Three-dimensional cerebral vascular image segmentation methods, devices, electronic equipment and storage media

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for three-dimensional cerebral vascular image segmentation, belonging to the field of medical image segmentation technology. The method includes: inputting a downsampled image of the three-dimensional cerebral vascular image to be segmented into a downsampled segmentation model to obtain a global segmented image of the downsampled image output by the downsampled segmentation model; performing window sampling on the global segmented image to determine a set of clipping windows for the global segmented image, and performing clipping processing on the three-dimensional cerebral vascular image based on the set of clipping windows to obtain multiple image sub-blocks; inputting the multiple image sub-blocks into a local segmentation model to obtain a local segmented image of each image sub-block output by the local segmentation model; and stitching together the various local segmented images to obtain the segmentation result of the three-dimensional cerebral vascular image. This invention can achieve an image segmentation method that integrates global and local semantics, effectively improving the segmentation accuracy of three-dimensional cerebral vascular images.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, and in particular to a three-dimensional cerebral vascular image segmentation method, device, electronic device, and storage medium. Background Technology

[0002] During intracranial arterial interventional surgery, three-dimensional digital silhouette angiography can achieve accurate segmentation and morphological analysis of intracranial arteries, providing accurate anatomical information of the intracranial arterial region, which is of great significance for preoperative planning and intraoperative navigation.

[0003] In existing technologies, due to the extremely large size of three-dimensional cerebral vascular images of intracranial arteries, and given the limitations of actual memory, it is difficult to fit the entire volume of intracranial arteries into a three-dimensional segmentation model. Therefore, the conventional approach is to apply the three-dimensional segmentation model to three-dimensional sub-blocks cropped from the original volume. Although this method can capture the dependencies of the image in three dimensions, it only utilizes information from local regions for segmentation, ignoring the global semantics of the entire volume, resulting in low segmentation accuracy for three-dimensional cerebral vascular images.

[0004] Therefore, how to better achieve segmentation of three-dimensional cerebrovascular images has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for three-dimensional cerebral vascular image segmentation, so as to better achieve the segmentation of three-dimensional cerebral vascular images.

[0006] This invention provides a three-dimensional cerebral vascular image segmentation method, comprising: The downsampled image of the three-dimensional cerebral vascular image to be segmented is input into the downsampling segmentation model to obtain the global segmented image of the downsampled image output by the downsampling segmentation model; the downsampling segmentation model is trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels; Window sampling is performed on the global segmented image to determine the set of cropping windows for the global segmented image, and the three-dimensional cerebral vascular image is cropped based on the set of cropping windows to obtain multiple image sub-blocks; The multiple image sub-blocks are input into the local segmentation model to obtain the local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained based on the three-dimensional cerebrovascular image samples and their corresponding segmentation image labels using a trained downsampling segmentation model. The segmented images are stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0007] According to the present invention, a three-dimensional cerebral vascular image segmentation method is provided, wherein the downsampling segmentation model includes a feature extraction network; after cropping the three-dimensional cerebral vascular image based on the cropping window set to obtain multiple image sub-blocks, the method further includes: Obtain the target feature image output by the last layer of the feature extraction network; The target feature image and the global segmentation image are input into the spatial fusion module to obtain the global feature image of the three-dimensional cerebral vascular image output by the spatial fusion module; The global feature image is stitched together with the multiple image sub-blocks to obtain multiple stitched image sub-blocks; Each of the stitched image sub-blocks is locally segmented and stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0008] According to a three-dimensional cerebral vascular image segmentation method provided by the present invention, the spatial fusion module is specifically used to: crop the global segmented image based on the cropping window set to obtain multiple cropping image sub-blocks, and crop the target feature image based on the cropping window set to obtain multiple feature image sub-blocks; Based on the multiple cropped image sub-blocks and the multiple feature image sub-blocks, a topological map of cerebral vascular tissue is determined; Feature extraction is performed on the topological map of the cerebral blood vessels to obtain the global feature image.

[0009] According to a three-dimensional cerebral vascular image segmentation method provided by the present invention, the step of determining a cerebral vascular tissue topology map based on the plurality of cropped image sub-blocks and the plurality of feature image sub-blocks includes: For any one of the plurality of cutout sub-blocks, determine the first number of cerebral vascular voxels in the first cutout sub-block, the second number of cerebral vascular voxels in any other cutout sub-block among the plurality of cutout sub-blocks, and the third number of the same vascular voxels in the first cutout sub-block and any other cutout sub-block. Based on the first quantity, the second quantity, and the third quantity, the adjacency degree between each of the cropped image sub-blocks is determined, and an adjacency matrix is ​​determined according to each of the adjacency degrees; Global average pooling is performed on each of the feature map sub-blocks to obtain the feature vector corresponding to each feature map sub-block; the cerebral vascular tissue topology map is established using the feature vector as the node and the adjacency matrix as the edge.

[0010] According to a three-dimensional cerebral vascular image segmentation method provided by the present invention, the step of extracting features from the cerebral vascular tissue topology map to obtain the global feature image includes: The cerebral vascular tissue topology map is input into a graph neural network to obtain a set of node vectors output by the graph neural network; the graph neural network is used to encode and fuse the information of adjacent topological nodes in the cerebral vascular tissue topology map. Each node vector in the node vector group is used as the weight vector of the corresponding cropping window in the cropping window set, and each weight vector is weighted and summed with the feature map sub-block of the corresponding cropping window to obtain the global blood vessel feature; the global blood vessel feature is upsampled to obtain the global feature image.

[0011] According to a three-dimensional cerebral vascular image segmentation method provided by the present invention, the step of window sampling of the globally segmented image to determine the clipping window set of the globally segmented image includes: The global segmentation image is converted into point cloud data; The point cloud data is sampled at the farthest point to obtain the sparse point cloud data corresponding to the global segmentation image; Centered on each point in the sparse point cloud data, a set of cropping windows for the global segmentation image is generated according to a preset window size.

[0012] According to a three-dimensional cerebral vascular image segmentation method provided by the present invention, before inputting the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model, the method further includes: The acquired three-dimensional cerebrovascular image samples and their corresponding segmented image labels are used as a set of training samples, and multiple sets of the training samples are obtained. The downsampling segmentation model is trained using multiple sets of training samples to obtain a well-trained downsampling segmentation model; The downsampled image of the three-dimensional cerebrovascular image sample is input into the trained downsampled segmentation model to obtain the global segmentation image of the three-dimensional cerebrovascular image sample, and based on the global segmentation image of the three-dimensional cerebrovascular image sample, the set of clipping windows corresponding to the three-dimensional cerebrovascular image sample is determined. Based on the set of clipping windows corresponding to the three-dimensional cerebrovascular image samples, the three-dimensional cerebrovascular image samples and their corresponding segmented image labels are clipped to obtain multiple sub-block segmentation masks and label sub-blocks corresponding to each sub-block segmentation mask. The local segmentation model is trained using the multiple sub-block segmentation masks and the labeled sub-blocks corresponding to each sub-block segmentation mask to obtain a trained local segmentation model.

[0013] The present invention also provides a three-dimensional cerebral vascular image segmentation device, comprising: The first segmentation module is used to input the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels. The first cropping module is used to perform window sampling on the global segmented image, determine the cropping window set of the global segmented image, and perform cropping processing on the three-dimensional cerebral vascular image based on the cropping window set to obtain multiple image sub-blocks; the second segmentation module is used to input the multiple image sub-blocks into a local segmentation model to obtain a local segmented image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a pre-trained downsampling segmentation model based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels; The first stitching module is used to stitch together the various local segmented images to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0014] The present invention also provides 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 three-dimensional cerebral vascular image segmentation method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the three-dimensional cerebral vascular image segmentation method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the three-dimensional cerebral vascular image segmentation method as described above.

[0017] The present invention provides a three-dimensional cerebral vascular image segmentation method, apparatus, electronic device, and storage medium. It trains a downsampling segmentation model using three-dimensional cerebral vascular image samples and their corresponding segmented image labels. Using this trained downsampling segmentation model, a local segmentation model is further trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels. This allows the downsampling segmentation model to perform global segmentation of the downsampling images of the three-dimensional cerebral vascular images. Through window sampling, an adaptive set of clipping windows for the three-dimensional morphology of cerebral blood vessels is obtained. Then, based on the clipping of the original three-dimensional cerebral vascular images, image sub-blocks containing global semantics are obtained. The local segmentation model then performs local segmentation on each image sub-block, and the various locally segmented images are stitched together to obtain the final segmentation result of the three-dimensional cerebral vascular image. This achieves an image segmentation method that integrates global and local semantics, effectively improving the segmentation accuracy of three-dimensional cerebral vascular images. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the three-dimensional cerebral vascular image segmentation method provided by the present invention.

[0020] Figure 2 This is one of the schematic diagrams illustrating the principle of the three-dimensional cerebral vascular image segmentation method provided by the present invention.

[0021] Figure 3 This is the second schematic diagram of the principle flow of the three-dimensional cerebral vascular image segmentation method provided by the present invention.

[0022] Figure 4 This is a flowchart illustrating the principle of the spatial fusion module in the three-dimensional cerebral vascular image segmentation method provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the three-dimensional cerebral vascular image segmentation device provided by the present invention.

[0024] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this 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 this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] In the description of the invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0027] The following is combined with Figures 1-6 This invention describes a three-dimensional cerebral vascular image segmentation method, apparatus, electronic device, and storage medium.

[0028] Figure 1 This is a flowchart illustrating the three-dimensional cerebral vascular image segmentation method provided by the present invention, as shown below. Figure 1 As shown, it includes steps 110, 120, 130 and 140.

[0029] Step 110: Input the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels.

[0030] Step 120: Window sampling is performed on the global segmented image to determine the set of cropping windows for the global segmented image, and the three-dimensional cerebral vascular image is cropped based on the set of cropping windows to obtain multiple image sub-blocks.

[0031] Step 130: Input multiple image sub-blocks into the local segmentation model to obtain the local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a pre-trained downsampling segmentation model based on three-dimensional cerebrovascular image samples and their corresponding segmentation image labels.

[0032] Step 140: The segmented images of each local area are stitched together to obtain the segmentation results of the three-dimensional cerebral vascular image.

[0033] Specifically, the three-dimensional cerebral vascular images described in the embodiments of the present invention refer to three-dimensional cerebral vascular images to be segmented, which can be cerebral vascular images of any three-dimensional modality, such as common CT scan images, magnetic resonance imaging (MRI) scan images, etc.

[0034] The downsampled images described in this embodiment of the invention refer to images obtained by downsampling three-dimensional cerebral vascular images.

[0035] The downsampling segmentation model described in this embodiment of the invention is obtained by training a deep neural network model based on three-dimensional cerebral vascular image samples and their corresponding segmented image labels. It is used for global segmentation of the input downsampling image. Specifically, the deep neural network model can be a three-dimensional U-Net network model, a three-dimensional segmentation model like SwinUNETR, or other neural network models capable of three-dimensional segmentation; this invention does not impose any specific limitations on this.

[0036] The clipping window set described in this embodiment of the invention refers to a set of windows that characterizes and covers the three-dimensional morphology of cerebral blood vessels. In other words, the combination of each window in the set at the corresponding position can cover the three-dimensional morphology of the entire intracranial artery and cerebral blood vessels.

[0037] The local segmentation model described in this embodiment of the invention is obtained by training a deep neural network based on a pre-trained downsampling segmentation model and its corresponding segmented image labels for three-dimensional cerebral vascular image samples. This model is used to further segment the input three-dimensional image sub-blocks to obtain finer segmentation results. Specifically, the deep neural network model can also be a three-dimensional U-Net network model, a three-dimensional segmentation model like SwinUNETR, or other neural network models capable of three-dimensional segmentation; this invention does not impose specific limitations on this.

[0038] In an embodiment of the present invention, a downsampling segmentation model can be trained in advance by collecting a large number of three-dimensional cerebrovascular image samples and their corresponding segmentation image labels to obtain a trained downsampling segmentation model. Then, the local segmentation model is trained again using the three-dimensional cerebrovascular image samples and their corresponding segmentation image labels, combined with the trained downsampling segmentation model, to finally obtain a trained local segmentation model.

[0039] In an embodiment of the present invention, in step 110, the three-dimensional cerebral vascular image to be segmented is first downsampled to obtain a corresponding downsampled image, which is then input into a trained downsampled segmentation model for global segmentation processing. This process yields a global segmented image of the downsampled cerebral vascular image, thereby achieving global segmentation of the downsampled image.

[0040] In an embodiment of the present invention, in step 120, dynamic window sampling can be used to sample the global segmented image obtained in step 110 to obtain a set of cropped windows for the global segmented image, thereby obtaining a set of windows for adaptive three-dimensional vascular morphology. Furthermore, this set of cropped windows can be used to crop the original-size three-dimensional cerebral vascular image to obtain multiple image sub-blocks.

[0041] Based on the above embodiments, as an optional embodiment, window sampling is performed on the global segmentation image to determine the cropping window set of the global segmentation image, including: Convert the global segmented image into point cloud data; Sampling of the farthest point in the point cloud data yields sparse point cloud data corresponding to the global segmentation image; Centered on each point in the sparse point cloud data, a set of cropping windows for global segmentation image is generated according to a preset window size.

[0042] Specifically, the preset window size described in the embodiments of the present invention refers to the pre-set three-dimensional window size for three-dimensional image segmentation, which can be set according to actual design requirements.

[0043] In an embodiment of the present invention, a dynamic window sampling method is used to perform window sampling on the global segmentation image to determine the cropping window set of the global segmentation image.

[0044] More specifically, the global segmented image output by the downsampling segmentation model is projected into 3D space and converted into corresponding point cloud data. Then, multiple rounds of farthest point sampling are performed on the point cloud data to collect point cloud data representing the 3D spatial morphology of cerebral blood vessels. In each round of sampling, already covered point cloud coordinates are removed. This process is iterated until all point cloud coordinates are covered, ultimately yielding sparse point cloud data corresponding to the global segmented image. Understandably, this sparse point cloud data can preserve the spatial morphology of cerebral blood vessels.

[0045] Furthermore, in an embodiment of the present invention, windows centered on each point in the sparse point cloud data are generated according to a preset window size, thereby forming a set of cropping windows for the global segmented image.

[0046] The method of this invention converts the global segmented image into a point cloud, uses dynamic window sampling to iteratively sample the farthest point cloud, obtains a sparse point cloud that can characterize the spatial morphology of cerebral blood vessels, and then generates a three-dimensional window set with points in each sparse point cloud as the center, thereby obtaining a window set containing the global semantics of the image. This is beneficial for making full use of the global semantics of the entire intracranial artery and cerebral blood vessel volume and improving the segmentation accuracy of three-dimensional cerebral blood vessel images.

[0047] Furthermore, in an embodiment of the present invention, in step 130, the multiple image sub-blocks obtained in step 120 are input into the trained local segmentation model to obtain the local segmented image of each image sub-block output by the local segmentation model, thereby achieving fine segmentation of the image sub-blocks.

[0048] Furthermore, in an embodiment of the present invention, in step 140, the various local segmented images are stitched together sequentially according to the positions of the aforementioned cropping windows, and finally the segmentation result of the three-dimensional cerebral vascular image can be obtained.

[0049] Figure 2 This is one of the schematic diagrams illustrating the principle of the three-dimensional cerebral vascular image segmentation method provided by the present invention, such as... Figure 2As shown, in an embodiment of the present invention, the three-dimensional cerebral vascular image to be segmented is downsampled to obtain a downsampled image, which is then input into a downsampled segmentation model to obtain a global segmented image of the downsampled image. Then, using a dynamic window sampling method, window sampling is performed on the global segmented image to generate a set of clipping windows for the global segmented image. This set of clipping windows is then used to clip the original three-dimensional cerebral vascular image, resulting in multiple image sub-blocks. Further, these multiple image sub-blocks are input into a local segmentation model for local segmentation, thereby obtaining a local segmented image for each image sub-block. Finally, the various locally segmented images are stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0050] The three-dimensional cerebral vascular image segmentation method of this invention trains a downsampling segmentation model using three-dimensional cerebral vascular image samples and their corresponding segmented image labels. Using this trained downsampling segmentation model, a local segmentation model is further trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels. This allows the downsampling segmentation model to perform global segmentation of the downsampling images of the three-dimensional cerebral vascular images. Through window sampling, a set of cropping windows adaptive to the three-dimensional morphology of cerebral blood vessels is obtained. Then, based on the cropping processing of the original three-dimensional cerebral vascular images, image sub-blocks containing global semantics are obtained. Each image sub-block is then locally segmented using the local segmentation model, and the various locally segmented images are stitched together to finally obtain the segmentation result of the three-dimensional cerebral vascular image. This method achieves an image segmentation approach that integrates global and local semantics, effectively improving the segmentation accuracy of three-dimensional cerebral vascular images.

[0051] Based on the above embodiments, as an optional embodiment, the downsampling segmentation model includes a feature extraction network; after cropping the three-dimensional cerebral vascular image based on a set of cropping windows to obtain multiple image sub-blocks, the method further includes: Obtain the target feature image output from the last layer of the feature extraction network; The target feature image and the global segmentation image are input into the spatial fusion module to obtain the global feature image of the three-dimensional cerebrovascular image output by the spatial fusion module. The global feature image is stitched together with multiple image sub-blocks to obtain multiple stitched image sub-blocks; Each stitched image sub-block is locally segmented and stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0052] It should be noted that an image segmentation model consists of a feature extraction part and a segmentation head part. The segmentation head part is the part of the model specifically used to perform image segmentation tasks. It is usually located at the end of the model, immediately after the feature extraction part, and is responsible for converting the feature maps extracted by the preceding network into pixel-level classification predictions.

[0053] Specifically, in embodiments of the present invention, the downsampling segmentation model includes a feature extraction network, which is part of the feature extraction portion. It is understood that the downsampling segmentation model also includes a segmentation head portion.

[0054] The target feature image described in this embodiment of the invention refers to the feature image output by the last layer of the feature extraction network in the downsampled image of a three-dimensional cerebral vascular image after it has been input into the downsampled segmentation model.

[0055] It is understandable that the global segmentation image is the segmented image that is finally output by the downsampled image of the three-dimensional cerebral vascular image after being input into the downsampled segmentation model.

[0056] The spatial fusion module described in this embodiment of the invention is used to extract global features from a three-dimensional cerebral vascular image based on the input target feature image and the global segmentation image, thereby obtaining a global feature image.

[0057] Figure 3 This is the second schematic diagram illustrating the principle and flow of the three-dimensional cerebral vascular image segmentation method provided by this invention, as shown below. Figure 3 As shown, in an embodiment of the present invention, the three-dimensional cerebral vascular image to be segmented is downsampled to obtain a corresponding downsampled image, which is then input into the downsampled segmentation model to obtain a global segmented image of the downsampled image. Then, dynamic window sampling is performed on the global segmented image to obtain a set of cropping windows for the global segmented image, and the original three-dimensional cerebral vascular image is cropped using this set of cropping windows to obtain multiple image sub-blocks.

[0058] Furthermore, the target feature image output from the last layer of the feature extraction network in the downsampling segmentation model is obtained, and the target feature image and the global segmentation image are input into the spatial fusion module. The target feature image and the global segmentation image are cropped using a set of cropping windows. Combined with the fusion processing of image features, the global feature image of the three-dimensional cerebral vascular image is extracted.

[0059] Furthermore, in an embodiment of the present invention, the global feature image extracted by the spatial fusion module is concat-stitched with multiple image sub-blocks obtained by cropping the original three-dimensional cerebral vascular image to obtain multiple stitched image sub-blocks. Then, each stitched image sub-block can be input into a local segmentation model for local segmentation, and the local segmentation results output by the local segmentation model are stitched together according to the position of the cropping window set to finally obtain the segmentation result of the three-dimensional cerebral vascular image.

[0060] The method of this invention introduces a spatial fusion module to perform feature segmentation and fusion on the feature image output by the feature extraction network in the downsampling segmentation model and the global segmentation image output by the model, thereby extracting global features of the three-dimensional cerebral vascular image, which is beneficial to further improve the segmentation accuracy of the three-dimensional cerebral vascular image.

[0061] Figure 4 This is a flowchart illustrating the principle of the spatial fusion module in the three-dimensional cerebral vascular image segmentation method provided by this invention, as follows: Figure 4 As shown, in an embodiment of the present invention, the spatial fusion module is specifically used for: The global segmented image is cropped based on the cropping window set to obtain multiple cropped image sub-blocks, and the target feature image is cropped based on the cropping window set to obtain multiple feature image sub-blocks. Based on multiple cropped image sub-blocks and multiple feature image sub-blocks, a topological map of cerebral vascular tissue is determined; Feature extraction was performed on the topological map of cerebral vascular tissue to obtain a global feature image.

[0062] Specifically, the cropped image sub-blocks described in the embodiments of the present invention refer to image sub-blocks obtained by cropping a global segmented image using a set of cropping windows.

[0063] The feature map sub-blocks described in the embodiments of the present invention refer to image sub-blocks obtained by cropping the target feature image using a set of cropping windows.

[0064] The cerebral vascular tissue topology map described in the embodiments of the present invention is a feature image used to characterize the topological regularity of cerebral vascular tissue.

[0065] In the embodiments of the present invention, the specific implementation of the spatial fusion module is as follows: After inputting the target feature image and the global segmentation image into the spatial fusion module, the two types of input images are cropped based on the aforementioned set of cropping windows. The global segmentation image is cropped to obtain multiple cropped image sub-blocks; at the same time, the target feature image is cropped to obtain multiple feature image sub-blocks.

[0066] Furthermore, in an embodiment of the present invention, a topological map of cerebral vascular tissue is determined based on multiple cropped image sub-blocks and multiple feature image sub-blocks.

[0067] Based on the above embodiments, as an optional embodiment, a topological map of cerebral vascular tissue is determined based on multiple cropped image sub-blocks and multiple feature image sub-blocks, including: For any one of the multiple clipping sub-blocks, determine the first number of cerebral vascular voxels in any one clipping sub-block, the second number of cerebral vascular voxels in any other clipping sub-block among the multiple clipping sub-blocks, and the third number of the same vascular voxels in any one clipping sub-block and any other clipping sub-block. Based on the first, second, and third quantities, the adjacency degree between each clipping image sub-block is determined, and the adjacency matrix is ​​determined according to each adjacency degree; Global Average Pooling (GAP) is performed on each feature map sub-block to obtain the feature vector corresponding to each feature map sub-block; A topological map of cerebral vascular tissue is constructed using eigenvectors as nodes and adjacency matrices as edges.

[0068] Specifically, the first quantity described in the embodiments of the present invention refers to the number of cerebrovascular voxels contained in any selected clipping sub-block among a plurality of clipping sub-blocks.

[0069] Any other cutout pattern sub-block described in the embodiments of the present invention refers to any cutout pattern sub-block among a plurality of cutout pattern sub-blocks, other than the selected cutout pattern sub-block.

[0070] The second quantity described in the embodiments of the present invention refers to the number of cerebrovascular voxels contained in any other cutout pattern sub-block.

[0071] The third quantity described in the embodiments of the present invention refers to the number of identical vascular voxels contained in any cutout sub-block and any other cutout sub-block.

[0072] Continue to refer to Figure 4 In embodiments of the present invention, for any one of a plurality of cutout image sub-blocks, the number of cerebral vascular voxels in that cutout image sub-block can be determined, i.e., a first number. Simultaneously, the number of cerebral vascular voxels in any other cutout image sub-block besides the selected one can be determined, i.e., a second number; furthermore, the number of identical vascular voxels contained in both the selected cutout image sub-block and any other cutout image sub-block can be determined, i.e., a third number.

[0073] Furthermore, in this embodiment, based on the first, second, and third quantities, a preset multiple of the third quantity (i.e., the number of voxels commonly contained), such as twice, can be calculated as the ratio of this multiple to the sum of the first and second quantities. This ratio determines the adjacency degree (CA) between each clipped image sub-block and is used as the adjacency degree between two window sub-blocks in the cerebral vascular tissue topology map. Subsequently, an adjacency matrix can be generated based on the adjacency degree of each clipped image sub-block relative to other clipped image sub-blocks.

[0074] Meanwhile, in this embodiment, GAP processing is performed on each feature map sub-block to obtain the feature vector corresponding to each feature map sub-block.

[0075] Furthermore, in an embodiment of the present invention, a topological map of cerebral vascular tissue is established by using the feature vector corresponding to each feature map sub-block as a node and the generated adjacency matrix as an edge.

[0076] The method of this invention considers the number of cerebral vascular voxels in each cropped image sub-block, mines the adjacency degree between each cropped image sub-block, constructs an adjacency matrix, and performs global average pooling on each feature image sub-block to obtain the corresponding feature vector. Thus, a cerebral vascular tissue topology map can be established with the feature vector as nodes and the adjacency matrix as edges, which is beneficial for more accurate extraction of global feature images of three-dimensional cerebral vascular images and further improves the segmentation accuracy of three-dimensional cerebral vascular images.

[0077] Furthermore, in an embodiment of the present invention, feature extraction is performed on the topological map of cerebral vascular tissue to obtain a global feature image.

[0078] Based on the above embodiments, as an optional embodiment, feature extraction is performed on the topological map of cerebral vascular tissue to obtain a global feature image, including: The topological map of cerebral vascular tissue is input into a graph neural network to obtain a set of node vectors output by the graph neural network; the graph neural network is used to encode and fuse the information of adjacent topological nodes in the topological map of cerebral vascular tissue. Each node vector in the node vector group is used as the weight vector of the corresponding clipping window in the clipping window set, and each weight vector is weighted and summed with the feature map sub-block of the corresponding clipping window to obtain the global features of the blood vessel. The global features of blood vessels are upsampled to obtain a global feature image.

[0079] Specifically, the Graph Neural Network (GNN) described in this embodiment of the invention is a type of neural network model used to process graph-structured data. It learns graph-structured data from a cerebral vascular tissue topology map using neural networks, thereby extracting and discovering features and patterns within the graph-structured data. Specifically, it is used to encode and fuse data information of adjacent topological nodes in a cerebral vascular tissue topology map, extracting node feature vectors.

[0080] Continue to refer to Figure 4 In an embodiment of the present invention, the aforementioned established cerebral vascular tissue topology map is input into a graph neural network for processing. After graph convolution processing, the information of adjacent topology nodes is encoded and fused to obtain a new set of node vectors, i.e., a node vector group.

[0081] Furthermore, in an embodiment of the present invention, each node vector in the node vector group is used as the weight vector of the corresponding cropping window in the cropping window set. Combined with the feature map sub-blocks obtained from the aforementioned cropping target feature image, each weight vector is weighted and summed with the corresponding feature map sub-block of the cropping window to obtain the encoded global vascular feature image. Then, this global vascular feature image is upsampled to obtain the global feature image of the three-dimensional cerebral vascular image.

[0082] In embodiments of the present invention, by using graph neural networks to process the established cerebral vascular tissue topology map, the features of the graph structure data in the topology map can be fully extracted and explored. Furthermore, by fusing the features with the features of the corresponding cropping window feature map sub-blocks, the accuracy of extracting global features of three-dimensional cerebral vascular images can be further improved.

[0083] The method of this invention introduces a spatial fusion module, which crops and blocks the global segmented image and the target feature image based on the cropping window set, uses local block information to build a topological map of cerebral vascular tissue, and extracts its features. This can accurately and effectively extract the global features of the three-dimensional cerebral vascular image, which is beneficial to improving the segmentation accuracy of the subsequent three-dimensional cerebral vascular image.

[0084] Based on the above embodiments, as an optional embodiment, before inputting the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampling segmentation model to obtain the global segmented image of the downsampled image output by the downsampling segmentation model, the method further includes: The acquired three-dimensional cerebrovascular image samples and their corresponding segmented image labels are used as a set of training samples to obtain multiple sets of training samples. The downsampling segmentation model is trained using multiple sets of training samples to obtain a well-trained downsampling segmentation model; The downsampled images of the three-dimensional cerebrovascular image samples are input into the trained downsampled segmentation model to obtain the global segmentation images of the three-dimensional cerebrovascular image samples. Based on the global segmentation images of the three-dimensional cerebrovascular image samples, the set of clipping windows corresponding to the three-dimensional cerebrovascular image samples is determined. Based on the set of clipping windows corresponding to the three-dimensional cerebrovascular image samples, the three-dimensional cerebrovascular image samples and their corresponding segmented image labels are clipped to obtain multiple sub-block segmentation masks and label sub-blocks corresponding to each sub-block segmentation mask. The local segmentation model is trained using the multiple sub-block segmentation masks and the label sub-blocks corresponding to each sub-block segmentation mask to obtain the trained local segmentation model.

[0085] Specifically, in the embodiments of the present invention, before inputting the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model, it is necessary to train the downsampled segmentation model and the local segmentation model.

[0086] In embodiments of the present invention, a large set of three-dimensional cerebral vascular image samples is collected, and each three-dimensional cerebral vascular image sample and its corresponding segmented image label are used as a set of training samples, thereby obtaining multiple sets of training samples. By using multiple sets of training samples to train the downsampling segmentation model, a trained downsampling segmentation model is obtained.

[0087] Specifically, in an embodiment of the present invention, for any training sample, its three-dimensional cerebral vascular image sample and its corresponding segmentation image label are downsampled to obtain a downsampled image corresponding to the three-dimensional cerebral vascular image sample and its corresponding downsampled label. Then, the downsampled image is input into the downsampled segmentation model of the aforementioned dynamic window inference method to generate a downsampled segmentation mask.

[0088] Then, in an embodiment of the present invention, based on the downsampling segmentation mask and the downsampling label, the loss function L of the downsampling segmentation model is calculated. d The network is trained using the stochastic gradient descent algorithm. The loss function L... d It can be represented as: Among them, Y d This refers to the aforementioned downsampling segmentation mask. This refers to the aforementioned downsampling label.

[0089] Subsequently, in an embodiment of the present invention, after the downsampling segmentation model is trained, the downsampling image of the three-dimensional cerebrovascular image sample is input into the trained downsampling segmentation model to obtain the global segmentation image corresponding to the sample. Based on dynamic window sampling, window sampling is performed on the global segmentation image to obtain the clipping window set corresponding to the three-dimensional cerebrovascular image sample.

[0090] Furthermore, in an embodiment of the present invention, based on the set of clipping windows, the three-dimensional cerebral vascular image samples and their corresponding segmented image labels are clipped to obtain multiple sub-block segmentation masks and label sub-blocks corresponding to each sub-block segmentation mask. Then, a loss function L can be calculated based on the sub-block segmentation masks and the corresponding label sub-blocks. s The local segmentation model is trained using the stochastic gradient descent algorithm, where the loss function L... s It can be represented as: Among them, Y s This refers to the aforementioned sub-block segmentation mask. This refers to the aforementioned tag sub-block.

[0091] In an embodiment of the present invention, the local segmentation model is iteratively trained by using the sub-block segmentation mask and the labeled sub-block corresponding to each sample, and finally a trained local segmentation model can be obtained.

[0092] Optionally, in the embodiments of the present invention, the specific implementation method of model training for the technical scenario of introducing a spatial fusion module is described as follows.

[0093] First, the downsampling segmentation model can be trained using the aforementioned training method. Then, the model parameters of the downsampling segmentation model are fixed, and the downsampling images of the three-dimensional cerebral vascular image samples are input into the downsampling segmentation model. The feature map of the last layer of the feature extraction network in the model is then output.

[0094] Then, in this embodiment, using the aforementioned clipping window set, sub-blocks of the three-dimensional cerebral vascular image samples and their labels, along with the aforementioned feature maps, are sampled in the vascular region to obtain image sub-blocks, label sub-blocks, and feature map sub-blocks, respectively. Furthermore, the label sub-blocks and feature map sub-blocks can be input into the spatial fusion module to generate fused features, i.e., global features. Next, the image sub-blocks and the fused features are concatenated and input into the local segmentation model of the aforementioned dynamic window inference method to generate corresponding sub-block segmentation masks. Furthermore, in this embodiment, referring to the aforementioned method, the loss function L is calculated based on the sub-block segmentation mask and label sub-block. s The spatial fusion module and the local segmentation model are trained using the stochastic gradient descent algorithm until training is complete.

[0095] The method of this invention trains the segmentation model using multiple sets of training samples, controls the loss value of each model within the convergence range, improves the model training accuracy, and thus improves the accuracy of the model in segmenting three-dimensional cerebral vascular images.

[0096] The three-dimensional cerebral vascular image segmentation device provided by the present invention is described below. The three-dimensional cerebral vascular image segmentation device described below can be referred to in correspondence with the three-dimensional cerebral vascular image segmentation method described above.

[0097] Figure 5 This is a schematic diagram of the structure of the three-dimensional cerebral vascular image segmentation device provided by the present invention, as shown below. Figure 5 As shown, it includes: a first segmentation module 510, a first trimming module 520, a second segmentation module 530, and a first splicing module 540 connected in sequence.

[0098] The first segmentation module 510 is used to input the downsampled image of the three-dimensional cerebrovascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on the three-dimensional cerebrovascular image samples and their corresponding segmented image labels.

[0099] The first cropping module 520 is used to perform window sampling on the global segmented image, determine the cropping window set of the global segmented image, and perform cropping processing on the three-dimensional cerebral vascular image based on the cropping window set to obtain multiple image sub-blocks.

[0100] The second segmentation module 530 is used to input multiple image sub-blocks into the local segmentation model to obtain the local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a pre-trained downsampling segmentation model based on three-dimensional cerebrovascular image samples and their corresponding segmentation image labels.

[0101] The first stitching module 540 is used to stitch together the various local segmented images to obtain the segmentation results of the three-dimensional cerebral vascular images.

[0102] The three-dimensional cerebral vascular image segmentation device described in this embodiment can be used to execute the above-described three-dimensional cerebral vascular image segmentation method embodiment. Its principle and technical effect are similar, and will not be repeated here.

[0103] The three-dimensional cerebrovascular image segmentation device of this invention trains a downsampling segmentation model using three-dimensional cerebrovascular image samples and their corresponding segmented image labels. Using this trained downsampling segmentation model, a local segmentation model is further trained based on the three-dimensional cerebrovascular image samples and their corresponding segmented image labels. This allows the downsampling segmentation model to perform global segmentation of the downsampling images of the three-dimensional cerebrovascular images. Through window sampling, a set of cropping windows adaptive to the three-dimensional morphology of cerebrovascular vessels is obtained. Then, based on the cropping processing of the original three-dimensional cerebrovascular images, image sub-blocks containing global semantics are obtained. The local segmentation model then performs local segmentation on each image sub-block, and the various locally segmented images are stitched together to finally obtain the segmentation result of the three-dimensional cerebrovascular image. This achieves an image segmentation method that integrates global and local semantics, effectively improving the segmentation accuracy of three-dimensional cerebrovascular images.

[0104] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logic instructions in the memory 630 to execute the three-dimensional cerebral vascular image segmentation method provided by the above methods. The method includes: inputting a downsampled image of the three-dimensional cerebral vascular image to be segmented into a downsampled segmentation model to obtain a global segmentation image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; performing window sampling on the global segmentation image to determine a set of clipping windows for the global segmentation image, and performing clipping processing on the three-dimensional cerebral vascular image based on the set of clipping windows to obtain multiple image sub-blocks; inputting the multiple image sub-blocks into a local segmentation model to obtain a local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a trained downsampled segmentation model based on the three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; and stitching together the various local segmentation images to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0105] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the three-dimensional cerebral vascular image segmentation method provided by the above methods. The method includes: inputting a downsampled image of the three-dimensional cerebral vascular image to be segmented into a downsampled segmentation model to obtain a global segmentation image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; performing window sampling on the global segmentation image to determine a set of clipping windows for the global segmentation image, and performing clipping processing on the three-dimensional cerebral vascular image based on the set of clipping windows to obtain multiple image sub-blocks; inputting the multiple image sub-blocks into a local segmentation model to obtain a local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a trained downsampled segmentation model based on the three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; and stitching together the various local segmentation images to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0107] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the three-dimensional cerebral vascular image segmentation method provided by the above methods. The method includes: inputting a downsampled image of a three-dimensional cerebral vascular image to be segmented into a downsampled segmentation model to obtain a global segmentation image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; performing window sampling on the global segmentation image to determine a set of clipping windows for the global segmentation image, and performing clipping processing on the three-dimensional cerebral vascular image based on the set of clipping windows to obtain multiple image sub-blocks; inputting the multiple image sub-blocks into a local segmentation model to obtain a local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained using a trained downsampled segmentation model based on the three-dimensional cerebral vascular image samples and their corresponding segmentation image labels; and stitching together the various local segmentation images to obtain the segmentation result of the three-dimensional cerebral vascular image.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional cerebral vascular image segmentation method, characterized in that, include: The downsampled image of the three-dimensional cerebral vascular image to be segmented is input into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model; The downsampling segmentation model is trained based on three-dimensional cerebrovascular image samples and their corresponding segmentation image labels. Window sampling is performed on the global segmented image to determine the set of cropping windows for the global segmented image, and the three-dimensional cerebral vascular image is cropped based on the set of cropping windows to obtain multiple image sub-blocks; The multiple image sub-blocks are input into the local segmentation model to obtain the local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained based on the three-dimensional cerebrovascular image samples and their corresponding segmentation image labels using a trained downsampling segmentation model. The segmented images are stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image; The downsampling segmentation model includes a feature extraction network; After cropping the three-dimensional cerebral vascular image based on the set of cropping windows to obtain multiple image sub-blocks, the method further includes: Obtain the target feature image output by the last layer of the feature extraction network; The target feature image and the global segmentation image are input into the spatial fusion module to obtain the global feature image of the three-dimensional cerebral vascular image output by the spatial fusion module; The global feature image is stitched together with the multiple image sub-blocks to obtain multiple stitched image sub-blocks; Each of the stitched image sub-blocks is locally segmented and stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image; The spatial fusion module is specifically used for: The global segmented image is cropped based on the cropping window set to obtain multiple cropped image sub-blocks, and the target feature image is cropped based on the cropping window set to obtain multiple feature image sub-blocks. Based on the multiple cropped image sub-blocks and the multiple feature image sub-blocks, a topological map of cerebral vascular tissue is determined; Feature extraction is performed on the topological map of the cerebral vascular tissue to obtain the global feature image; The process of determining the cerebral vascular tissue topology map based on the plurality of cropped image sub-blocks and the plurality of feature image sub-blocks includes: For any one of the plurality of cutout sub-blocks, determine the first number of cerebral vascular voxels in the first cutout sub-block, the second number of cerebral vascular voxels in any other cutout sub-block among the plurality of cutout sub-blocks, and the third number of the same vascular voxels in the first cutout sub-block and any other cutout sub-block. Based on the first quantity, the second quantity, and the third quantity, the adjacency degree between each of the cropped image sub-blocks is determined, and an adjacency matrix is ​​determined according to each of the adjacency degrees; Global average pooling is performed on each of the feature map sub-blocks to obtain the feature vector corresponding to each feature map sub-block; A topological map of the cerebral vascular tissue is constructed using the feature vectors as nodes and the adjacency matrix as edges.

2. The three-dimensional cerebral vascular image segmentation method according to claim 1, characterized in that, The step of extracting features from the topological map of the cerebral vascular tissue to obtain the global feature image includes: The cerebral vascular tissue topology map is input into a graph neural network to obtain a set of node vectors output by the graph neural network; the graph neural network is used to encode and fuse the information of adjacent topological nodes in the cerebral vascular tissue topology map. Each node vector in the node vector group is used as the weight vector of the corresponding cropping window in the cropping window set, and each weight vector is weighted and summed with the feature map sub-block of the corresponding cropping window to obtain the global features of the blood vessel. The global features of the blood vessels are upsampled to obtain the global feature image.

3. The three-dimensional cerebral vascular image segmentation method according to claim 1, characterized in that, The step of performing window sampling on the global segmentation image to determine the cropping window set of the global segmentation image includes: The global segmentation image is converted into point cloud data; The point cloud data is sampled at the farthest point to obtain the sparse point cloud data corresponding to the global segmentation image; Centered on each point in the sparse point cloud data, a set of cropping windows for the global segmentation image is generated according to a preset window size.

4. The three-dimensional cerebral vascular image segmentation method according to any one of claims 1-3, characterized in that, Before inputting the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model, the method further includes: The acquired three-dimensional cerebrovascular image samples and their corresponding segmented image labels are used as a set of training samples, and multiple sets of the training samples are obtained. The downsampling segmentation model is trained using multiple sets of training samples to obtain a well-trained downsampling segmentation model; The downsampled image of the three-dimensional cerebrovascular image sample is input into the trained downsampled segmentation model to obtain the global segmentation image of the three-dimensional cerebrovascular image sample, and based on the global segmentation image of the three-dimensional cerebrovascular image sample, the set of clipping windows corresponding to the three-dimensional cerebrovascular image sample is determined. Based on the set of clipping windows corresponding to the three-dimensional cerebrovascular image samples, the three-dimensional cerebrovascular image samples and their corresponding segmented image labels are clipped to obtain multiple sub-block segmentation masks and label sub-blocks corresponding to each sub-block segmentation mask. The local segmentation model is trained using the multiple sub-block segmentation masks and the labeled sub-blocks corresponding to each sub-block segmentation mask to obtain a trained local segmentation model.

5. A three-dimensional cerebral vascular image segmentation device, characterized in that, include: The first segmentation module is used to input the downsampled image of the three-dimensional cerebral vascular image to be segmented into the downsampled segmentation model to obtain the global segmented image of the downsampled image output by the downsampled segmentation model; the downsampled segmentation model is trained based on the three-dimensional cerebral vascular image samples and their corresponding segmented image labels. The first cropping module is used to perform window sampling on the global segmented image, determine the cropping window set of the global segmented image, and perform cropping processing on the three-dimensional cerebral vascular image based on the cropping window set to obtain multiple image sub-blocks. The second segmentation module is used to input the multiple image sub-blocks into the local segmentation model to obtain the local segmentation image of each image sub-block output by the local segmentation model; the local segmentation model is trained based on the three-dimensional cerebrovascular image samples and their corresponding segmentation image labels using a trained downsampling segmentation model. The first stitching module is used to stitch together the various local segmented images to obtain the segmentation result of the three-dimensional cerebral vascular image; The downsampling segmentation model includes a feature extraction network; It also includes a module for determining the segmentation result, which is specifically used for: Obtain the target feature image output by the last layer of the feature extraction network; The target feature image and the global segmentation image are input into the spatial fusion module to obtain the global feature image of the three-dimensional cerebral vascular image output by the spatial fusion module; The global feature image is stitched together with the multiple image sub-blocks to obtain multiple stitched image sub-blocks; Each of the stitched image sub-blocks is locally segmented and stitched together to obtain the segmentation result of the three-dimensional cerebral vascular image; The spatial fusion module specifically includes: The cropping submodule is used to crop the global segmented image based on the cropping window set to obtain multiple cropped image sub-blocks, and to crop the target feature image based on the cropping window set to obtain multiple feature image sub-blocks. The topology map determination module is used to determine the topology map of cerebral vascular tissue based on the plurality of cropped map sub-blocks and the plurality of feature map sub-blocks; The feature extraction module is used to extract features from the topological map of the cerebral vascular tissue to obtain the global feature image; The topology graph determination module is specifically used for: For any one of the plurality of cutout sub-blocks, determine the first number of cerebral vascular voxels in the first cutout sub-block, the second number of cerebral vascular voxels in any other cutout sub-block among the plurality of cutout sub-blocks, and the third number of the same vascular voxels in the first cutout sub-block and any other cutout sub-block. Based on the first quantity, the second quantity, and the third quantity, the adjacency degree between each of the cropped image sub-blocks is determined, and an adjacency matrix is ​​determined according to each of the adjacency degrees; Global average pooling is performed on each of the feature map sub-blocks to obtain the feature vector corresponding to each feature map sub-block; A topological map of the cerebral vascular tissue is constructed using the feature vectors as nodes and the adjacency matrix as edges.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the three-dimensional cerebral vascular image segmentation method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the three-dimensional cerebral vascular image segmentation method as described in any one of claims 1 to 4.

Images