Method for training a blood vessel segmentation model and related products

Abstract

The application discloses a method for training a blood vessel segmentation model and related products, which comprises the following steps: obtaining a three-dimensional angiogram image for training, the three-dimensional angiogram image containing a blood vessel network composed of a plurality of vessel segments; giving each vessel segment a unique segment identifier to obtain a first segmentation mask of the three-dimensional angiogram image composed of segmentation masks of each vessel segment; determining the geometric features and semantic features of each vessel segment based on the segmentation mask of each vessel segment; constructing a first capsule vector and a first graph structure of all vessel segments based on the geometric features and semantic features of each vessel segment; and inputting the three-dimensional angiogram image, the first segmentation mask, the first capsule vector and the first graph structure as training data into the blood vessel segmentation model to train the same. The scheme of the application can realize high-reliability blood vessel segmentation recognition and provide strong support for accurate diagnosis of blood vessel diseases.

Description

Technical Field

[0001] This application generally relates to the field of medical device technology. More specifically, this application relates to a method, apparatus, and computer-readable storage medium for training a vascular segmentation model. Furthermore, this application also relates to an apparatus and computer-readable storage medium for vascular segmentation. Background Technology

[0002] In the diagnostic process of cerebrovascular diseases, precise digital subtraction angiography (DSA) vascular segmentation technology is a crucial step. By meticulously and accurately segmenting blood vessels, it provides doctors with key information on lesion localization, helping them to clearly grasp the specific location, extent, and severity of vascular lesions, thereby significantly improving the reliability of cerebrovascular disease diagnosis.

[0003] Currently, traditional vascular segmentation methods, such as threshold-based segmentation, morphological manipulation, and region growing, primarily rely on pixel-level features. These methods exhibit several insurmountable technical challenges when dealing with complex vascular networks. First, in actual imaging, factors such as image noise, uneven contrast agent filling, and excessively small vessel diameters often lead to continuous vessels being incorrectly segmented into multiple independent segments, severely disrupting the inherent continuity of the vascular structure—a problem of discontinuity. Second, when different vessels overlap or are adjacent in the projection direction, traditional methods often struggle to accurately distinguish them, resulting in different vessels being incorrectly connected into the same segment—a problem of adhesion. Third, existing methods generally lack the ability to effectively model the spatial structural relationships of vessels, failing to fully utilize the key features of vascular directionality and continuity, thus failing to ensure structural consistency during segmentation.

[0004] In view of this, this application provides a scheme for DSA vessel segmentation to achieve highly reliable cerebral vessel segmentation identification, providing strong support for the accurate diagnosis of cerebrovascular diseases. Summary of the Invention

[0005] In order to at least solve one or more of the technical problems mentioned above, this application proposes a scheme for DSA vessel segmentation in the following aspects.

[0006] In a first aspect, this application provides a method for training a vascular segmentation model, comprising: acquiring a three-dimensional angiography image for training, the three-dimensional angiography image containing a vascular network composed of multiple vascular segments; assigning a unique segment identifier to each vascular segment to obtain a first segmentation mask of the three-dimensional angiography image composed of segmentation masks of each vascular segment; determining the geometric and semantic features of each vascular segment based on the segmentation mask of each vascular segment; constructing a first capsule vector and a first graph structure for all vascular segments based on the geometric and semantic features of each vascular segment; and inputting the three-dimensional angiography image, the first segmentation mask, the first capsule vector, and the first graph structure as training data into the vascular segmentation model for training.

[0007] In some embodiments, the geometric features include the first endpoint, the last endpoint, the center point, the average diameter, and the arc length, and the semantic features include the anatomical type; and determining the geometric features of each vascular segment based on the segmentation mask of each vascular segment includes: performing a binarization operation on the three-dimensional angiography image to obtain a binarized angiography image; using a skeleton extraction algorithm to obtain the skeleton point set of the vascular network based on the binarized angiography image; determining the centerline point set of each vascular segment based on the segmentation mask of each vascular segment and the skeleton point set of the vascular network; and calculating the geometric features of each vascular segment based on the centerline point set of each vascular segment.

[0008] In some embodiments, based on the geometric and semantic features of each vascular segment, a first capsule vector and a first graph structure for all vascular segments are constructed; the geometric and semantic features of each vascular segment are structured and encoded to obtain the first capsule vector for all vascular segments; based on the first and last endpoints of each vascular segment, the distance between the last endpoint of each vascular segment and the first endpoints of other vascular segments is calculated; based on the distance and a predetermined threshold, the connection relationship between each vascular segment is determined to obtain the first graph structure.

[0009] In some embodiments, inputting the three-dimensional angiography image, the first segmentation mask, the first capsule vector, and the first graph structure as training data into the blood vessel segmentation model for training includes: inputting the three-dimensional angiography image into the blood vessel segmentation model to perform a segmentation operation to obtain a second segmentation mask, a second capsule vector, and a second graph structure; determining a segmentation loss, a capsule vector reconstruction loss, and a structural consistency loss constituting the prediction loss based on the first segmentation mask, the first capsule vector, the first graph structure, the second segmentation mask, the second capsule vector, and the second graph structure; and updating the parameters of the blood vessel segmentation model based on the prediction loss to achieve training.

[0010] In some embodiments, the blood vessel segmentation model includes a convolutional neural network, a graph neural network, and a segmentation inference network.

[0011] In some embodiments, inputting the three-dimensional angiography image into the vascular segmentation model to perform segmentation operations to obtain a second segmentation mask, a second capsule vector, and a second graph structure includes: inputting the three-dimensional angiography image into the convolutional neural network to perform preliminary segmentation to obtain a first feature map and a preliminary segmentation mask, and determining a second capsule vector and a second graph structure based on the preliminary segmentation mask; using the second capsule vector as node features and the second graph structure as edge information, and inputting the node features and the edge information into the graph neural network for structured inference to obtain structural features of each vascular segment; mapping the structural features of each vascular segment to all voxels corresponding to that vascular segment to obtain a second feature map, and fusing the first feature map and the second feature map to obtain a fused feature map; and inputting the fused feature map into the segmentation inference network to perform segmentation inference to obtain the second segmentation mask.

[0012] In a second aspect, this application provides an apparatus for training a vascular segmentation model, comprising: a processor; and a memory storing program instructions for training the vascular segmentation model, wherein when the program is executed by the processor, the method described according to the foregoing first aspect and several embodiments is implemented.

[0013] In a third aspect, this application provides an apparatus for segmenting blood vessels, comprising: a processor; and a memory storing program instructions for segmenting blood vessels, wherein when the program instructions are executed by the processor, the apparatus performs the following operations: acquiring a three-dimensional angiography image to be segmented; inputting the three-dimensional angiography image into a blood vessel segmentation model trained according to the method described in the first aspect and several embodiments above to perform a segmentation operation, thereby obtaining a blood vessel segmentation result.

[0014] In some embodiments, when the program instructions are executed by the processor, the device further performs the following operations: performs boundary smoothing and / or isolated vessel segment removal on the vessel segmentation results to obtain the final vessel segmentation results.

[0015] In the fourth aspect, there are computer-readable instructions stored thereon for training a blood vessel segmentation model and for segmenting blood vessels, which, when executed by one or more processors, implement the operations implemented by the method according to the first aspect and the plurality of embodiments described above, as well as the operations implemented by the device according to the third aspect and the embodiments described above.

[0016] Using the above-described scheme for DSA vessel segmentation, this embodiment of the application assigns a unique segment identifier to each vessel segment to obtain a first segmentation mask. Based on this mask, the geometric and semantic features of the vessel segment are determined, and then a first capsule vector and a first graph structure are constructed. These training data are used to train the vessel segmentation model. The first capsule vector preserves the geometric and semantic features of the vessel segment through vector encoding, and the first graph structure reconstructs the structural relationships between vessel segments through a dynamic connection mechanism. This allows the trained vessel segmentation model to effectively utilize the directional and continuous characteristics of vessels, accurately distinguishing adjacent or overlapping vessels, and avoiding incorrect segmentation of continuous vessels or incorrect connection of different vessels. This effectively overcomes the problems of disconnection, adhesion, and lack of structural consistency in existing technologies, achieving highly reliable cerebral vessel segmentation identification and providing strong support for the accurate diagnosis of cerebrovascular diseases. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein:

[0018] Figure 1 An exemplary flowchart of a method for training a blood vessel segmentation model according to an embodiment of this application is shown;

[0019] Figure 2 An exemplary process of segmentation operation in a blood vessel segmentation model according to an embodiment of this application is shown;

[0020] Figure 3 A structural block diagram of an apparatus for training a blood vessel segmentation model according to an embodiment of this application is shown;

[0021] Figure 4 An exemplary process for segmenting blood vessels, according to an embodiment of this application, is illustrated. Detailed Implementation

[0022] The technical solutions of the embodiments of this application 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 application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should be noted that in this specification, the terms "first," "second," etc., are used only to distinguish one feature from another and do not imply any limitation on the features. Therefore, without departing from the teachings of this application, the first segmentation mask discussed below may also be referred to as the second segmentation mask or the third segmentation mask.

[0025] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0026] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0027] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0028] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0029] Figure 1 An exemplary flowchart of a method 100 for training a blood vessel segmentation model according to an embodiment of this application is shown. It is understood that method 100 can be executed by any suitable device with data processing capabilities, such as, but not limited to, terminal devices and servers.

[0030] like Figure 1As shown, in step S101, a three-dimensional angiography image for training, namely a 3D DSA image (hereinafter referred to as the original 3D DSA image, original image, or original image), can be acquired. This 3D DSA image contains a vascular network composed of multiple vascular segments. In practice, the format of the three-dimensional angiography image can be, for example, .nii.gz format.

[0031] In the data preparation stage, in order to reduce the interference of image noise on model training and improve the quality of training data, the 3D angiography images used for training can be denoised. For example, Gaussian filtering or nonlocal mean filtering can be used. Through denoising, the interference of noise on the model's learning of vascular features can be significantly reduced, ensuring the integrity and accuracy of vascular structures in the training data. This provides high-quality input for the model's subsequent feature extraction and segmentation tasks, thereby improving the model's training effect and the final accuracy of vascular segmentation.

[0032] Next, in step S102, each blood vessel segment can be assigned a unique segment identifier to obtain a segment mask of the three-dimensional angiography image composed of the segment masks of each blood vessel segment (which can be called the first segment mask for easy distinction).

[0033] Each segment identifier corresponds to an independent vascular segment, used to mark the actual segmental boundaries of a blood vessel. For example, in a cerebral vascular network, different cerebral vascular segments are labeled with different labels, such as label1 representing the internal carotid artery segment and label2 representing the middle cerebral artery segment. In practice, medical professionals can manually delineate the boundaries of different branches or independent parts in the vascular network to ensure that each segment identifier corresponds to a continuous and complete vascular segment.

[0034] Segmentation masks are voxel-level annotations of the spatial boundaries of blood vessel segments. By assigning a segment identifier to each voxel corresponding to a blood vessel segment, the morphology, location, and extent of the blood vessel segment in three-dimensional space are precisely delineated. Therefore, the size of the segmentation mask for a 3D angiography image is consistent with the size of the 3D angiography image. Each voxel in the segmentation mask corresponds to the same spatial location in the 3D angiography image, and a unique segment identifier identifies which blood vessel segment the voxel at that location belongs to.

[0035] This consistency in size ensures that when training the blood vessel segmentation model, the model's predictions can be accurately compared with the actual annotations on a voxel-by-voxel basis, and the voxel-level segmentation loss can be calculated. This effectively guides model optimization and improves the accuracy of blood vessel segmentation.

[0036] Next, in step S103, the geometric and semantic features of each blood vessel segment can be determined based on the segmentation mask of each blood vessel segment.

[0037] In embodiments of this application, geometric features may include, but are not limited to, the first endpoint, the last endpoint, the center point, the average diameter, and the arc length. Semantic features may include anatomical type, which can identify the anatomical category to which the vessel segment belongs. These features can constitute the true capsule vector, used to calculate the capsule vector reconstruction loss, to constrain the accuracy of the capsule vectors extracted by the model.

[0038] In step S103, the geometric features of each vascular segment are determined based on the segmented mask of each segment. Specifically, the following operations can be performed: First, the three-dimensional angiography image is binarized to obtain a binarized angiography image, highlighting the vascular structure and eliminating interference from non-vascular information. Next, based on the binarized angiography image, a three-dimensional skeleton extraction algorithm is used to refine the vascular network, removing redundant voxels while retaining its core topology, and obtaining the skeleton point set of the vascular network composed of continuous spatial points. Subsequently, based on the segmented mask of each vascular segment and the skeleton point set of the vascular network, the centerline point set of each vascular segment is determined. Finally, based on the centerline point set of each vascular segment, the geometric features of each vascular segment are calculated.

[0039] Furthermore, when determining the centerline point set for each vascular segment based on its segmentation mask and the skeleton point set of the vascular network, the following operations can be performed for each vascular segment (referred to as the current vascular segment for ease of description): traverse each point in the skeleton point set and determine whether its corresponding voxel belongs to the segmentation mask of the current vascular segment in the first segmentation mask (i.e., determine whether the segment identifier of the voxel is equal to the segment identifier of the current vascular segment). If it does, the point can be determined as the centerline point of the current vascular segment; otherwise, it can be determined as not being the centerline point of the current vascular segment. Thus, by collecting all skeleton points belonging to the current vascular segment, the centerline point set of the current vascular segment can be constructed. After determining the centerline point set for each vascular segment, the geometric features of each vascular segment can be calculated based on the centerline point set of each vascular segment.

[0040] First, determining the initial and final endpoints requires an image search algorithm: The centerline point set is modeled as a graph (each point is a node, and adjacent points are connected by edges). All possible endpoints in the graph (such as nodes or branch points with a connectivity of 1) are traversed. A shortest path algorithm (such as Dijkstra's algorithm) is used to calculate the shortest path between any two endpoints, and the longest path (i.e., the "longest-shortest path") is selected. The two endpoints of this path are defined as the initial and final endpoints, and the direction of their connection accurately reflects the main extension direction of the vascular segment. Further, the midpoint of the Euclidean distance between the initial and final endpoints in three-dimensional space is calculated; this midpoint is the center point of the vascular segment.

[0041] In addition, other geometric features of the vascular segment are calculated as follows: the average diameter needs to be calculated by combining its minimum bounding box and centerline—first, the minimum bounding box of the vascular segment in three-dimensional space (i.e., the smallest cuboid that can completely contain all voxels of the vascular segment) is determined by segmented masking; centerline points are uniformly sampled within the bounding box; the average distance from each sampling point to the vascular boundary (determined based on the binarized image) is calculated, and then multiplied by 2 to obtain the average diameter; the arc length is the total length of the centerline point set, obtained by accumulating the Euclidean distance between adjacent points; the anatomical type needs to be labeled by combining the overall morphology of the vascular segment within the minimum bounding box (such as direction, branching pattern) and cerebral vascular anatomy atlas (such as the A1 segment of the anterior cerebral artery, the M2 segment of the middle cerebral artery, etc.). The minimum bounding box plays a spatial constraint role in this process, which can effectively eliminate the interference of background and other vascular segments on feature calculation and ensure the accuracy of the results.

[0042] Next, in step S104, the first capsule vector and the first graph structure of all vascular segments can be constructed based on the geometric and semantic features of each vascular segment.

[0043] Specifically, the first capsule vector of all vascular segments can be obtained by structurally encoding the geometric and semantic features of each vascular segment.

[0044] For each vascular segment, its geometric and semantic features are integrated. These multi-dimensional features are mapped into a fixed-dimensional vector through a preset encoding rule. This vector is the first capsule vector of the corresponding vascular segment. The geometric features retain the spatial morphology and size information of the vascular segment through vector components, while the semantic features encode the anatomical attributes of the vascular segment through specific components. This allows each first capsule vector to completely and structurally represent the core features of the corresponding vascular segment.

[0045] Furthermore, when obtaining the first graph structure, the distance between the tail endpoint of each vascular segment and the head endpoints of other vascular segments can be calculated, for example, using Euclidean distance. Then, based on this distance and a predetermined threshold, the connection relationships between the vascular segments are determined to obtain the first graph structure.

[0046] Specifically, if the distance is less than a predetermined threshold, it is determined that the two blood vessel segments are connected; otherwise, it is determined that the two blood vessel segments are not connected. In practice, the specific value of the predetermined threshold can be set according to the characteristics of the blood vessel structure, and this application does not impose specific limitations on it.

[0047] Finally, in step S105, the three-dimensional angiography image, the first segmentation mask, the first capsule vector, and the first graph structure can be input as training data into the blood vessel segmentation model for training.

[0048] Specifically, a 3D angiography image can be input into a vessel segmentation model to perform segmentation operations, obtaining a second segmentation mask, a second capsule vector, and a second graph structure. Then, based on the first segmentation mask, the first capsule vector, the first graph structure, the second segmentation mask, the second capsule vector, and the second graph structure, the segmentation loss, capsule vector reconstruction loss, and structural consistency loss constituting the prediction loss can be determined. Furthermore, the parameters of the vessel segmentation model can be updated based on this prediction loss to enable training.

[0049] The aforementioned segmentation loss takes the second segment mask as input and the difference between them can be measured using cross-entropy loss or Dice loss. The capsule vector reconstruction loss takes the second capsule vector and the first capsule vector as input and the difference can be calculated using mean squared error (MSE). The structural consistency loss calculates the cosine similarity of the capsule vectors of blood vessel segments with connectivity in the graph structure, encouraging similar structures to have consistent vector directions.

[0050] After determining the segmentation loss, capsule vector reconstruction loss, and structural consistency loss, different weights can be assigned to the three loss distributions, and the predicted loss can be obtained by calculating their weighted sum.

[0051] Next, when updating the parameters of the vessel segmentation model based on the predicted loss, the gradient of the parameters of each layer of the vessel segmentation model can be calculated using the backpropagation algorithm. This gradient reflects the degree of influence of the parameters on the loss. Furthermore, optimizers (such as Adam or SGD) can be used to adjust the parameters of each layer based on the gradient.

[0052] In embodiments of this application, the aforementioned vascular segmentation model may include a convolutional neural network, a graph neural network, and a segmentation inference network. Based on this, a three-dimensional angiography image is input into the vascular segmentation model to perform segmentation operations, thereby obtaining a second segmentation mask, a second capsule vector, and a second graph structure. Specifically, the following can be performed: Figure 2 The segmentation operation 200 in the blood vessel segmentation model shown.

[0053] like Figure 2 As shown, in step S201, the three-dimensional angiography image can be input into a convolutional neural network to perform preliminary segmentation to obtain a first feature map and a preliminary segmentation mask, and based on the preliminary segmentation mask, the second capsule vector and the second map structure are determined.

[0054] Specifically, the convolutional neural network can be 3D U-Net. Its Encoder part extracts multi-scale voxel-level features (such as the grayscale of blood vessels, boundaries, center lines, local morphology, and other low- to mid-level features) through multi-layer convolution and downsampling. The Decoder part restores the spatial resolution through upsampling and skip connections, and outputs a preliminary segmentation mask and voxel-level visual feature map of the same size as the original image (which can be called the first feature map for easy distinction).

[0055] It is understood that the process of determining the second capsule vector and the second graph structure based on the preliminary segmentation mask is the same as the process of determining the first capsule vector and the first graph structure based on the first segmentation mask described in steps S103 and S104 above, and they can be referred to each other. For the sake of brevity, this disclosure will not elaborate on them one by one.

[0056] In step S202, the second capsule vector can be used as node features, and the second graph structure can be used as edge information. The node features and edge information are then input into the graph neural network for structured reasoning to obtain the structural features of each blood vessel segment.

[0057] Specifically, graph neural networks (GNNs) learn the connectivity between blood vessel segments through a message passing mechanism. The core idea is to allow each node (blood vessel segment) to aggregate information from its neighboring nodes to update its own features to reflect the global connectivity and output the structural features of each blood vessel segment.

[0058] After GNN inference, the structural feature vector of each vascular segment not only retains its own geometric properties (from the capsule vector), but also encodes the global connectivity with other vascular segments (such as whether they belong to the same branch, the rationality of the connection, and their position in the vascular network), providing a basis for "structural consistency" for subsequent feature fusion and accurate segmentation.

[0059] In step S203, the structural features of each blood vessel segment can be mapped onto all voxels corresponding to that blood vessel segment to obtain a second feature map, and the first feature map and the second feature map are fused to obtain a fused feature map.

[0060] Specifically, the structural features originally represented by blood vessel segments are mapped onto each voxel in 3D space, forming a multi-channel GNN feature map with the same size as the original image (which can be called the second feature map for easy distinction), thereby achieving spatial matching between the structural features and the voxel-level visual feature map (i.e., the first feature map) extracted by the convolutional neural network.

[0061] For example, if the structural feature vector of a certain blood vessel segment contains the information that it is "closely connected to the upstream blood vessel", all voxels covered by it will be assigned a value reflecting this information on the corresponding channel, so that the subsequent network can combine local visual features and global structural features to determine the blood vessel segment to which the voxel belongs.

[0062] In step S204, the fused feature map is input into the segmentation inference network to perform segmentation inference in order to obtain the second segmentation mask.

[0063] In the embodiments of this application, the segmentation inference network may include multiple 3D convolutional modules, each of which may be composed of a 3D convolutional layer, a batch normalization layer and an activation function connected in sequence.

[0064] The 3D convolutional layer uses a three-dimensional convolutional kernel, which can extract features in the spatial dimension from the input fused feature map and capture the local correlation information of blood vessel segments in three-dimensional space. The batch normalization layer standardizes the features output by the convolutional layer, which accelerates the training convergence speed of the network and alleviates the overfitting problem. The activation function layer (such as the ReLU function) introduces non-linear transformation capability into the network, enabling the network to learn more complex feature mapping relationships.

[0065] In addition to convolutional modules, segmentation inference networks can also include upsampling layers. Upsampling layers use deconvolution operations or interpolation methods to restore the low-resolution feature maps processed by convolution to the same spatial size as the input 3D angiography image, ensuring that the final output segmentation result accurately corresponds to each voxel position of the original image.

[0066] In the final stage of the segmentation inference network, the feature map, after multiple rounds of processing, can be compressed through a 1×1×1 3D convolutional layer (or fully connected layer). Then, the probability of each voxel belonging to a different blood vessel segment is calculated using the softmax function, generating a voxel-level classification probability map. Further, the argmax function can be applied to the voxel-level classification probability map to select the segment identifier corresponding to the channel with the highest probability for each voxel, generating the final segmentation mask (i.e., the second segmentation mask). In the second segmentation mask, each voxel is uniquely labeled as either a blood vessel segment or background.

[0067] The above combination Figure 1 and Figure 2This application describes a method for training a vascular segmentation model. The method assigns a unique segment identifier to each vascular segment to obtain a first segmentation mask. Based on this mask, the geometric and semantic features of the vascular segment are determined, and a first capsule vector and a first graph structure are constructed. These training data are then used to train the vascular segmentation model. The first capsule vector preserves the geometric and semantic features of the vascular segment through vector encoding. The first graph structure reconstructs the structural relationships between vascular segments through a dynamic connection mechanism. This allows the trained vascular segmentation model to effectively utilize the directional and continuous characteristics of blood vessels, accurately distinguishing adjacent or overlapping vessels, and avoiding incorrect segmentation of continuous vessels or incorrect connection of different vessels. This effectively overcomes problems such as disconnection, adhesion, and lack of structural consistency in existing technologies, achieving highly reliable cerebral vascular segmentation recognition and providing strong support for the accurate diagnosis of cerebrovascular diseases.

[0068] Next, combine Figure 3 This application provides an exemplary description of a device 300 for training a blood vessel segmentation model, as exemplified in its embodiments. Figure 3 As shown, the electronic device 300 in this application embodiment may include a processor 301, a memory 302, and a communication bus 303.

[0069] In specific embodiments, the processor 301 described above can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), CPU, controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and this embodiment does not specifically limit it.

[0070] In this embodiment, the communication bus 303 is used to establish communication between the processor 301 and the memory 302; the memory 302 stores program instructions for training the blood vessel segmentation model; when the processor 301 executes the program instructions stored in the memory 302, it implements the combination of this application. Figures 1 to 2 The method described is for training a blood vessel segmentation model.

[0071] The above combination Figure 3 This paper describes an electronic device for training a blood vessel segmentation model that can be used to execute the present application. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation methods and implementation entities of this application are not limited thereto, but can be modified without departing from the spirit of this application.

[0072] Furthermore, embodiments of this application also provide a device for segmenting blood vessels, which, in conjunction with the above description... Figure 3 The described electronic device 300 for training a blood vessel segmentation model has a similar device structure. It is understood that the description of the various embodiments in this disclosure emphasizes the differences between the embodiments, while their similarities or corresponding aspects can be referred to mutually. For the sake of brevity, this disclosure will not elaborate further.

[0073] In the device for segmenting blood vessels according to this application, the memory stores program instructions for segmenting blood vessels. When these program instructions are executed by the processor, the device can perform the following: Figure 4 The operation shown is for segmenting blood vessels.

[0074] like Figure 4 As shown, in step S401, a three-dimensional angiography image of the segment to be obtained can be acquired.

[0075] In step 402, the 3D angiography image can be input into the trained vascular segmentation model to perform segmentation operations and obtain the vascular segmentation result. It can be understood that the trained vascular segmentation model refers to the model based on the preceding text combined with... Figures 1 to 2 The method described for training a blood vessel segmentation model has been successfully trained.

[0076] In addition, in order to eliminate noise interference and correct unreasonable local structures to obtain better vascular segmentation results, in step S403, boundary smoothing and / or isolated vascular segment removal can also be performed on the vascular segmentation results obtained in step S402 to obtain the final vascular segmentation results.

[0077] When performing boundary smoothing, a combination of "erosion-dilation" operations from 3D morphology can be used. First, erosion is used to shrink the blood vessel boundary and remove fine burrs. Then, dilation is used to restore the original size of the blood vessel, while smoothing the boundary contour and reducing jagged undulations, making the blood vessel contour more consistent with a realistic smooth morphology.

[0078] When removing isolated vessel segments, segments that are too small or have no reasonable connection to the main vascular network can be removed. In other words, an isolated vessel segment is defined as a segment that is not connected to any other vessel segment and has an area smaller than a predetermined threshold. In practice, the area of ​​a vessel segment can be represented by the number of voxels it contains, while the predetermined threshold can be set based on actual clinical experience; this application does not impose specific limitations on it.

[0079] The above combination Figure 4 This paper describes a device for vascular segmentation. The device obtains vascular segmentation results by inputting a 3D angiographic image to be segmented into a trained vascular segmentation model. Because the trained model preserves the geometric and semantic features of vascular segments through capsule vectors and vector encoding during training, and reconstructs the structural relationships between vascular segments through a dynamic graph connection mechanism, it can effectively utilize the directional and continuous features of blood vessels to accurately distinguish adjacent or overlapping vessels when segmenting new 3D angiographic images. This avoids the incorrect segmentation of continuous vessels or the incorrect connection of different vessels, effectively overcoming problems such as disconnection, adhesion, and lack of structural consistency in existing technologies. This achieves highly reliable cerebral vascular segmentation and recognition, providing strong support for the accurate diagnosis of cerebrovascular diseases.

[0080] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium. This computer-readable storage medium stores program instructions for training a blood vessel segmentation model and / or for blood vessel segmentation, which can be used to implement the embodiments of this application. Figures 1 to 2 The described method and / or combination for training a blood vessel segmentation model Figure 4 The operation performed by the device described for segmenting blood vessels.

[0081] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0082] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a blood vessel segmentation model, characterized in that, include: Acquire three-dimensional angiographic images for training, the three-dimensional angiographic images containing a vascular network composed of multiple vascular segments; Each vessel segment is assigned a unique segment identifier to obtain the first segment mask of the three-dimensional angiography image, which is composed of the segment masks of each vessel segment. Based on the segmentation mask of each blood vessel segment, the geometric and semantic features of each blood vessel segment are determined. Based on the geometric and semantic features of each vascular segment, the first capsule vector and the first graph structure of all vascular segments are constructed. The three-dimensional angiography image, the first segmentation mask, the first capsule vector, and the first graph structure are input as training data into the blood vessel segmentation model for training. The training process involves inputting the three-dimensional angiography image, the first segmentation mask, the first capsule vector, and the first graph structure as training data into the vessel segmentation model. The three-dimensional angiography image is input into the blood vessel segmentation model to perform segmentation operations, so as to obtain the second segmentation mask, the second capsule vector, and the second graph structure. Based on the first segmentation mask, the first capsule vector, the first graph structure, the second segmentation mask, the second capsule vector, and the second graph structure, the segmentation loss, capsule vector reconstruction loss, and structural consistency loss constituting the prediction loss are determined. Based on the predicted loss, the parameters of the blood vessel segmentation model are updated to enable training. The vascular segmentation model includes a convolutional neural network, a graph neural network, and a segmentation inference network; the three-dimensional angiography image is input into the vascular segmentation model to perform segmentation operations, thereby obtaining a second segmentation mask, a second capsule vector, and a second graph structure, including: The three-dimensional angiography image is input into the convolutional neural network to perform preliminary segmentation to obtain a first feature map and a preliminary segmentation mask, and based on the preliminary segmentation mask, a second capsule vector and a second map structure are determined. The second capsule vector is used as node features, and the second graph structure is used as edge information. The node features and the edge information are then input into the graph neural network for structured reasoning to obtain the structural features of each blood vessel segment. The structural features of each vascular segment are mapped onto all voxels corresponding to that vascular segment to obtain a second feature map. The first feature map and the second feature map are then fused to obtain a fused feature map. The fused feature map is input into the segmentation inference network to perform segmentation inference and obtain the second segmentation mask.

2. The method according to claim 1, characterized in that, The geometric features include the first endpoint, the last endpoint, the center point, the average diameter, and the arc length; the semantic features include the anatomical type. as well as Based on the segmentation mask of each vascular segment, the geometric features of each vascular segment are determined, including: The three-dimensional angiography image is binarized to obtain a binarized angiography image; Based on the binarized angiography image, a skeleton extraction algorithm is used to obtain the skeleton point set of the vascular network; Based on the segmentation mask of each vascular segment and the skeleton point set of the vascular network, the centerline point set of each vascular segment is determined; the geometric features of each vascular segment are calculated based on the centerline point set of each vascular segment.

3. The method according to claim 2, characterized in that, Based on the geometric and semantic features of each vascular segment, the first capsule vector and the first graph structure of all vascular segments are constructed. The geometric and semantic features of each vascular segment are structured and encoded to obtain the first capsule vector of all vascular segments. Based on the first and last endpoints of each vascular segment, calculate the distance between the last endpoint of each vascular segment and the first endpoints of other vascular segments; Based on the distance and a predetermined threshold, the connection relationship between each blood vessel segment is determined to obtain the first graph structure.

4. A device for training a blood vessel segmentation model, characterized in that, include: processor; as well as A memory storing program instructions for training a blood vessel segmentation model, which, when executed by a processor, enables the implementation of the method according to any one of claims 1-3.

5. A device for segmenting blood vessels, comprising: processor; A memory storing program instructions for vessel segmentation, which, when executed by the processor, cause the device to perform the following operations: Obtain the three-dimensional angiography image to be segmented; The three-dimensional angiography image is input into the blood vessel segmentation model trained by the method according to any one of claims 1-3 to perform segmentation operations and obtain blood vessel segmentation results.

6. The device according to claim 5, characterized in that, when the program instructions are executed by the processor, the device further performs the following operations: Boundary smoothing and / or isolated vessel segment removal are performed on the vessel segmentation results to obtain the final vessel segmentation results.

7. A computer-readable storage medium storing computer-readable instructions for training a blood vessel segmentation model and / or for segmenting blood vessels, wherein the computer-readable instructions, when executed by one or more processors, implement the operations implemented by the method according to any one of claims 1-3 and / or the apparatus according to any one of claims 5-6.

Images