structure is constructed. S6, based on the full slice image structure, uses graph neural network for deformation field modeling, and models the deformation field learning and registration problem as an optimization learning problem of graph neural network, and uses the optimized graph neural network to output the image block-level discrete node displacement; S7 converts the output image block-level discrete node displacements into a continuous deformation field to achieve high-resolution non-rigid registration. The continuous deformation field is used to perform spatial transformation on the moving image to obtain a registration result aligned with the fixed image.

[0008] Preferably, the step of performing low-resolution coarse rigid registration based on the first image and the corresponding first tissue region mask to obtain a rigid transformation matrix specifically includes: The first image and the first tissue region mask are downsampled again, and the background region in the downsampled first image is removed based on the downsampled first tissue region mask to obtain the second image; Key feature points are extracted from the second image. Similarity matching is performed on the feature descriptors of the second image corresponding to the fixed image and the second image corresponding to the moving image. After screening the effective matching point pairs using a cross-validation strategy, the effective matching point pairs are estimated using a random sampling consensus algorithm. The rigid transformation matrix between the second image of the fixed image and the moving image is then calculated.

[0009] Preferably, obtaining the first tissue region mask corresponding to the first image through color threshold filtering specifically includes: Based on the brightness quantile threshold, the first image is divided into dark tissue regions and medium brightness tissue regions, and background regions with saturation below the set threshold are removed from both types of tissue regions. The two types of tissue regions filtered by saturation threshold are merged to obtain an initial binary mask. Then, elliptical structuring elements are used to perform morphological operations to transform the initial binary mask into a smooth and continuous first tissue region mask.

[0010] Preferably, the deformation field modeling based on the full-slice graph structure using a graph neural network is expressed as follows: , in, Parameterized by a graph neural network, information is propagated between neighboring nodes through a message passing mechanism; For image block-level discrete node displacement; It is a full slice diagram structure.

[0011] Preferably, the full slice structure , is represented as: , , in, For a set of nodes, Let be the set of edges. , These represent the sets of nodes in the moving image and the stationary image, respectively. These represent the sets of edges in the moving image and the fixed image, respectively; For the set of edges across slices, For the fixed image, the first 1 node For the moving image Each node, with its cross-slice edge, is only used for similarity calculations between the next two slices and does not participate in the information transmission process of the graph neural network.

[0012] Preferably, the node update process in a graph neural network specifically includes: Extracting depth feature vectors from image patches The depth feature vector and node position information are used together as the node input features. ; Input features to nodes Perform feature mapping to obtain the initial features of the nodes. ; During the t-th round of message passing, node Based on the node's own characteristics Intra-image neighbor node features and the spatial relationship between nodes Generate aggregated features , is represented as: ,in, For message functions; Represents nodes within a moving image The set of neighboring nodes; According to the data passed to the node Aggregation features With the characteristics of the node itself Update # Wheel node features , is represented as: ,in, For the first Round node update function, used to fuse node's own features with aggregated features; According to the updated version Wheel node features Generate the corresponding spatial displacement parameters .

[0013] Preferably, an unsupervised learning strategy is employed, combining feature similarity and smoothing regularization loss to train the graph neural network, wherein the total loss function is calculated as follows: , , , In the formula: The loss is the consistency loss of node features; To smooth out the loss of the regularization term; These are weight parameters; The set of edges within the moving image; For nodes Spatial predicted displacement; For nodes The corresponding node in the moving image; For nodes At the node corresponding to the fixed image; This is for deformation operations on moving images; To extract the depth feature vectors of nodes in a moving image; This is to extract the depth feature vectors of nodes in a fixed image.

[0014] Preferably, the output image block-level discrete node displacements are interpolated to obtain a continuous deformation field, specifically as follows: For each pixel Based on the corresponding bottom left corner node Corresponding displacement Top left node Corresponding displacement bottom right node Corresponding displacement Top right node Corresponding displacement Pixel points are calculated using bilinear interpolation. The corresponding displacement is calculated using the following expression: ,in, , ; Pixels calculated based on bilinear interpolation Corresponding displacement The continuous deformation field was calculated. The expression is: ,in, Represents the pixels in the source image The mapping position in the target image.

[0015] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement any of the methods described above.

[0016] According to a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0017] Compared with the prior art, the present invention has the following advantages: 1) This invention proposes for the first time a deformation field learning method based on a full-slice graph structure. The image patch is constructed as a node in the graph structure, and the connection relationship between the nodes is constructed based on the physical spatial adjacency relationship, thereby forming a graph model that reflects the spatial topology of the full-slice image. This method can improve the stability and biological rationality of non-rigid registration.

[0018] 2) This invention segments whole-slice pathological images into multiple image patches, using these patches as basic processing units. A pre-trained model extracts high-dimensional semantic features from these patches for subsequent registration modeling. This method significantly reduces the number of basic units involved in registration calculations, lowers memory requirements, and improves processing efficiency for ultra-large-scale medical images. Furthermore, extracting high-dimensional semantic features from image patches reduces interference from staining differences, local artifacts, and other noise on the registration results.

[0019] 3) To address the issue that image blocks and graph structures are discrete representations, this invention further extends node-level displacements into a continuous deformation field covering the entire slice area through continuous function reconstruction. Based on this deformation field, non-rigid deformation is applied to the moving image, enabling the method to have both computational efficiency advantages and meet the requirements of medical images for deformation continuity and medical rationality. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the specific implementation process in the embodiment. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] Example like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a method for registering whole pathological slide images, which includes: S1, acquire the fixed image and the moving image to be registered, wherein the fixed image and the moving image are pathological whole slide images.

[0024] S2, downsample the fixed image and the moving image respectively to obtain the corresponding first image, and obtain the first tissue region mask corresponding to the first image.

[0025] In one embodiment of the present invention, the original fixed image and the moving image (40X field of view) are downsampled by 10 times in the first stage to obtain the corresponding first image (low-resolution image).

[0026] The first image is converted from the RGB color space to the HSV color space, and a first tissue region mask corresponding to the first image is obtained by color thresholding. Specifically: Adaptive threshold segmentation is performed using the statistical distribution of the luminance channel V and the saturation channel S. By calculating the 40% and 90% quantiles of luminance, the first image is divided into dark tissue regions and medium-luminance tissue regions, and background regions with saturation below a set threshold are removed from both types of tissue regions. The two types of tissue regions filtered by saturation threshold are merged to obtain an initial binary mask. Elliptical structuring elements are then used to perform morphological opening and closing operations to remove noise and fill holes, transforming the initial binary mask into a smooth and continuous first tissue region mask.

[0027] S3, based on the first image and the corresponding first tissue region mask, performs low-resolution coarse rigid registration to obtain a coarse-resolution rigid transformation matrix, specifically including: The first image and the corresponding first tissue region mask are downsampled by 10 times again. The background region of the downsampled first image is removed based on the downsampled first tissue region mask to obtain the second image. Key feature points are extracted from the second image. Similarity matching is performed on the feature descriptors of the second image corresponding to the fixed image and the second image corresponding to the moving image. After filtering valid matching point pairs using a cross-validation strategy, a random sampling consensus algorithm is used to estimate the valid matching point pairs, and the rigid transformation matrix between the fixed image and the moving image is calculated.

[0028] In one embodiment of the present invention, the SIFT feature extraction algorithm is used to extract key feature points from a second image.

[0029] In one embodiment of the present invention, the Brute-Force matcher is used to match the feature descriptors of the fixed image and the moving image, and the Euclidean distance (L2 norm) is used as the feature similarity measure. All matching results are sorted according to the matching distance, and a certain proportion of high-quality matching points with smaller distances are selected as valid matching points.

[0030] In one embodiment of the present invention, the Random Sample Consensus Algorithm (RANSAC) is used to robustly estimate valid matching point pairs to eliminate erroneous matching points, and the two-dimensional affine rigid transformation matrix between the two images is calculated: , in, It is a linear transformation matrix. This is the translation vector. The transformation matrix describes the translation, rotation, and scaling relationships between the moving image and the fixed image. Finally, the estimated rigid transformation matrix is ​​saved for subsequent high-resolution image registration initialization, improving the stability and computational efficiency of high-resolution fine registration.

[0031] S4, after mapping the first tissue region mask to the original resolution, a second tissue region mask is obtained, and the sliced ​​tissue region is extracted from the second tissue region mask.

[0032] In one embodiment of the present invention, the first tissue region mask is mapped to the original resolution using nearest neighbor interpolation to obtain the second tissue region mask.

[0033] S5, generate multiple sampling points at intervals within the sliced ​​organization area, and extract image patches of fixed size centered on the sampling points from the initial moving image and the fixed image. Define the sampling points of each image patch as nodes, define the pixel positions of the sampling points as node position information, and define the displacement vectors between two adjacent nodes as edge features to construct a full slice image structure.

[0034] In this embodiment, the full slice diagram structure , is represented as: , , in, For a set of nodes, Let be the set of edges. , These represent the sets of nodes in the moving image and the stationary image, respectively. These represent the sets of edges in the moving image and the fixed image, respectively; For the set of edges across slices, For the fixed image, the first 1 node For the moving image Each section, with its cross-slice edges, is only used for similarity calculations between subsequent slices and does not participate in the information transmission process of the graph neural network.

[0035] The specific construction process is as follows: In one embodiment of the present invention, the size of the image patch can be 512. 512, 1024 1024 pixels or other preset size. Each image patch is treated as a node. The spatial coordinates of a node are defined as the pixel position of that sampling point in the image: .

[0036] The pre-trained machine learning model CHIEF is used to extract the deep feature vectors of image patches, and these vectors, along with node location information, are used as node input features.

[0037] Within each slice, based on the spatial location of nodes, the spatial distance relationship between nodes is calculated using the k-nearest neighbor algorithm, and each node is connected to its k nearest neighbors. The value of k can be 4, 8, or other preset values. The set of neighborhood points of node i within the moving image is: Fixed nodes within the image The set of neighborhood points is: .

[0038] The characteristic of an edge is defined as the displacement vector between two neighboring nodes: .

[0039] The above nodes and edges constitute the graph structure of the moving and fixed images: ,in, Represents a set of nodes. Represents a set of edges, used to describe the spatial adjacency relationships between nodes.

[0040] Calculate the Euclidean distance between nodes across slices in the moving image and the fixed image, and select the k nearest neighbors: ; Based on cross-slice node adjacency relationships Construct cross-slice edges: ; Form a full-slice map structure that includes both intra-slice and cross-slice maps. : .

[0041] S6, based on the full slice image structure, uses a graph neural network for deformation field modeling, and models the deformation field learning and registration problem as an optimization learning problem of the graph neural network. The optimized graph neural network is used to perform high-resolution non-rigid registration and outputs the image block-level discrete node displacement.

[0042] In this embodiment, a graph neural network is used for deformation field modeling, which is represented as: , in, Parameterized by a graph neural network, information is propagated between neighboring nodes through a message passing mechanism; For image block-level discrete node displacement; It is a full slice diagram structure.

[0043] The node update process in a graph neural network specifically includes: Depth feature vectors are extracted from image patches, and these depth feature vectors, along with node location information, are used as node input features. ; Input features to nodes Perform feature mapping to obtain the initial features of the nodes. ; During the t-th round of message passing, node Based on the node's own characteristics Neighbor node characteristics and the spatial relationship between nodes Generate aggregated features , is represented as: ,in, For message functions; Indicates the movement of nodes within a slice The set of neighboring nodes.

[0044] According to the data passed to the node Aggregation features With the characteristics of the node itself Update the node features in round t+1. , is represented as: ,in, This is the node update function for the t-th round, used to fuse the node's own features with aggregated features; Based on the updated node characteristics of round t+1 The corresponding spatial displacement parameters are generated by reading the function. .

[0045] In one embodiment of the present invention, an unsupervised learning strategy is adopted, which combines feature similarity and smoothing regularization loss to train the graph neural network.

[0046] 1) Feature similarity calculation: Based on the spatial displacement predicted by the nodes Calculate the position of the node after the transformation: , in, For nodes The original coordinates in the original slice, These are the position coordinates after deformation. For each deformed moving image node... Find the nearest neighbor node in a fixed image. Define the node feature consistency loss: ; In the formula: Represents nodes in a moving slice Node features include location with depth features ; Represents the nodes in a fixed slice and a moving slice. The nearest neighbor node; This is for deformation operations on moving images; To obtain the depth feature vector of the moving image node; This is to obtain the depth feature vectors of fixed image nodes.

[0047] 2) Smoothing regularization loss: ; The regularization term constrains the displacement changes between adjacent nodes, ensuring the continuity and smoothness of the deformation field in space. Feature similarity and the smoothing regularization term together constitute the total loss function. : .

[0048] In the formula: These are the weight parameters.

[0049] Finally, the parameters of the graph neural network are updated through backpropagation, and multiple rounds of learning and optimization are performed to make the model more capable of generating realistic and reasonable deformation fields.

[0050] S7 converts the output image block-level discrete node displacements into a continuous deformation field to achieve high-resolution non-rigid registration. The continuous deformation field is used to perform spatial transformation on the moving image to obtain a registration result aligned with the fixed image.

[0051] Since graph neural networks only output patch-level discrete node displacements, a continuous deformation field needs to be constructed to achieve pixel-level registration.

[0052] In one embodiment of the present invention, the specific implementation process is as follows: For each pixel Based on the corresponding bottom left corner node Corresponding displacement Top left node Corresponding displacement bottom right node Corresponding displacement Top right node Corresponding displacement Pixel points are calculated using bilinear interpolation. The corresponding displacement is calculated using the following expression: ,in, , ; Pixels calculated based on bilinear interpolation Corresponding displacement The continuous deformation field was calculated. The expression is: ,in, Represents the pixels in the source image The mapping position in the target image.

[0053] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0054] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0055] The processing unit executes the various methods and processes described above, such as methods S1 to S8. For example, in some embodiments, methods S1 to S8 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1 to S8 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S8 by any other suitable means (e.g., by means of firmware).

[0056] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0057] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0058] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for registering whole-section pathological images, characterized in that, include: S1, acquire the fixed image and the moving image to be registered, wherein the fixed image and the moving image are pathological whole slide images; S2, downsample the fixed image and the moving image respectively to obtain the corresponding first image, and obtain the first tissue region mask corresponding to the first image; S3. Based on the first image and the corresponding first tissue region mask, perform low-resolution coarse rigid registration to obtain the rigid transformation matrix at coarse resolution. S4, after mapping the first tissue region mask to the original resolution, a second tissue region mask is obtained, and the sliced ​​tissue region is extracted from the second tissue region mask; S5, multiple sampling points are generated at intervals within the sliced ​​tissue area. Image blocks of a fixed size are extracted from the initial moving image and the fixed image, with the sampling points as the center. The sampling points of each image block are defined as nodes, and the pixel positions of the sampling points are defined as node position information. Then, the rigid transformation matrix is ​​mapped to the original resolution and applied to the node coordinates of the moving image. The displacement vector between two adjacent nodes is defined as the edge feature, and the full slice image structure is constructed. S6, based on the full slice image structure, uses graph neural network for deformation field modeling, and models the deformation field learning and registration problem as an optimization learning problem of graph neural network, and uses the optimized graph neural network to output the image block-level discrete node displacement; S7 converts the output image block-level discrete node displacements into a continuous deformation field to achieve high-resolution non-rigid registration. The continuous deformation field is used to perform spatial transformation on the moving image to obtain a registration result aligned with the fixed image.

2. The method for registering pathological whole-section images according to claim 1, characterized in that, The step of performing low-resolution coarse rigid registration based on the first image and the corresponding first tissue region mask to obtain a rigid transformation matrix specifically includes: The first image and the first tissue region mask are downsampled again, and the background region in the downsampled first image is removed based on the downsampled first tissue region mask to obtain the second image; Key feature points are extracted from the second image. Similarity matching is performed on the feature descriptors of the second image corresponding to the fixed image and the second image corresponding to the moving image. After screening the effective matching point pairs using a cross-validation strategy, the effective matching point pairs are estimated using a random sampling consensus algorithm. The rigid transformation matrix between the second image of the fixed image and the moving image is then calculated.

3. The method for registering whole-section pathological images according to claim 2, characterized in that, Obtaining the first tissue region mask corresponding to the first image through color threshold filtering specifically includes: Based on the brightness quantile threshold, the first image is divided into dark tissue regions and medium brightness tissue regions, and background regions with saturation below the set threshold are removed from both types of tissue regions. The two types of tissue regions filtered by saturation threshold are merged to obtain an initial binary mask. Then, elliptical structuring elements are used to perform morphological operations to transform the initial binary mask into a smooth and continuous first tissue region mask.

4. The method for registering pathological whole-section images according to claim 1, characterized in that, The deformation field modeling based on the full-slice graph structure and using a graph neural network is expressed as follows: ， in, Parameterized by a graph neural network, information is propagated between neighboring nodes through a message passing mechanism; For image block-level discrete node displacement; It is a full slice diagram structure.

5. The method for registering whole pathological slide images according to claim 1, characterized in that, The full slice structure , is represented as: ， ， in, For a set of nodes, Let be the set of edges. , These represent the sets of nodes in the moving image and the stationary image, respectively. These represent the sets of edges in the moving image and the fixed image, respectively; For the set of edges across slices, For the fixed image, the first 1 node For the moving image Each node, with its cross-slice edge, is only used for similarity calculations between the next two slices and does not participate in the information transmission process of the graph neural network.

6. The method for registering pathological whole-section images according to claim 4, characterized in that, The node update process in a graph neural network specifically includes: Extracting depth feature vectors from image patches The depth feature vector and node position information are used together as the node input features. ; Input features to nodes Perform feature mapping to obtain the initial features of the nodes. ; During the t-th round of message passing, node Based on the node's own characteristics Intra-image neighbor node features and the spatial relationship between nodes Generate aggregated features , is represented as: ,in, For message functions; Represents nodes within a moving image The set of neighboring nodes; According to the data passed to the node Aggregation features With the characteristics of the node itself Update # Wheel node features , is represented as: ,in, For the first Round node update function, used to fuse node's own features with aggregated features; According to the updated version Wheel node features Generate the corresponding spatial displacement parameters .

7. The method for registering pathological whole-section images according to claim 4, characterized in that, An unsupervised learning strategy is employed, combining feature similarity and smoothing regularization loss to train the graph neural network. The total loss function is calculated as follows: ， ， ， In the formula: The loss is the consistency loss of node features; To smooth out the loss of the regularization term; These are weight parameters; The set of edges within the moving image; For nodes Spatial predicted displacement; For nodes The corresponding node in the moving image; For nodes At the node corresponding to the fixed image; This is for deformation operations on moving images; To extract the depth feature vectors of nodes in a moving image; This is to extract the depth feature vectors of nodes in a fixed image.

8. The method for registering pathological whole-section images according to claim 1, characterized in that, The output image block-level discrete node displacements are interpolated to obtain a continuous deformation field, specifically: For each pixel Based on the corresponding bottom left corner node Corresponding displacement Top left node Corresponding displacement bottom right node Corresponding displacement Top right node Corresponding displacement Pixel points are calculated using bilinear interpolation. The corresponding displacement is calculated using the following expression: ,in, , ; Pixels calculated based on bilinear interpolation Corresponding displacement The continuous deformation field was calculated. The expression is: ,in, Represents the pixels in the source image The mapping position in the target image.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. 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 method as described in any one of claims 1 to 8.

Images