Unsupervised multi-view automatic segmentation method, device, medium and product for cerebral vessels

Abstract

This application discloses an unsupervised multi-view automatic segmentation method, device, medium, and product for cerebral blood vessels, relating to the field of image segmentation. The method includes: constructing a Frangi-filtered view group for enhanced images based on multimodal magnetic resonance three-dimensional images (such as T1WI, TOF-MRA, etc.) using a multi-scale Frangi-filtering method; constructing a Sato-filtered view group using a multi-scale Sato-filtering method; constructing a diffusion-enhanced view group using a multi-scale vascular diffusion enhancement method; and generating a vascular classification result view group using a Dirichlet process mixture model. A multi-view fusion method based on atlas segmentation is then used to fuse these view groups to obtain a segmented cerebral blood vessel image, and the segmentation result is optimized through post-processing. This application can obtain accurate and stable cerebral blood vessel segmentation images unsupervised and can also be applied to magnetic resonance angiography images with different modalities and scanning parameters, exhibiting strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image segmentation, and in particular to a brain vessel unsupervised multi-view automatic segmentation method, device, medium and product. BACKGROUND

[0002] Brain vessels, as the energy source of the central nervous system, play an important role in brain function homeostasis. Brain vessel aging is a key link of brain aging and an important factor causing neurodegenerative diseases. The development of magnetic resonance angiography technology has realized the non-invasive and repeated observation of individual brain vessel structure in clinical research. However, due to the complexity of the brain vessel network structure, in actual work, radiologists usually make subjective qualitative judgments on the magnetic resonance vessel image based on professional knowledge and experience, which leads to the fact that the population-level research and clinical precise diagnosis of brain vessel structure are still limited at the present stage.

[0003] At present, the three-dimensional segmentation and reconstruction method of brain vessels can be roughly divided into two categories. One is an unsupervised machine learning method, and its representative methods are Frangi filtering and Sato filtering. These methods are based on the understanding of a single feature of brain vessel signals, and often cannot accurately and stably identify brain vessel structures under multiple scales and multiple modalities. The other is a method based on deep neural networks. Due to the complexity of the brain vessel network structure and the lack of high-quality standardized brain vessel multi-modal magnetic resonance annotation data sets, at present, there is still a lack of high-quality standardized brain vessel multi-modal magnetic resonance annotation data sets. In addition, due to the limitations of model architecture design and magnetic resonance angiography principles, the existing methods generally have poor segmentation and recognition effects on small blood vessels. In addition, due to the semi-quantitative nature of magnetic resonance imaging (i.e., signal intensity is affected by equipment and parameters rather than absolute value), different scanning equipment and scanning parameter settings of different centers will cause significant changes in image contrast and quality, leading to artifacts of magnetic resonance angiography signals and non-uniformity of different scale vessel signals. Therefore, due to the lack of robustness in different scenarios, the application and promotion of many existing methods are limited. SUMMARY

[0004] In order to solve the above problems, the present application provides a brain vessel unsupervised multi-view automatic segmentation method, device, medium and product.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] In a first aspect, the present application provides a brain vessel unsupervised multi-view automatic segmentation method, comprising:

[0007] Acquire multimodal 3D magnetic resonance images of the brain; the multimodal 3D magnetic resonance images include T1-weighted imaging (T1WI) and time-of-flight magnetic resonance angiography (TOF-MRA).

[0008] Based on the multimodal magnetic resonance three-dimensional image, a Frangi filter view group is constructed using the multi-scale Frangi filter method.

[0009] A Sato filter view group is constructed based on the multimodal magnetic resonance three-dimensional image using a multi-scale Sato filter method.

[0010] A multi-scale vascular diffusion enhancement method was used to construct a multi-scale vascular diffusion enhancement view group based on the multimodal magnetic resonance three-dimensional images;

[0011] A Dirichlet process mixture model was used to construct a vascular classification result perspective group based on the multimodal magnetic resonance three-dimensional images;

[0012] A multi-view fusion method based on map segmentation is used to fuse the Frangi filter view group, the Sato filter view group, the multi-scale vascular diffusion enhancement view group, and the vascular classification result view group. After post-processing optimization, a cerebral vascular segmentation image is obtained.

[0013] Optionally, a Frangi filter view group is constructed based on the multimodal magnetic resonance three-dimensional image using a multi-scale Frangi filter method, including:

[0014] In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points;

[0015] Perform eigenvalue decomposition on the Hessian matrix and determine the absolute values ​​of the eigenvalues ​​obtained from the decomposition.

[0016] The eigenvalues ​​are sorted in descending order according to their absolute values ​​to obtain a sequence of absolute values ​​of the eigenvalues.

[0017] The vascular enhancement response value is constructed based on the absolute value of all feature values ​​in the absolute value sequence of feature values, and the probability information of the existence of tubular continuum structure at different scales of each voxel point in the multimodal magnetic resonance three-dimensional image space is determined based on the vascular enhancement response value.

[0018] Based on probability information, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale, and the three-dimensional blood vessel probability distribution matrix is ​​used as the Frangi filter view at the corresponding scale.

[0019] The Frangi filter perspective group is formed based on the Frangi filter perspectives at all scales.

[0020] Optionally, a Sato filter view group is constructed based on the multimodal magnetic resonance three-dimensional image using a multi-scale Sato filter method, including:

[0021] In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points;

[0022] The Hessian matrix is ​​decomposed into eigenvalues, and the eigenvalues ​​are sorted to form an eigenvalue sequence.

[0023] Sato vascular enhancement response values ​​are constructed based on all feature values ​​in the feature value sequence, and the Sato probability value of each voxel point in the multimodal magnetic resonance three-dimensional image space having a tubular continuum structure at different scales is determined based on the Sato vascular enhancement response values.

[0024] Based on the Sato probability value, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale, and the three-dimensional blood vessel probability distribution matrix is ​​used as the Sato filtering perspective at the corresponding scale.

[0025] The Sato filter perspective group is formed based on the Sato filter perspectives at all scales.

[0026] Optionally, a multi-scale vascular diffusion enhancement view group is constructed based on the multimodal magnetic resonance three-dimensional image using a multi-scale vascular diffusion enhancement method, including:

[0027] In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points;

[0028] Based on the Hessian matrix, different scales are selected, and smoothing and enhancement processing are performed based on the Frangi filtering results to establish vascular response functions at different scales.

[0029] A three-dimensional diffusion tensor is constructed using the vascular response function and the Hessian matrix;

[0030] The three-dimensional diffusion tensor is used to perform specific diffusion enhancement along the blood vessel axis. After a preset diffusion iteration time, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale. The three-dimensional blood vessel probability distribution matrix is ​​used as the multi-scale blood vessel diffusion enhancement perspective at the corresponding scale.

[0031] The multi-scale vascular diffusion enhancement perspective group is formed based on multi-scale vascular diffusion enhancement perspectives at all scales.

[0032] Optionally, the vascular response function at different scales can be expressed as:

[0033] ;

[0034] In the formula, exp() represents the exponential function. , , and All of these are hyperparameters. , and These are all eigenvalues ​​of the Hessian matrix, and e is the natural logarithm. Used to distinguish between disc-shaped and linear structures Used to identify spherical structures = Indicates structural strength.

[0035] Optionally, a Dirichlet process mixture model is used to construct a vascular classification result perspective group based on the multimodal magnetic resonance three-dimensional images, including:

[0036] The Dirichlet distribution is obtained by random sampling from the Dirichlet process, and the probability density function of the mixture distribution is generated.

[0037] Categorical variables are sampled from a mixed distribution, and voxel signal values ​​are sampled from the corresponding multivariate Gaussian distribution based on the categorical variables; the categorical variables are used as cluster labels.

[0038] The Gibbs sampling method is used to obtain blood vessel classification results based on the probability density function and the voxel signal value;

[0039] The vascular classification result perspective group is constructed based on the vascular classification results.

[0040] Optionally, a multi-view fusion method based on atlas segmentation is used to fuse the Frangi filter view group, the Sato filter view group, the multi-scale vascular diffusion enhancement view group, and the vascular classification result view group to obtain a cerebral vascular segmentation image, including:

[0041] For each perspective blood vessel segmentation image in the Frangi filter perspective group, the Sato filter perspective group, the multi-scale blood vessel diffusion enhancement perspective group, and the blood vessel classification result perspective group, a similarity matrix between voxels is constructed based on a Gaussian kernel function using a sparse representation method.

[0042] A unified similarity matrix is ​​obtained based on the similarity matrix using a weighted average method. ;

[0043] Based on the unified similarity matrix Determination matrix And based on a unified similarity matrix Sum-degree matrix Determine the graph Laplacian matrix; the graph Laplacian matrix is ; In the formula, For degree matrix elements, To unify the elements of the similarity matrix, and In this formula, the variable is represented. and Indicates the row and column indices of the matrix;

[0044] Perform eigenvalue decomposition on the graph Laplacian matrix, and sort the resulting eigenvalues ​​in descending order, retaining the first few eigenvalues. Each eigenvalue and its relation to the previous The feature vectors corresponding to each feature value are used to adjust the sensitivity of blood vessel recognition, thereby obtaining the brain blood vessel segmentation image; The number of eigenvalues. For parameters.

[0045] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described unsupervised multi-view automatic segmentation method for cerebral blood vessels.

[0046] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described unsupervised multi-view automatic segmentation method for cerebral blood vessels.

[0047] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described unsupervised multi-view automatic segmentation method for cerebral blood vessels.

[0048] According to the specific embodiments provided in this application, this application has the following technical effects:

[0049] This application provides an unsupervised multi-view automatic segmentation method, device, medium, and product for cerebral blood vessels. It comprehensively considers multiple essential common features of cerebral blood vessel tissue in multimodal magnetic resonance images. Starting from the continuity of cerebral blood vessel signals at different scales, differences in signal distribution, and tubular geometric features, it selects and constructs different feature extraction methods (multi-scale Frangi filtering method, multi-scale Sato filtering method, multi-scale vascular diffusion enhancement method, and Dirichlet process hybrid model). It adopts a multi-view fusion method to integrate cross-scale segmentation information from multiple feature perspectives. Without relying on large-scale labeled data, it merges them into a unified segmentation result. The information between different perspectives can interact and corroborate each other, so that the final cerebral blood vessel segmentation image has the advantages of high accuracy and strong generalization. It effectively solves the problems of inaccurate recognition of multi-scale and multimodal cerebral blood vessel structures, poor segmentation effect of small blood vessels, and insufficient robustness caused by signal artifacts and inconsistencies in existing methods. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating an unsupervised multi-view automatic segmentation method for cerebral blood vessels provided in an embodiment of this application;

[0052] Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0053] 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 embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0054] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] In one exemplary embodiment, this application provides an unsupervised multi-view automatic segmentation method for cerebral blood vessels. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the application of this method on a server is used as an example. Figure 1 As shown, the method includes:

[0056] Step 100: Acquire multimodal 3D magnetic resonance images of the brain. These images include T1-weighted imaging and TOF-MRA images.

[0057] Step 101: Construct a Frangi filter view group based on multimodal magnetic resonance three-dimensional images using the multi-scale Frangi filtering method.

[0058] Step 102: Construct a Sato filter view group based on multimodal magnetic resonance three-dimensional images using the multi-scale Sato filtering method.

[0059] Step 103: Using a multi-scale vascular diffusion enhancement method, a multi-scale vascular diffusion enhancement view group is constructed based on multi-modal magnetic resonance three-dimensional images.

[0060] Step 104: Using the Dirichlet process mixture model based on multimodal magnetic resonance three-dimensional images, construct a perspective group for vascular classification results.

[0061] Step 105: A multi-view fusion method based on map segmentation is used to fuse the viewpoints of the Frangi filter viewpoint group, the Sato filter viewpoint group, the multi-scale vascular diffusion enhancement viewpoint group, and the vascular classification result viewpoint group. The cerebral vascular segmentation image is obtained through post-processing optimization.

[0062] In another exemplary embodiment of this application, the cerebral vascular signals in the three-dimensional images of multimodal magnetic resonance imaging (T1-weighted imaging images and TOF-MRA magnetic resonance images) all exhibit tubular continuum structures in three-dimensional space. Based on this essential characteristic of cerebral blood vessels, this application employs a multiscale Frangi filtering method to enhance and extract cerebral vascular signals to form segmented Frangi filter view groups. Therefore, in this embodiment, the implementation process of step 101 includes:

[0063] 11. In the three-dimensional image space of multimodal magnetic resonance imaging, obtain voxel points and construct a voxel-level Hessian matrix based on the voxel points. For example:

[0064] Calculate the second-order partial derivatives of voxel points in the three-dimensional image space of multimodal magnetic resonance imaging, and construct the voxel-level Hessian matrix, expressed as:

[0065] .

[0066] In the formula, It is a Hessian matrix. Given the spatial coordinates of a voxel point, To define at coordinates A multivariable function in three-dimensional space, This is the partial derivative operator.

[0067] 12. Perform eigenvalue decomposition on the Hessian matrix and determine the absolute values ​​of the eigenvalues ​​obtained. The eigenvalue decomposition formula is as follows:

[0068] .

[0069] In the formula, Λ is a diagonal matrix, and its diagonal elements are the eigenvalues ​​of the Hessian matrix, arranged in descending order of absolute value. , , ...... This is the eigenvector matrix, where each column vector represents the eigenvector corresponding to the respective eigenvalue. This represents the transpose of the eigenvector.

[0070] Take the absolute value of the eigenvalues ​​of the Hessian matrix and label it as . , , ...When taking the three eigenvalues ​​with the largest absolute values, we have:

[0071] .

[0072] 13. Sort the eigenvalues ​​in descending order of their absolute values ​​to obtain a sequence of absolute values ​​of the eigenvalues.

[0073] 14. Based on the absolute value sequence of eigenvalues, the first 3 ( , , The absolute values ​​corresponding to the eigenvalues ​​are used to calculate the vascular enhancement response value, and based on this response value, the probability information of the existence of tubular continuum structures at different scales for each voxel point in the multimodal magnetic resonance three-dimensional image space is determined. Among these, the structural strength statistics... The calculation formula is:

[0074] , , .

[0075] In the formula, Represents the Hessian matrix The Euclidean norm, This is a statistic used to distinguish between disc-shaped and linear structures. This is a statistic used to identify spherical structures.

[0076] The formula for calculating the probability information of tubular continuum structures is:

[0077] .

[0078] Among them, three-dimensional vector = Represents a voxel in three-dimensional space. This represents the probability value of blood vessels on a voxel. , , and These are all hyperparameters that need to be preset and are used to control the sensitivity of the Frangi filter to spherical, tubular, and disk-shaped structures and background noise. = Indicates structural strength.

[0079] 15. Based on probability information, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale, and the three-dimensional blood vessel probability distribution matrix is ​​used as the Frangi filter view at the corresponding scale.

[0080] The classic Frangi filter method for segmenting blood vessels, when handling cross-scale vessel integration and recognition, employs a simple strategy of maximizing the probability, leading to significant information loss. To address this issue, this embodiment adopts a two-step strategy of cross-scale filtering and fusion separation, preserving probability information at each scale and forming Frangi filter viewpoint groups according to different filtering scales. The specific steps for generating the Frangi filter viewpoint group images are as follows:

[0081] (1) Compare the original image (i.e., the multimodal magnetic resonance three-dimensional image) with the scale of s The convolution operation is performed using a 3D Gaussian convolution kernel, and the specific operation process is as follows:

[0082] .

[0083] in, The result of the convolution operation. The scale is The result of the convolution operation of the three-dimensional Gaussian convolution kernel function. To act on voxels The scale above is The three-dimensional Gaussian convolution kernel function has the following specific form:

[0084] .

[0085] in, The dimension of the Gaussian function, for example When =3, it represents three-dimensional space.

[0086] (2) Change the scale At a given scale List ( 1, 2,… Probabilistic calculations for tubular continuum structures are performed under the condition of […]. Each scale... A three-dimensional blood vessel probability distribution matrix with the same dimensions as the original image is generated, which serves as the Frangi filter perspective at this scale.

[0087] 16. Based on the Frangi filter perspectives at all scales, a Frangi filter perspective group is formed for use in multi-view fusion.

[0088] In another exemplary embodiment of this application, step 102 is mainly for constructing a second-view image of blood vessel segmentation using a multi-scale Sato filtering method. Specifically, after performing a convolution operation on the original image using a three-dimensional Gaussian convolution kernel function of scale s [see step (1) above], the implementation steps of step 102 include:

[0089] 21. In the space of multimodal magnetic resonance three-dimensional images (T1-weighted imaging images and TOF-MRA magnetic resonance images), voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points. The specific process is the same as step 11 above.

[0090] 22. Perform eigenvalue decomposition on the Hessian matrix and sort the eigenvalues ​​to form an eigenvalue sequence. Before sorting, do not take the absolute value of the eigenvalues; instead, sort them directly considering their signs. We have:

[0091] .

[0092] 23. Based on the eigenvalue sequence ( , , The Sato probability value is calculated using eigenvalues, and based on this probability value, the probability information of the existence of a tubular continuum structure at different scales for each voxel in the three-dimensional image space of multimodal magnetic resonance is determined. The formula for calculating the Sato probability value of the tubular continuum structure is as follows:

[0093] .

[0094] In the formula, Representing coordinates The probability value of blood vessels at a voxel point. The hyperparameter in this formula. Once preset, it is used to control the sensitivity of the Sato filter to background noise.

[0095] 24. Based on the Sato probability value, generate a three-dimensional vascular probability distribution matrix at each scale that is the same dimension as the three-dimensional image of the multimodal magnetic resonance (T1-weighted imaging image and TOF-MRA magnetic resonance image), and use the three-dimensional vascular probability distribution matrix as the Sato filtering perspective at the corresponding scale.

[0096] The classic Sato filtering (Multiscale Sato Filtering) method for segmenting blood vessels, when handling cross-scale vessel integration and recognition, suffers from significant information loss due to its simple strategy of maximizing probabilities. To address this issue, this embodiment also employs a two-step strategy of cross-scale filtering and fusion separation in the Sato viewpoint construction process. This retains the probability information at each scale, forming Sato filter viewpoint groups according to different filtering scales for later unified fusion. Based on this, the specific workflow is as follows:

[0097] Select a series of different Gaussian kernel scales ( 1, 2,… At their respective scales, according to the Sato probability value calculation formula for the tubular continuum structure described above, each scale... A three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated as the Sato filter perspective at this scale.

[0098] 25. A Sato filter perspective group is formed based on the Sato filter perspective at all scales.

[0099] In another exemplary embodiment of this application, the cerebral vascular structure has obvious continuity characteristics. Based on this characteristic, this application employs a multiscale vessel enhancement diffusion method, using a Hessian matrix to construct a three-dimensional diffusion tensor, allowing the vessel segmentation label to diffuse and grow along the axial direction of the vessel, suppressing the vertical extension of the label, and ensuring the continuity of the label during vessel segmentation. Based on this, the implementation process of step 103 includes:

[0100] 31. In the three-dimensional image space of multimodal magnetic resonance (T1-weighted imaging image and TOF-MRA magnetic resonance image), voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points. The specific process is the same as step 11 above.

[0101] 32. Based on the Hessian matrix, different scales are selected, and the Frangi filter provided above is smoothed and enhanced to establish vascular response functions at different scales, thereby significantly reducing noise and interference from non-vascular structures. The vascular response functions at different scales are expressed as follows:

[0102] .

[0103] In the formula, exp() represents the exponential function. Here, the hyperparameters... , , and After being preset, it is used to control the smoothness of the vascular response function. e It is the natural logarithm.

[0104] 33. Construct a three-dimensional diffusion tensor using the vascular response function and the Hessian matrix. Wherein:

[0105] (1) Perform eigenvalue decomposition on the Hessian matrix to obtain the eigenvector matrix. .

[0106] (2) Construct the diffusion eigenvalue matrix based on the vascular response function. The matrix The elements on the diagonal are:

[0107] .

[0108] .

[0109] In the formula, For vascular response function, , and Both represent matrices Elements on the diagonal. and As hyperparameters, they need to be preset to control the intensity and speed of diffusion in different directions, as well as the sensitivity to vascular tissue.

[0110] 34. A three-dimensional diffusion tensor is used for axial-specific diffusion of labeled blood vessels. After a preset time t, a distribution matrix with the same dimension as the multimodal MRI three-dimensional image is generated at each scale, serving as the multi-scale vascular diffusion enhancement perspective at the corresponding scale. The diffusion formula is expressed as:

[0111] .

[0112] In the formula, This is a Nabla operator, and its output is a vector consisting of the partial derivatives of the multivariate function with respect to each variable. Indicates diffusion time. The result of the convolution operation mentioned earlier is the original image after being processed by a process with a scale of [value missing]. The processed image obtained by performing a convolution operation on a 3D Gaussian convolution kernel is represented as a multivariate function in 3D space; the dimension of the Gaussian function is... In the above formula, represents a three-dimensional diffusion tensor, determined by the vascular response function and the Hessian matrix. This represents the divergence operation, and the output is the average rate of change of the local signal in the three-dimensional space of the convolved image.

[0113] Select a series of different Gaussian kernel scales ( 1, 2,… ), at their respective scales, according to the above diffusion formula, after After time diffusion, at each scale A three-dimensional vascular probability distribution matrix of the same dimension as the multimodal magnetic resonance three-dimensional image is generated, which serves as the multi-scale vascular diffusion enhancement perspective at this scale.

[0114] 35. Based on the multi-scale vascular diffusion enhancement perspective at all scales, a multi-scale vascular diffusion enhancement perspective group is constructed.

[0115] In another exemplary embodiment of this application, another fundamental characteristic of cerebral vascular tissue is that it has a relatively stronger signal in magnetic resonance angiography images compared to non-vascular tissue, forming signal contrast. Based on this characteristic of cerebral vascular signals, this application performs unsupervised clustering analysis based on the signal intensity values ​​of all voxels to identify vascular tissue. However, hyperparameters in unsupervised clustering analysis (such as the number of categories) can significantly affect the clustering effect. To address this issue, this application utilizes a Dirichlet Process Mixture Model to adaptively determine the number of categories based on the image's inherent properties, thereby achieving optimal clustering. The input information of the Dirichlet Process Mixture Model is the magnetic resonance signal intensity value of each voxel in the TOF-MRA magnetic resonance angiography image. Based on this, the implementation process of step 104 provided above may include:

[0116] 41. Obtain the Dirichlet distribution by random sampling from a Dirichlet process, and generate the probability density function of the Dirichlet distribution. For example, from a process with hyperparameters of... The base distribution is The Dirichlet distribution is obtained through random sampling during the Dirichlet process. ,have:

[0117] .

[0118] In the formula, This indicates a random sampling operation.

[0119] Dirichlet distribution obtained from sampling The probability density function (with respect to categorical variables) )for:

[0120] .

[0121] In the formula, Let be the probability density function. For about Beta distribution, for Dimensions For the first Dirichlet distribution The corresponding hyperparameters.

[0122] 42. Randomly sample the values ​​of the categorical variable from the Dirichlet distribution (denoted as ). This value serves as a parameter for the multivariate distribution of each voxel, and is derived from the parameter... The distribution of voxel signals generated in a multivariate distribution.

[0123] 43. The Gibbs sampling method (which has been implemented in the PyMC Python package) is used for iterative calculations to obtain blood vessel classification results based on the probability density function and voxel signal values.

[0124] 44. Construct a three-dimensional classification probability map based on the vessel classification results, and build a perspective group for the vessel classification results. The vessel classification results obtained by the Dirichlet process mixture model are used as the fourth perspective for vessel segmentation.

[0125] In another exemplary embodiment of this application, based on the prior knowledge of different basic signal features of cerebral blood vessels in magnetic resonance images, four cerebral blood vessel enhancement segmentation view groups are constructed (corresponding to the view groups constructed in steps 101, 102, 103, and 104). Based on this, this application utilizes a multi-view fusion method based on atlas segmentation to fuse the view images of the four cerebral blood vessel enhancement segmentation view groups, ultimately constructing a unified three-dimensional cerebral blood vessel segmentation image (i.e., the final cerebral blood vessel segmentation image obtained in this application). 51. In the cerebral blood vessel segmentation images of each view group (Frangi filter view group, Sato filter view group, multi-scale vascular diffusion enhancement view group, and vascular classification result view group) using the multi-view fusion method, a similarity matrix between voxels in each view group is constructed using a sparse representation method.

[0126] 52. Based on the similarity matrix between voxels from various perspectives Obtain the unified similarity matrix ,

[0127] .

[0128] In the formula, Representing the Viewpoint voxels and voxels Similarity between them Representing the Viewpoint voxels and voxels The similarity between them. , Representing the Viewpoint voxels and voxels The probability value of blood vessels, Representing the In the similarity matrix of perspectives, the first The similarity value of each element. The total number of viewpoints. The number of voxels.

[0129] 53. Determine the degree matrix based on the unified similarity matrix, and then determine the graph Laplacian matrix based on the unified similarity matrix and the degree matrix. The graph Laplacian matrix is: ; In the formula, For degree matrix elements, To unify the elements of the similarity matrix, and In this formula, the variable is represented. and Indicates the row and column numbers of the matrix.

[0130] For example, considering the correlation between information from different perspectives, the similarity matrices of each perspective are weighted and summed to obtain a unified similarity matrix. ,for:

[0131] .

[0132] In the formula, Representing the A similarity matrix of perspectives Representative voxels Uniform similarity, Representative voxels and voxels The uniform similarity. Weight values ​​for each perspective. Similarity matrix from this perspective With the unified similarity matrix The norm between them is determined by:

[0133] .

[0134] In the formula, Representative norm.

[0135] 54. Perform eigenvalue decomposition on the graph Laplacian matrix, sort the resulting eigenvalues ​​in descending order, and retain the first few eigenvalues. By using eigenvalues ​​and the corresponding eigenvectors of the retained eigenvalues ​​to adjust the sensitivity of blood vessel recognition, a brain blood vessel segmentation image is obtained. The number of eigenvalues. The parameter is . The final result of cerebral blood vessel segmentation is expressed as:

[0136] .

[0137] In the formula, The final result of cerebral blood vessel segmentation, Indicates intermediate parameters.

[0138] Based on the above description, the unsupervised multi-view automatic segmentation method for cerebral blood vessels provided in this application can achieve accurate and stable identification and segmentation of blood vessels in multimodal magnetic resonance angiography data under unlabeled conditions. This advantage stems from the multi-view fusion unsupervised machine learning strategy employed in this application. It combines prior knowledge with mathematical modeling, comprehensively considering multiple essential common features of cerebral vascular tissue in multimodal magnetic resonance images. Starting from the continuity of cerebral vascular signals at different scales, differences in signal distribution, and tubular geometric features, different feature extraction methods are selected and constructed. A multi-view fusion method is used to integrate cross-scale segmentation information from multiple feature perspectives, merging them into a unified segmentation result. Information from each perspective can interactively corroborate each other. Therefore, this segmentation result has the advantages of accuracy, strong generalization, and independence from large-scale labeled data.

[0139] Furthermore, to address the artifacts in magnetic resonance angiography signals and the inconsistency of vascular signals at different scales, this application sets up a "two-stage" algorithm that separates multi-scale filtering signal extraction and information integration, increasing the viewpoint image information that can be used for integration, and achieving stable and accurate segmentation performance on blood vessels at different scales.

[0140] Furthermore, this application summarizes and analyzes the essential characteristics widely present in cerebral vascular signals in multimodal magnetic resonance imaging (MRI) images—continuity, tubular structure, and signal distribution contrast. These essential characteristics are used as prior knowledge for segmentation algorithms. Different signal extraction and recognition methods from machine learning are constructed to generate segmentation viewpoint groups corresponding to different cerebral vascular features. Finally, multi-viewpoint fusion technology is used to integrate and unify the segmentation results of multiple feature viewpoint groups, enabling the unsupervised acquisition of accurate and stable cerebral vascular segmentation images for application in MRI angiography images with different modalities and scanning parameters.

[0141] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores unsupervised multi-view automatic segmentation data of cerebral blood vessels. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an unsupervised multi-view automatic segmentation method for cerebral blood vessels.

[0142] Those skilled in the art will understand that Figure 2 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0143] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0144] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0147] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for unsupervised multi-view automatic segmentation of cerebral blood vessels, characterized in that, include: Acquire multimodal 3D magnetic resonance images of the brain; Multimodal magnetic resonance three-dimensional images include T1-weighted imaging images and TOF-MRA magnetic resonance images; Based on the multimodal magnetic resonance three-dimensional image, a Frangi filter view group is constructed using the multi-scale Frangi filter method. A Sato filter view group is constructed based on the multimodal magnetic resonance three-dimensional image using a multi-scale Sato filter method. A multi-scale vascular diffusion enhancement method was used to construct a multi-scale vascular diffusion enhancement view group based on the multimodal magnetic resonance three-dimensional images; A Dirichlet process mixture model was used to construct a vascular classification result perspective group based on the multimodal magnetic resonance three-dimensional images; A multi-view fusion method based on map segmentation is used to fuse the Frangi filter view group, the Sato filter view group, the multi-scale vascular diffusion enhancement view group, and the vascular classification result view group. The cerebral vascular segmentation image is obtained by post-processing optimization. Based on the multimodal magnetic resonance 3D image, a Frangi filter view group is constructed using the multi-scale Frangi filter method, including: In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points; Perform eigenvalue decomposition on the Hessian matrix and determine the absolute values ​​of the eigenvalues ​​obtained from the decomposition. The eigenvalues ​​are sorted in descending order according to their absolute values ​​to obtain a sequence of absolute values ​​of the eigenvalues. The vascular enhancement response value is constructed based on the absolute value of all feature values ​​in the absolute value sequence of feature values, and the probability information of the existence of tubular continuum structure at different scales of each voxel point in the multimodal magnetic resonance three-dimensional image space is determined based on the vascular enhancement response value. Based on probabilistic information, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale. This three-dimensional blood vessel probability distribution matrix is ​​then used as the Frangi filter viewpoint at the corresponding scale. Specifically, the multimodal magnetic resonance three-dimensional image is convolved with a three-dimensional Gaussian convolution kernel of scale s, and the scale is adjusted accordingly. At a given scale Probability calculations for tubular continuum structures are performed under list conditions, at each scale. A three-dimensional blood vessel probability distribution matrix with the same dimensions as the multimodal magnetic resonance three-dimensional image is generated as the Frangi filter perspective at this scale; The Frangi filter perspective group is formed based on Frangi filter perspectives at all scales; wherein, a two-step strategy of cross-scale filtering and fusion separation is adopted to retain the probability information at each scale and to form Frangi filter perspective groups according to different filtering scales. We adopt a multi-view fusion unsupervised machine learning strategy, combining prior knowledge with mathematical modeling. Starting from the continuity of cerebral vascular signals at different scales, the differences in signal distribution, and tubular geometric features, we select and construct different feature extraction methods. We then use a multi-view fusion method to integrate cross-scale segmentation information from multiple feature perspectives and merge them into a unified segmentation result.

2. The unsupervised multi-view automatic segmentation method for cerebral blood vessels according to claim 1, characterized in that, Based on the multimodal magnetic resonance 3D image, a Sato filter view group is constructed using a multi-scale Sato filter method, including: In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points; The Hessian matrix is ​​decomposed into eigenvalues, and the eigenvalues ​​are sorted to form an eigenvalue sequence. Sato vascular enhancement response values ​​are constructed based on all feature values ​​in the feature value sequence, and the Sato probability value of each voxel point in the multimodal magnetic resonance three-dimensional image space having a tubular continuum structure at different scales is determined based on the Sato vascular enhancement response values. Based on the Sato probability value, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale, and the three-dimensional blood vessel probability distribution matrix is ​​used as the Sato filtering perspective at the corresponding scale. The Sato filter perspective group is formed based on the Sato filter perspectives at all scales.

3. The unsupervised multi-view automatic segmentation method for cerebral blood vessels according to claim 1, characterized in that, A multi-scale vascular diffusion enhancement method is employed based on the aforementioned multimodal magnetic resonance three-dimensional images to construct a multi-scale vascular diffusion enhancement view group, including: In the three-dimensional image space of multimodal magnetic resonance, voxel points are obtained, and a voxel-level Hessian matrix is ​​constructed based on the voxel points; Based on the Hessian matrix, different scales are selected, and smoothing and enhancement processing are performed based on the Frangi filtering results to establish vascular response functions at different scales. A three-dimensional diffusion tensor is constructed using the vascular response function and the Hessian matrix; The three-dimensional diffusion tensor is used to perform specific diffusion enhancement along the blood vessel axis. After a preset diffusion iteration time, a three-dimensional blood vessel probability distribution matrix with the same dimension as the multimodal magnetic resonance three-dimensional image is generated at each scale. The three-dimensional blood vessel probability distribution matrix is ​​used as the multi-scale blood vessel diffusion enhancement perspective at the corresponding scale. The multi-scale vascular diffusion enhancement perspective group is formed based on multi-scale vascular diffusion enhancement perspectives at all scales.

4. The unsupervised multi-view automatic segmentation method for cerebral blood vessels according to claim 3, characterized in that, Vascular response functions at different scales are expressed as follows: ； In the formula, exp() represents the exponential function. , , and All of these are hyperparameters. , and These are all eigenvalues ​​of the Hessian matrix, and e is the natural logarithm. Used to distinguish between disc-shaped and linear structures Used to identify spherical structures = Indicates structural strength.

5. The unsupervised multi-view automatic segmentation method for cerebral blood vessels according to claim 1, characterized in that, A Dirichlet process mixture model is used to construct a vascular classification result perspective group based on the multimodal 3D magnetic resonance images, including: The Dirichlet distribution is obtained by random sampling from the Dirichlet process, and the probability density function of the mixture distribution is generated. Categorical variables are sampled from a mixed distribution, and voxel signal values ​​are sampled from the corresponding multivariate Gaussian distribution based on the categorical variables; the categorical variables are used as cluster labels. The Gibbs sampling method is used to obtain blood vessel classification results based on the probability density function and the voxel signal value; The vascular classification result perspective group is constructed based on the vascular classification results.

6. The unsupervised multi-view automatic segmentation method for cerebral blood vessels according to claim 1, characterized in that, A multi-view fusion method based on atlas segmentation is used to fuse the Frangi filter view group, the Sato filter view group, the multi-scale vascular diffusion enhancement view group, and the vascular classification result view group to obtain a cerebral vascular segmentation image, including: For each perspective blood vessel segmentation image in the Frangi filter perspective group, the Sato filter perspective group, the multi-scale blood vessel diffusion enhancement perspective group, and the blood vessel classification result perspective group, a similarity matrix between voxels is constructed based on a Gaussian kernel function using a sparse representation method. A unified similarity matrix is ​​obtained based on the similarity matrix using a weighted average method. ; Based on the unified similarity matrix Determination matrix And based on a unified similarity matrix Sum-degree matrix Determine the graph Laplace matrix; the graph Laplace matrix is ; In the formula, For degree matrix elements, To unify the elements of the similarity matrix, and In this formula, the variable is represented. and Indicates the row and column indices of the matrix; Perform eigenvalue decomposition on the graph Laplacian matrix, and sort the resulting eigenvalues ​​in descending order, retaining the first few eigenvalues. Each eigenvalue and its relation to the previous The feature vectors corresponding to each feature value are used to adjust the sensitivity of blood vessel recognition, thereby obtaining the brain blood vessel segmentation image; The number of eigenvalues. For parameters.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the unsupervised multi-view automatic segmentation method for cerebral vessels according to any one of claims 1-6.

8. A 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 unsupervised multi-view automatic segmentation method for cerebral blood vessels as described in any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the unsupervised multi-view automatic segmentation method for cerebral blood vessels as described in any one of claims 1-6.

Images