Cerebrovascular segmentation method and device based on geometric knowledge and topological structure fusion

Abstract

The invention provides a cerebral vessel segmentation method based on geometric knowledge and topological structure fusion, and the method comprises the following steps: S1, inputting a brain image into a vascular skeleton extraction model, and obtaining a vascular skeleton diagram; s2, simplifying a skeleton path of the vascular skeleton diagram to obtain skeleton key points, generating an adjacent matrix, and constructing a sparse structure diagram according to the skeleton key points and the adjacent matrix; and S3, inputting the brain image and the sparse structure diagram into a pre-trained first U-Net model, and segmenting blood vessels in the brain image to obtain a cerebrovascular contour diagram. Through combination of the topological information of the blood vessels and the deep convolutional neural network, the segmentation effect on fine blood vessels and complex blood vessels is improved, and the cerebrovascular segmentation effect is remarkable.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a cerebral blood vessel segmentation method and device based on the fusion of geometric knowledge and topological structure. Background Art

[0002] Accurate segmentation of cerebral vascular structures is crucial for the early diagnosis, treatment planning, and surgical navigation of cerebrovascular diseases. For example, the early detection of diseases such as cerebral aneurysms, vascular stenosis, and stroke relies on automated, refined segmentation of cerebral vascular images. However, vascular segmentation faces numerous challenges due to the complex morphology, small diameters, and large intensity variations of brain vessels. In particular, blurred vessel edges and reduced contrast in lesioned areas further complicate automated segmentation.

[0003] The U-Net model and its variants are often used for blood vessel segmentation due to their excellent performance in medical image segmentation. However, these methods rely on local pixel features, resulting in limited segmentation accuracy when identifying fine blood vessels, and their effectiveness in cerebral vascular segmentation is generally poor.

[0004] To this end, it is urgent to develop an improved vascular segmentation method based on the U-Net model to accurately segment cerebral blood vessels. Summary of the Invention

[0005] (1) Technical issues to be solved

[0006] In view of the problems existing in the above-mentioned technologies, the present invention aims to address them, at least to a certain extent. To this end, the present invention aims to propose a cerebral vascular segmentation method and device based on the fusion of geometric knowledge and topological structure, which improves the segmentation of fine and complex blood vessels and has a significant effect on cerebral vascular segmentation.

[0007] (2) Technical solution

[0008] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0009] In a first aspect, the present invention provides a cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure, comprising the following steps:

[0010] S1. Input the brain image into the vascular skeleton extraction model to obtain the vascular skeleton map;

[0011] S2. Simplify the skeleton path of the vascular skeleton graph to obtain skeleton key points, generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix;

[0012] S3, inputting the brain image and the sparse structure map into the pre-trained first U-Net model, segmenting the blood vessels in the brain image, and obtaining a cerebral blood vessel contour map;

[0013] In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: extract features corresponding to skeleton key nodes from the feature map of the encoder according to the sparse structure map, construct a first sparse feature map based on the features corresponding to the sparse structure map and the skeleton key nodes, perform deep feature extraction on the first sparse feature map through the graph attention network to obtain a second sparse feature map, connect the feature map of the encoder and the second sparse feature map to obtain a fused feature map, perform a convolution operation on the fused feature map to obtain a convolution feature map, input the convolution feature map into the CBAM module, output the attention map, connect the convolution feature map and the attention map to generate a smooth feature map, and connect the smooth feature map to the corresponding layer of the decoder.

[0014] Optionally, the blood vessel skeleton extraction model is a pre-trained second U-Net model.

[0015] Optionally, simplifying the skeleton path of the blood vessel skeleton graph includes: recursively simplifying the skeleton path of the blood vessel skeleton graph using a Douglas-Peucker algorithm.

[0016] Optionally, the attention coefficient α between two skeleton key nodes in the graph attention network g (v i ,v j ) is calculated as:

[0017]

[0018] Where, v i and v j Represent two skeleton key nodes respectively, and is the feature representation of the corresponding skeleton key node, sim represents the similarity of the features between two skeleton key nodes, and softmax is the activation function.

[0019] Optionally, the first U-Net model is a TransU-Net model.

[0020] Optionally, before S1, it also includes:

[0021] A1. Obtain a training dataset, which includes brain images and the true values of the vascular skeletons corresponding to the brain images;

[0022] A2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

[0023] Optionally, the loss function includes:

[0024]

[0025] Where S is the vascular skeleton output by the second U-Net model based on brain images, is the vascular skeleton output by the second U-Net model based on the brain image after affine transformation, and MMSE is the mean square error.

[0026] Optionally, before S1, it also includes:

[0027] B1. Obtain a training dataset, which includes brain images, sparse structure maps, and ground truth values of cerebral vascular contour maps corresponding to the brain images;

[0028] B2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

[0029] Alternatively, the loss function is expressed as:

[0030]

[0031] Where, is the cross entropy loss, is the DICE loss, β1 and β2 are weight coefficients.

[0032] In a second aspect, the present invention provides a cerebral blood vessel segmentation device based on the fusion of geometric knowledge and topological structure, comprising:

[0033] A skeleton extraction module is used to input brain images into the vascular skeleton extraction model to obtain a vascular skeleton map;

[0034] The structure graph generation module is used to simplify the skeleton path of the vascular skeleton graph, obtain skeleton key points, generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix;

[0035] The blood vessel segmentation module is used to input the brain image and the sparse structure map into the pre-trained first U-Net model to segment the blood vessels in the brain image and obtain the cerebral blood vessel contour map;

[0036] In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: extract features corresponding to skeleton key nodes from the feature map of the encoder according to the sparse structure map, construct a first sparse feature map based on the features corresponding to the sparse structure map and the skeleton key nodes, perform deep feature extraction on the first sparse feature map through the graph attention network to obtain a second sparse feature map, connect the feature map of the encoder and the second sparse feature map to obtain a fused feature map, perform a convolution operation on the fused feature map to obtain a convolution feature map, input the convolution feature map into the CBAM module, output the attention map, connect the convolution feature map and the attention map to generate a smooth feature map, and connect the smooth feature map to the corresponding layer of the decoder.

[0037] (3) Beneficial effects

[0038] The beneficial effects of the present invention are:

[0039] The present invention proposes a cerebral vascular segmentation method and device based on the fusion of geometric knowledge and topological structure. This method combines multiple stages, including vascular skeleton extraction, sparse structure graph generation, and geometric information-enhanced image segmentation. During the first U-Net model's brain image segmentation, vascular geometry information is introduced by embedding the sparse structure graph into a skip layer. The graph attention network (GAT) explicitly models the topological structure of cerebral vessels, allowing the network to capture complex geometric features and combine them with local pixel features. This approach reduces distortion caused by resolution scaling and improves the ability to segment fine vessels. By concatenating the encoder's feature map with a second sparse feature map and performing a convolution operation, the sparse topological features and dense convolutional features are effectively fused, enhancing the model's ability to capture cerebral vascular details. Combined with the CBAM attention mechanism, background noise and redundant information are significantly suppressed, reducing the incidence of missegmentation. This method achieves geometric information-enhanced cerebral vascular segmentation, particularly for identifying complex and fine vessels. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention is described with the aid of the following drawings:

[0041] Figure 1 FIG. 4 is a flowchart of processing a brain image according to the cerebral blood vessel segmentation method of Example 1. DETAILED DESCRIPTION

[0042] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure, including the following steps:

[0045] S1. Input the brain image into the vascular skeleton extraction model to obtain the vascular skeleton map.

[0046] Preferably, the blood vessel skeleton extraction model is a pre-trained second U-Net model.

[0047] S2. Simplify the skeleton path of the vascular skeleton graph to obtain skeleton key points and generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix.

[0048] Preferably, simplifying the skeleton path of the blood vessel skeleton graph includes: recursively simplifying the skeleton path of the blood vessel skeleton graph using a Douglas-Peucker algorithm.

[0049] Specifically, in this embodiment, the sparse structure graph constructed based on the skeleton key points and the adjacency matrix is represented as G(V, E, C), where G represents the sparse structure graph, V represents the skeleton key point set, E represents the connection relationship between the skeleton key points, and C represents the image pixel coordinates corresponding to the skeleton key points.

[0050] S3. Input the brain image and the sparse structure map into the pre-trained first U-Net model, segment the blood vessels in the brain image, and obtain a cerebral blood vessel contour map.

[0051] In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: the features corresponding to the skeleton key nodes are extracted from the feature map of the encoder according to the sparse structure map, the first sparse feature map is constructed according to the features corresponding to the sparse structure map and the skeleton key nodes, the first sparse feature map is deep-featured extracted through the graph attention network GAT to obtain the second sparse feature map, the feature map of the encoder and the second sparse feature map are connected to obtain a fused feature map, a convolution operation is performed on the fused feature map to obtain a convolution feature map, the convolution feature map is input into the CBAM module, the attention map is output, the convolution feature map and the attention map are connected to generate a smooth feature map, and the smooth feature map is connected to the corresponding layer of the decoder.

[0052] In the first U-Net model's processing of brain images, vascular geometry information was introduced by embedding a sparse structure graph into the skip layer. The graph attention network (GAT) explicitly modeled the topological structure of brain vessels, allowing the network to capture complex geometric features and combine them with local pixel features. This approach reduced distortion caused by resolution scaling and improved the ability to segment fine vessels. By concatenating the encoder's feature map with the second sparse feature map and performing a convolution operation, the sparse topological features and dense convolutional features were effectively fused, enhancing the model's ability to capture detailed brain vascular information. Combined with the CBAM attention mechanism, background noise and redundant information were significantly suppressed, reducing the incidence of missegmentation. This results in geometrically enhanced brain vascular segmentation, with particular excellence in identifying complex and fine vessels.

[0053] Preferably, the attention coefficient α between two skeleton key nodes in the graph attention network g (v i ,v j ) is calculated as:

[0054]

[0055] Where, v i and v j Represent two skeleton key nodes respectively, and is the feature representation of the corresponding skeleton key node, sim represents the similarity of the features between two skeleton key nodes, and softmax is the activation function.

[0056] Preferably, in this embodiment, performing a convolution operation on the fused feature map includes: performing a spatially separable convolution operation on the fused feature map.

[0057] Preferably, the first U-Net model is a TransU-Net model. The TransU-Net model combines multi-head self-attention with convolution operations to achieve global dependency modeling and local detail capture of blood vessels, thereby improving the effect of blood vessel segmentation.

[0058] In summary, this embodiment constructs a multi-stage brain vascular segmentation process, combining multiple stages of vascular skeleton extraction, sparse structure map generation, and geometric information enhanced image segmentation. It effectively solves the problem of complex microvascular structure recognition and processing in brain vascular segmentation, and performs outstandingly in the recognition and segmentation of microvessels in high-density vascular areas.

[0059] Specifically, before S1, cerebral vascular segmentation methods based on the fusion of geometric knowledge and topological structure also include:

[0060] A1. Obtain a training dataset, which includes brain images and true values of vascular skeletons corresponding to the brain images.

[0061] Specifically, obtaining the true value of the vascular skeleton corresponding to the brain image includes: extracting the binary segmentation result of the brain image using a traditional image processing algorithm to obtain the true value of the vascular skeleton. Preferably, the Hilditch operator is used to extract the binary segmentation result of the brain image to obtain the true value of the vascular skeleton.

[0062] A2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

[0063] Preferably, the loss function includes:

[0064]

[0065] Where S is the vascular skeleton output by the second U-Net model based on brain images, This is the vascular skeleton output by the second U-Net model based on the affine-transformed brain image. MMSE is the mean squared error. By introducing structural consistency regularization during training, the affine transformation improves the robustness of the network. This method significantly improves the accuracy and stability of skeleton extraction, especially when dealing with different postures or complex structures of brain vessels.

[0066] Specifically, before S1, cerebral vascular segmentation methods based on the fusion of geometric knowledge and topological structure also include:

[0067] B1. Obtain a training dataset, which includes brain images, sparse structure maps, and true values of cerebral vascular contour maps corresponding to the brain images.

[0068] B2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

[0069] Preferably, the loss function includes a cross entropy loss function and a DICE coefficient loss function, thereby improving the blood vessel segmentation accuracy of the second U-Net model.

[0070] Further preferably, the loss function is expressed as:

[0071]

[0072] Where, is the cross entropy loss, is the DICE loss, β1 and β2 are weight coefficients. By combining the cross entropy loss With DICE loss The joint loss function optimizes the network segmentation performance and further improves the blood vessel segmentation accuracy of the second U-Net model.

[0073] Experimental results show that the cerebral vascular segmentation method provided by this embodiment outperforms existing technologies in key indicators such as accuracy, sensitivity, and F1 score, and is particularly outstanding in identifying small blood vessels and improving vascular connectivity. By introducing geometric knowledge into the skip layer module, segmentation accuracy and overall structural integrity are significantly improved. In addition, this method effectively reduces missegmentation and breakage when processing complex vascular regions, demonstrating its strong adaptability in the segmentation of high-density vascular regions. In summary, the present invention provides an efficient and accurate solution for cerebral vascular segmentation and has broad application potential.

[0074] Example 2

[0075] This embodiment provides a cerebral blood vessel segmentation device based on the fusion of geometric knowledge and topological structure, including:

[0076] The skeleton extraction module is used to input the brain image into the vascular skeleton extraction model to obtain the vascular skeleton map.

[0077] The structure graph generation module is used to simplify the skeleton path of the vascular skeleton graph, obtain skeleton key points, generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix.

[0078] The blood vessel segmentation module is used to input the brain image and the sparse structure map into the pre-trained first U-Net model, segment the blood vessels in the brain image, and obtain the cerebral blood vessel contour map.

[0079] In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: extract features corresponding to skeleton key nodes from the feature map of the encoder according to the sparse structure map, construct a first sparse feature map based on the features corresponding to the sparse structure map and the skeleton key nodes, perform deep feature extraction on the first sparse feature map through the graph attention network to obtain a second sparse feature map, connect the feature map of the encoder and the second sparse feature map to obtain a fused feature map, perform a convolution operation on the fused feature map to obtain a convolution feature map, input the convolution feature map into the CBAM module, output the attention map, connect the convolution feature map and the attention map to generate a smooth feature map, and connect the smooth feature map to the corresponding layer of the decoder.

[0080] This embodiment constructs a multi-module brain vascular segmentation process, combining multiple modules such as a skeleton extraction module, a structure graph generation module, and a geometric information enhanced vascular segmentation module. It effectively solves the problem of complex structure processing in brain vascular segmentation, especially in high-density vascular areas.

[0081] It should be noted that the specific functions of each module in the cerebral blood vessel segmentation device based on the fusion of geometric knowledge and topological structure provided in this embodiment, as well as the processing flow of cerebral blood vessel segmentation, can be referred to the detailed description of the cerebral blood vessel segmentation processing method provided in the above-mentioned embodiment 1, and will not be repeated here.

[0082] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.

[0084] It should be noted that, in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third etc. is for convenience only and does not indicate any order. These words may be understood as part of the component name.

[0085] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0086] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments after learning the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0087] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.

Claims

1. A cerebral vascular segmentation method based on the fusion of geometric knowledge and topological structure, characterized in that: The following steps are involved: S1. Input the brain image into the vascular skeleton extraction model to obtain the vascular skeleton map; S2. Simplify the skeleton path of the vascular skeleton graph to obtain skeleton key points, generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix; S3, inputting the brain image and the sparse structure map into the pre-trained first U-Net model, segmenting the blood vessels in the brain image, and obtaining a cerebral blood vessel contour map; In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: extract features corresponding to skeleton key nodes from the feature map of the encoder according to the sparse structure map, construct a first sparse feature map based on the features corresponding to the sparse structure map and the skeleton key nodes, perform deep feature extraction on the first sparse feature map through the graph attention network to obtain a second sparse feature map, connect the feature map of the encoder and the second sparse feature map to obtain a fused feature map, perform a convolution operation on the fused feature map to obtain a convolution feature map, input the convolution feature map into the CBAM module, output the attention map, connect the convolution feature map and the attention map to generate a smooth feature map, and connect the smooth feature map to the corresponding layer of the decoder.

2. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 1 is characterized in that: The vascular skeleton extraction model is a pre-trained second U-Net model.

3. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 1 is characterized in that: The skeleton path of the blood vessel skeleton graph is simplified, including: recursively simplifying the skeleton path of the blood vessel skeleton graph using a Douglas-Peucker algorithm.

4. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 1, characterized in that: Attention coefficient α between two skeleton key nodes in graph attention network g (v i ,v j ) is calculated as: Where, v i and v j Represent two skeleton key nodes respectively, and is the feature representation of the corresponding skeleton key node, sim represents the similarity of the features between two skeleton key nodes, and softmax is the activation function.

5. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 1, characterized in that: The first U-Net model is the TransU-Net model.

6. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 2, characterized in that: Before S1, it also included: A1. Obtain a training dataset, which includes brain images and the true values of the vascular skeletons corresponding to the brain images; A2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

7. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 6, characterized in that: The loss functions include: Where S is the vascular skeleton output by the second U-Net model based on brain images, is the vascular skeleton output by the second U-Net model based on the brain image after affine transformation, and MMSE is the mean square error.

8. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 1, characterized in that: Before S1, it also included: B1. Obtain a training dataset, which includes brain images, sparse structure maps, and ground truth values of cerebral vascular contour maps corresponding to the brain images; B2. Train the second U-Net model based on the training data set to minimize the preset loss function and obtain a vascular skeleton extraction model.

9. The cerebral blood vessel segmentation method based on the fusion of geometric knowledge and topological structure according to claim 8, characterized in that: The loss function is expressed as: Where, is the cross entropy loss, is the DICE loss, β1 and β2 are weight coefficients.

10. A cerebral blood vessel segmentation device based on the fusion of geometric knowledge and topological structure, characterized in that: include: A skeleton extraction module is used to input brain images into the vascular skeleton extraction model to obtain a vascular skeleton map; The structure graph generation module is used to simplify the skeleton path of the vascular skeleton graph, obtain skeleton key points, generate an adjacency matrix, and construct a sparse structure graph based on the skeleton key points and the adjacency matrix; The blood vessel segmentation module is used to input the brain image and the sparse structure map into the pre-trained first U-Net model to segment the blood vessels in the brain image and obtain the cerebral blood vessel contour map; In the first U-Net model, the encoder extracts feature maps based on brain images; the processing process of the jump layer is as follows: extract features corresponding to skeleton key nodes from the feature map of the encoder according to the sparse structure map, construct a first sparse feature map based on the features corresponding to the sparse structure map and the skeleton key nodes, perform deep feature extraction on the first sparse feature map through the graph attention network to obtain a second sparse feature map, connect the feature map of the encoder and the second sparse feature map to obtain a fused feature map, perform a convolution operation on the fused feature map to obtain a convolution feature map, input the convolution feature map into the CBAM module, output the attention map, connect the convolution feature map and the attention map to generate a smooth feature map, and connect the smooth feature map to the corresponding layer of the decoder.

Images