A magnetic resonance image semantic segmentation method for cerebral arteriovenous malformations

Abstract

The application discloses a kind of magnetic resonance image semantic segmentation methods for cerebral arteriovenous malformation, first to the magnetic resonance angiography MRA image of cerebral arteriovenous malformation is preprocessed, construct low-order statistics model by the image after preprocessing, obtain the class likelihood probability of its pixel, based on maximum likelihood probability calculation initial segmentation result, again using Markov random field and reverse multi-scale filtering operator constructs high-order statistics model, obtain fine segmentation label set and fuzzy segmentation label set, finally fusion field adaptive method, contrast learning method and weak supervision learning method, training is carried out to double-branch network model, obtain the double-branch network model for cerebral arteriovenous malformation blood vessel segmentation, complete magnetic resonance image semantic segmentation.The method of the application fuses statistics segmentation method and deep learning method, realizes lesion blood vessel high-performance segmentation, solves the problems of large consumption of computing resources, long calculation time, less training data and poor segmentation effect in existing segmentation technology.

Description

Technical Field

[0001] This invention belongs to the field of image semantic segmentation technology, specifically relating to a method for semantic segmentation of magnetic resonance images for cerebral arteriovenous malformations. Background Technology

[0002] Arteriovenous malformations (AVMs) are the most common type of cerebrovascular malformation, with a prevalence of approximately 16–23 per 100,000 people. The majority of cases occur in individuals aged 10–40, indicating a trend towards younger onset. Patients may experience aneurysm risk, epilepsy, and neurological damage. In severe cases, AVMs can lead to rupture of blood vessels, causing intracranial hemorrhage and resulting in high rates of disability and mortality. The pathological characteristic of AVMs is the absence of capillary connections between arteries and veins, forming a cluster of malformed blood vessels with uneven diameters and thicknesses. Blood flowing through these malformed vessels bypasses surrounding tissues, directly and rapidly entering the veins from the arteries. This places immense pressure on the inner walls of the arteries and veins, causing them to thin and become brittle, ultimately potentially leading to cerebral rupture. Due to the complexity of AVM structures, it is difficult to distinguish the locations of feeding arteries and draining veins, hindering subsequent diagnosis, surgical planning, and intraoperative guidance. Therefore, generating a three-dimensional visualization model of the blood vessels is necessary. This would free surgeons from the experience-based mental construction of AVM geometric models, allowing them to focus more on observing the actual three-dimensional structure of the vessels, analyzing the severity of the condition, developing detailed treatment plans, and predicting postoperative outcomes.

[0003] Blood vessel segmentation is a crucial step in obtaining accurate 3D reconstruction models of blood vessels. However, due to the non-uniformity of pixel intensity in blood vessel images and the complexity of the blood vessel structure itself, it is difficult to achieve satisfactory segmentation results in brain blood vessel images. Oversegmentation and undersegmentation are also common problems, making this task still extremely challenging. Currently, numerous scholars both domestically and internationally have conducted research on cerebral vascular segmentation. These studies are mainly divided into two categories: model-based and data-driven segmentation methods. Model-based segmentation methods can be summarized as statistical methods, deformation model methods, and tracking methods. Model-based methods can directly and automatically segment images using statistical features or probability distributions without additional model training. However, they make many assumptions about the images, resulting in low accuracy for segmentation results with rich details, and their computational complexity is high, leading to long execution times. Data-driven segmentation methods are mainly based on machine learning. They train neural network models using a large amount of image data and labels. The trained network can achieve fast semantic segmentation of images. However, this method requires the collection and labeling of a large amount of medical image data. Due to the rarity of AVM cases and the complexity of AVM MRI images, manually labeling a large number of images to train the model is difficult to achieve. Therefore, data-driven segmentation methods urgently need to solve the problems of limited sample size and difficulty in obtaining labels. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a semantic segmentation method for magnetic resonance images of cerebral arteriovenous malformations (AVMs). By integrating statistical methods and deep learning methods, a semantic segmentation algorithm for cerebral blood vessels is obtained, which solves the problems of high computational resource consumption, long computation time, limited training data, and poor segmentation results in existing segmentation techniques for AVMs.

[0005] The technical solution adopted in this invention is: a semantic segmentation method for magnetic resonance images of cerebral arteriovenous malformations, the specific steps of which are as follows:

[0006] S1. Preprocessing of magnetic resonance angiography (MRA) images of cerebral arteriovenous malformations;

[0007] S2. Based on the preprocessed magnetic resonance angiography (MRA) images from step S1, a low-order statistical model is constructed to obtain the class likelihood probability of pixels, and the initial segmentation result is calculated based on the maximum likelihood probability.

[0008] S3. Based on the initial segmentation results of step S2, a high-order statistical model is constructed using Markov random fields and inverse multi-scale filtering operators to obtain the precise segmentation label set and the fuzzy segmentation label set.

[0009] S4. Based on the precise segmentation label set and fuzzy segmentation label set obtained in step S3, the dual-branch network model is trained by integrating the domain adaptive method, the contrastive learning method and the weakly supervised learning method to obtain a dual-branch network model for segmentation of arteriovenous malformations in the brain, thus completing the semantic segmentation of magnetic resonance images.

[0010] Furthermore, step S1 is specifically as follows:

[0011] First, gamma correction is applied to the initial magnetic resonance angiography (MRA) images to adjust the image brightness and contrast, highlighting the target vascular region. Then, median filtering is used to reduce noise while preserving image details. Next, the FSL skull removal algorithm is used to extract the brain image. Finally, Min-Max normalization is applied to the images to scale the image intensity to the range of 0 to 255, eliminating intensity differences between images.

[0012] Furthermore, step S2 is specifically as follows:

[0013] Based on the preprocessed magnetic resonance angiography (MRA) images obtained in step S1, a low-order statistical model is constructed to perform preliminary statistical segmentation and obtain initial segmentation results.

[0014] The low-order statistical model is a finite mixture model based on the image pixel intensity histogram, combining Rayleigh and Gaussian distributions.

[0015] First, using prior knowledge of magnetic resonance angiography (MRA) images, the pixel intensity histogram of the MRA images is divided into three parts according to pixel intensity: low pixel intensity regions correspond to cerebrospinal fluid, bones, and background air; medium intensity regions correspond to brain tissue, including gray matter and white matter, as well as part of the eye; and high intensity regions correspond to blood vessels and subcutaneous tissue.

[0016] Then, based on the finite mixture model, the EM algorithm is used to perform high-precision fitting of the pixel intensity distribution.

[0017] The high-intensity region distribution function of the finite mixture model is used to calculate the vessel category likelihood function, while the other distribution functions of the finite mixture model are combined to calculate the background category likelihood function.

[0018] The finite mixture model is given by a Rayleigh distribution f. R (y s ) and three Gaussian distributions f Gi (y s The function is constructed from (i = 1, 2, 3) and its expression is as follows:

[0019]

[0020] Among them, w R and w Gi (i = 1, 2, 3) represent the weights of each distribution function on the fitting effect, and satisfy the weight normalization constraint. y s The intensity value of pixel s is expressed by the following functions based on its Rayleigh and Gaussian distributions:

[0021]

[0022] Among them, the weight parameters and σ i μ i The β distribution parameters are initialized using histogram distribution features and K-Means clustering algorithm, and the pixel distribution is fitted using the expectation-maximization (EM) algorithm. After a set number of iterations of the EM algorithm, the parameters are updated based on the previous iteration, and finally an accurate approximation of the pixel distribution is obtained.

[0023] Then, based on the distribution function of the fitted finite mixture model, the blood vessel category X of each pixel point is calculated. V and background category X B The likelihood probability p(y) s |x s The expression is as follows:

[0024]

[0025] Where, x s ∈{X V ,X B} represents the possible classification categories of pixel s. The likelihood probability of all pixels in the image constitutes the likelihood probability map, and the value of each pixel in the likelihood probability map is the likelihood probability of that pixel.

[0026] Finally, the category of pixel s is obtained by maximum likelihood probability estimation and used as the classification result of the low-order statistical model, i.e., the initial segmentation result. The expression is as follows:

[0027]

[0028] Furthermore, step S3 is specifically as follows:

[0029] S31. Construct a higher-order statistical model using Markov random field and inverse multi-scale filtering operator. Input the initial segmentation results and likelihood probability map obtained in step S2 into the higher-order statistical model to obtain the accurate segmentation label set.

[0030] First, based on the inverse multi-scale filtering operator, the intensity response of the blood vessel edge is calculated. Then, the image is flipped pixel by pixel intensity, setting the background as a high-intensity region and the blood vessel as a low-intensity region.

[0031] I'(s) = 255 - I(s)

[0032] Where I(s) represents the original image at s=(x1,x2) T The pixel intensity value at the location, T represents the transpose operation, and I'(s) represents the corresponding pixel intensity value after flipping. The image after flipping is subjected to multi-scale filtering to obtain the enhancement of its target edges. The expression of the Hessian matrix H(s,σ) of its filtering response is as follows:

[0033]

[0034]

[0035] Where * denotes convolution operation, and σ denotes the scale of the Gaussian operator; G(s,σ)=(2πσ 2 ) -1 exp(-s T s / 2σ 2 ) represents a Gaussian kernel with a scale of σ.

[0036] For each pixel, the Hessian matrix H(s,σ) can be decomposed into two eigenvalues: |λ 1,σ |≤|λ 2,σ |, Eigenvalue relationships are used to distinguish different structural features; then the filter operator response value Vσ The expression is as follows:

[0037]

[0038] Where, λ ρ,σ The regularization term used to reduce noise sensitivity is expressed as follows:

[0039]

[0040] Here, κ represents a pre-set parameter used to control the regularization strength, and its value is between 0 and 1.

[0041] When applying multiple filtering operators of different scales, the maximum response is selected as the final filtered response value of pixel s. It represents the set of positive real numbers.

[0042] Then, the inverse multi-scale filter response value is fused with the Markov random field to obtain the pixel-level potential function, as shown in the following expression:

[0043] E(x s ,x r ) = w a E1(x s ,x r )+l*w b V(r)

[0044] Where, x s ∈{X V ,X B} represents the possible classification category of the center pixel s; x r This represents the classification result at pixel r, within the 26-neighborhood of pixel s, obtained from the initial segmentation based on a low-order statistical model, where r ∈ η(s); E1(x s ,x r ) represents the category x of the center pixel s. s Its neighboring pixel classification category x r The relationship between the two labels is such that if the labels are the same, the value is -1, otherwise it is 1; l∈{-1,1} represents the ability of the multi-scale filtering response V(r) to suppress or enhance subsequent prior probabilities. s =X V Then l=1 suppresses subsequent prior probabilities, and conversely l=-1 enhances subsequent prior probabilities; w a and w b These represent the weights, used to balance the effects of the Markov random field operator and the multi-scale filtering operator.

[0045] By summing the potential energy functions related to all neighborhoods, we obtain the energy function U(x) of the center pixel s. sThe expression is as follows:

[0046]

[0047] Based on the Hammersley-Clifford theorem, the prior probability p(x) of pixel s with respect to each category is obtained. s The expression is as follows:

[0048]

[0049] Where Z represents the normalization constant, which is obtained by summing all x. s Category of molecules We obtain the prior probability p(x) s Normalize to the interval (0, 1).

[0050] Based on the likelihood and prior probabilities obtained from the aforementioned low-order and high-order statistical models, the posterior probability of the center pixel s is calculated using Bayes' theorem. Then, using the maximum a posteriori probability algorithm, the class with the highest posterior probability is selected as the final segmentation result for pixel s. The expression is as follows:

[0051]

[0052] in, This indicates a direct proportional relationship; by comparing pixel intensity y s Under the condition of vessel type X V and background category X B The final segmentation result is determined by the posterior probability, with the category label corresponding to the highest posterior probability being selected. The accurate segmentation label set B is obtained by performing maximum a posteriori probability estimation on all pixels of the image.

[0053] S32. Based on the precise segmentation label set obtained in step S31 and the likelihood probability map obtained in step S2, a fuzzy segmentation label set is obtained through weighted maximum likelihood probability estimation and Boolean difference operation.

[0054] The expanded segmentation result, i.e., the expanded foreground label set A, is obtained by performing weighted maximum likelihood probability estimation on all pixels in the image, as shown in the following expression:

[0055] p(y s |X V )>λ*p(y s |X B )

[0056] Where λ represents the weighting weight, and λ < 1 expands the blood vessel category region and shrinks the background category region.

[0057] Then, a Boolean difference operation is performed between the expanded foreground label set A and the precise segmentation label set B obtained in step S31. The Boolean difference operation result (AB) is used as the fuzzy segmentation label set that needs to be re-labeled using deep learning methods in the future.

[0058] Furthermore, step S4 is specifically as follows:

[0059] S41. Based on contrastive learning and domain adaptation algorithms, the dual-branch model is trained using a precise segmentation label set and a public dataset;

[0060] Normal brain magnetic resonance angiography (MRA) images after Min-Max normalization are introduced as the source domain in domain adaptive learning. This source domain dataset is derived from a public dataset, while the AVM dataset is used as the target domain. The domain adaptive framework is constructed by combining the source domain and the target domain.

[0061] Then, by combining contrastive learning and domain adaptation algorithms to extract domain-invariant features, two bi-branch models are trained using source and target domain datasets respectively.

[0062] In this dual-branch model, a parallel network layer is added to the bottleneck layer of the U-Net skeleton network. The bottleneck layer consists of two parallel network layers: an independent layer and a shared layer. The shared layer is used to extract domain-invariant features between the source and target domains. The two dual-branch models share a shared layer for domain-invariant feature extraction. The independent layer extracts domain-specific features for their respective datasets. Each model has its own independent layer.

[0063] In the domain adaptive learning process, a contrastive learning method is used, where the outputs of independent layers and shared layers are set to be negative pairs, and the output features of shared layers in different domains are set to be positive pairs. Through subsequent model training, the similarity between positive pairs of features is maximized, and the similarity between negative pairs of features is minimized.

[0064] The comprehensive loss function is obtained by weighting and combining the comparative loss function, the loss function of the source domain data, and the loss function of the target domain data. The parameters of the network model are updated with the goal of minimizing the comprehensive loss function. After a pre-set number of training iterations, the optimal dual-branch model during the training process is output.

[0065] S42. Based on the model trained in step S41, use an iterative pseudo-label update method to select high-confidence pseudo-labels to fine-tune the next generation of the model.

[0066] The magnetic resonance angiography (MRA) images from the target domain AVM dataset are input into the model trained in step S41 for prediction and segmentation. First, the class probability Prob = Net(G) of each pixel in the image is calculated using the dual-branch network model of the target domain.

[0067] Here, Net represents a two-branch network model trained on the target domain dataset G.

[0068] After obtaining the current Prob, initialize the uncertain mask M in the current iteration based on the uncertain mask M and pseudo-label Ps from the previous iteration. Cur and pseudo-tags Ps Cur .

[0069] Then, a pre-set segmentation threshold is used to filter high-confidence pixel labels, selecting categories that meet the condition (Prob > threshold) as the current pseudo-labels, and setting their uncertainty mask M. Cur A value of 0 for the corresponding pixel position indicates that the label at that pixel position is determined; otherwise, an uncertain mask M is set. Cur A pixel value of 1 indicates that the label at that pixel position is uncertain.

[0070] Finally, comparing the previous generation and the current pseudo-labels, if the pseudo-label category changes, the mask M will be uncertain. Cur Setting the corresponding pixel value to 1 indicates that the label at that pixel position is uncertain, resulting in the final pseudo-label set Ps. Cur and uncertain mask M Cur The pseudo-label set and uncertain mask of the current iteration serve as the input conditions Ps and M for the next iteration.

[0071] S43. Based on the accurate segmentation label set and the updated pseudo-label set, a new round of model fine-tuning is carried out. After model training and pseudo-label updating for a preset number of iterations, a deep learning model for segmentation of cerebral arteriovenous malformation vessels is obtained.

[0072] The model fine-tuning employs a weakly supervised learning method, which integrates the precise segmentation of the label set and the pseudo-label set for model training, and uses the precise segmentation of the label set and the pseudo-label set as the gold standard in the calculation of the loss function.

[0073] S44. Input the magnetic resonance angiography (MRA) images from the AVM into the trained dual-branch model targeting the target domain to complete the semantic segmentation of the images.

[0074] The beneficial effects of this invention are as follows: The method of this invention first preprocesses the magnetic resonance angiography (MRA) images of cerebral arteriovenous malformations (AVMs). A low-order statistical model is constructed using the preprocessed images to obtain the class likelihood probability of each pixel. The initial segmentation result is calculated based on the maximum likelihood probability. Then, a high-order statistical model is constructed using Markov random fields and inverse multi-scale filtering operators to obtain a fine-grained segmentation label set and a fuzzy segmentation label set. Finally, a domain-adaptive method, a contrastive learning method, and a weakly supervised learning method are integrated to train a dual-branch network model, resulting in a dual-branch network model for segmenting blood vessels in cerebral AVMs, thus completing the semantic segmentation of the magnetic resonance image. This invention integrates statistical segmentation methods and deep learning methods to achieve high-performance segmentation of lesion blood vessels, solving the problems of high computational resource consumption, long computation time, limited training data, and poor segmentation results in existing segmentation techniques. Compared with existing segmentation techniques, the method of this invention uses statistical segmentation methods to initially obtain precise segmentation label sets and fuzzy segmentation label sets, eliminating the need for professional doctors to manually label a large number of blood vessels, reducing labeling costs and requirements, and greatly alleviating the labeling burden. At the same time, it combines contrastive learning methods and domain adaptation methods to solve the problem of the limited number of cerebral arteriovenous malformation cases, and uses existing normal blood vessel datasets to fully extract domain-invariant features, improving the model's generalization ability and segmentation effect. In addition, the weakly supervised learning method based on pseudo-label updates introduces fuzzy region information into the model training process and ensures the historical continuity and reliability of its pseudo-labels, improving the model's understanding of features in difficult-to-segment regions. Attached Figure Description

[0075] Figure 1 This is a flowchart of a magnetic resonance imaging semantic segmentation method for arteriovenous malformations of the brain according to the present invention.

[0076] Figure 2 These are the original and preprocessed magnetic resonance angiography (MRA) images in the embodiments of the present invention.

[0077] Figure 3 This is a flowchart of a low-order statistical model in an embodiment of the present invention.

[0078] Figure 4 This is a diagram showing the effect of a low-order statistical model in an embodiment of the present invention.

[0079] Figure 5 This is a flowchart of a high-order statistical model in an embodiment of the present invention.

[0080] Figure 6 This is a diagram illustrating the effect of a higher-order statistical model in an embodiment of the present invention.

[0081] Figure 7 This is a diagram of the deep learning model training framework in an embodiment of the present invention.

[0082] Figure 8 This is a pseudocode diagram of the pseudo-label update algorithm in an embodiment of the present invention.

[0083] Figure 9 This is a diagram showing the deep learning semantic segmentation result in an embodiment of the present invention. Detailed Implementation

[0084] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0085] like Figure 1 The flowchart shown below illustrates a method for semantic segmentation of magnetic resonance images for arteriovenous malformations of the brain. The specific steps are as follows:

[0086] S1. Preprocessing of magnetic resonance angiography (MRA) images of cerebral arteriovenous malformations;

[0087] S2. Based on the preprocessed magnetic resonance angiography (MRA) images from step S1, a low-order statistical model is constructed to obtain the class likelihood probability of pixels, and the initial segmentation result is calculated based on the maximum likelihood probability.

[0088] S3. Based on the initial segmentation results of step S2, a high-order statistical model is constructed using Markov random fields and inverse multi-scale filtering operators to obtain the precise segmentation label set and the fuzzy segmentation label set.

[0089] S4. Based on the precise segmentation label set and fuzzy segmentation label set obtained in step S3, the dual-branch network model is trained by integrating the domain adaptive method, the contrastive learning method and the weakly supervised learning method to obtain a dual-branch network model for segmentation of arteriovenous malformations in the brain, thus completing the semantic segmentation of magnetic resonance images.

[0090] In this embodiment, step S1 is specifically as follows:

[0091] Preprocessing magnetic resonance angiography (MRA) images to improve image quality is helpful for subsequent statistical segmentation.

[0092] First, gamma correction is applied to the initial magnetic resonance angiography (MRA) images to adjust image brightness and contrast, highlighting the target vascular region. Then, median filtering is used to reduce noise while preserving image details. Next, the FSL skull removal algorithm is used to extract the brain image. Finally, Min-Max normalization is applied to the images, scaling the image intensity to the range of 0 to 255 to eliminate intensity differences between images and improve the segmentation effect of subsequent statistical multi-level segmentation models.

[0093] This embodiment uses partial magnetic resonance angiography (MRA) images and preprocessed magnetic resonance angiography (MRA) images, as shown below. Figure 2As shown, the highlighted areas of blood vessels in the image are more prominent, and the contrast with the background is enhanced, suppressing non-brain areas and reducing image interference noise.

[0094] In this embodiment, step S2 is specifically as follows:

[0095] Based on the preprocessed magnetic resonance angiography (MRA) images obtained in step S1, a low-order statistical model is constructed to perform preliminary statistical segmentation and obtain initial segmentation results.

[0096] The low-order statistical model is a finite mixture model based on the image pixel intensity histogram, combining Rayleigh and Gaussian distributions.

[0097] In this embodiment, the implementation process of the low-order statistical model is as follows: Figure 3 As shown, the overall distribution of the pixel histogram of the magnetic resonance image is a high-low-high-low distribution, and the brain structure in the magnetic resonance image is divided into three parts based on the pixel intensity.

[0098] First, using prior knowledge of magnetic resonance angiography (MRA) images, the pixel intensity histogram of the MRA images is divided into three parts according to pixel intensity: low pixel intensity regions correspond to cerebrospinal fluid, bones, and background air; medium intensity regions correspond to brain tissue, including gray matter and white matter, as well as part of the eye; and high intensity regions correspond to blood vessels and subcutaneous tissue (i.e., the target blood vessel region).

[0099] To extract the likelihood probability of each pixel, the EM algorithm is used to fit the pixel intensity distribution with high accuracy based on the finite mixture model.

[0100] The high-intensity region distribution function of the finite mixture model is used to calculate the vessel category likelihood function, while the other distribution functions of the finite mixture model are combined to calculate the background category likelihood function.

[0101] The finite mixture model is given by a Rayleigh distribution f. R (y s ) and three Gaussian distributions f Gi (y s The function is constructed from (i = 1, 2, 3) and its expression is as follows:

[0102]

[0103] Among them, w R and w Gi (i = 1, 2, 3) represent the weights of each distribution function on the fitting effect, and satisfy the weight normalization constraint. y sThe intensity value of pixel s is expressed by the following functions based on its Rayleigh and Gaussian distributions:

[0104]

[0105] Among them, the weight parameters and σ i μ i The β distribution parameters are initialized using histogram distribution features and K-Means clustering algorithm, and the pixel distribution is fitted using the expectation-maximization (EM) algorithm. After a set number of iterations, an accurate approximation of the pixel intensity distribution is obtained, improving the fitting effect. Finally, the foreground and background are classified based on the obtained likelihood probability.

[0106] Then, based on the distribution function of the fitted finite mixture model, the blood vessel category X of each pixel point is calculated. V and background category X B The likelihood probability p(y) s |x s The expression is as follows:

[0107]

[0108] Where, x s ∈{X V ,X B} represents the possible classification categories of pixel s. The likelihood probability of all pixels in the image constitutes the likelihood probability map, and the value of each pixel in the likelihood probability map is the likelihood probability of that pixel.

[0109] Finally, the category of pixel s is obtained by maximum likelihood probability estimation and used as the segmentation result of the low-order statistical model, i.e., the initial segmentation result. The expression is as follows:

[0110]

[0111] like Figure 4 As shown in the figure, this embodiment displays the effect of the low-order statistical model, wherein... Figure 4 (a) is the original magnetic resonance angiography (MRA) image. Figure 4 (b) is the pixel distribution histogram corresponding to the image. Figure 4 (c) is the histogram distribution fitted using a finite mixture model. Figure 4 (d) shows the initial segmentation results of blood vessels using a low-order statistical model.

[0112] Optimizing the initial segmentation results obtained in step S2 yields more accurate blood vessel labels, which can reduce the noise impact of training samples on the deep learning model.

[0113] For segmentation of arteriovenous malformations in the brain, the segmentation results of edge pixels are not ideal due to the blurred edges and complex shapes of the blood vessels. Subsequent deep learning is used to further determine the edge pixels, thereby improving the overall segmentation effect.

[0114] In this embodiment, step S3 is specifically as follows:

[0115] A higher-order statistical model, incorporating Markov random fields and inverse multi-scale filtering operators, refines the initial segmentation results obtained from the lower-order statistical model, eliminating classification labels with low confidence. The resulting precise segmentation label set and fuzzy segmentation label set are used for subsequent deep learning model training. In this embodiment, the intensity response of blood vessel edges is calculated based on the inverse multi-scale filtering operator, and the continuity of the segmentation results is optimized by combining Markov random fields to reduce the prior probability of hard-to-segment regions. Based on the prior probability and likelihood probability, the initial segmentation results of the lower-order statistical model are optimized using Bayes' theorem and maximum a posteriori probability estimation to obtain a precise segmentation label set. Weighted maximum likelihood probability estimation is used to obtain foreground labels for a wider region, and Boolean difference operations are used to obtain a fuzzy segmentation label set.

[0116] In this embodiment, the implementation flowchart of the higher-order statistical model is as follows: Figure 5 As shown.

[0117] S31. Construct a higher-order statistical model using Markov random field and inverse multi-scale filtering operator. Input the initial segmentation results and likelihood probability map obtained in step S2 into the higher-order statistical model to obtain the accurate segmentation label set.

[0118] First, based on the inverse multi-scale filtering operator, the intensity response of blood vessel edges is calculated (especially at the edge of lesions, where the filtering response ranges from 0 to 1. Since the operator is applied to the inverse image, it suppresses bright blood vessels in the original image, thus improving its detection performance in blurred and uncertain edge background regions). Then, the image is flipped pixel by pixel intensity, setting the background as a high-intensity region and the blood vessels as a low-intensity region.

[0119] I'(s) = 255 - I(s)

[0120] Where I(s) represents the original image at s=(x1,x2) T The pixel intensity value at the location, T represents the transpose operation, and I'(s) represents the corresponding pixel intensity value after flipping. The image after flipping is subjected to multi-scale filtering to obtain the enhancement of its target edges. The expression of the Hessian matrix H(s,σ) of its filtering response is as follows:

[0121]

[0122]

[0123] Where * denotes convolution operation, and σ denotes the scale of the Gaussian operator; G(s,σ)=(2πσ 2 ) -1 exp(-s T s / 2σ 2 ) represents a Gaussian kernel with a scale of σ.

[0124] By measuring the changes in pixel intensity in an image region using G(s,σ), local structures are identified. Since the image is flipped, the target region becomes the background. By exploring the pixel changes between the background and the foreground, the edges of the background target region are enhanced. This allows for the identification of low-confidence vascular regions and background regions, facilitating subsequent classification optimization.

[0125] For each pixel, the Hessian matrix H(s,σ) can be decomposed into two eigenvalues: |λ 1,σ |≤|λ 2,σ Different eigenvalue relationships are used to distinguish different structural features; for example, based on |λ... 2,σ |>>|λ 1,σ |and λ 2,σ If the value is less than 0, the tubular structure can be identified; therefore, the filter operator response value V... σ The expression is as follows:

[0126]

[0127] Where, λ ρ,σ The regularization term used to reduce noise sensitivity is expressed as follows:

[0128]

[0129] Here, κ represents a pre-set parameter used to control the regularization strength, and its value is between 0 and 1.

[0130] When applying multiple filtering operators of different scales, the maximum response is selected as the final filtered response value of pixel s. It represents the set of positive real numbers.

[0131] Then, the inverse multi-scale filter response value is fused with a Markov random field (the Markov random field is used to consider the spatial continuity of the label distribution of the initial segmentation result; if the labels of the center pixel and its neighboring pixels are inconsistent, the label classification probability needs to be reduced) to obtain the pixel-level potential function, as shown in the following expression:

[0132] E(x s ,x r ) = w a E1(x s ,xr )+l*w b V(r)

[0133] Where, x s ∈{X V ,X B} represents the possible classification category of the center pixel s, x r This represents the classification result at pixel r, within the 26-neighborhood of pixel s, obtained from the initial segmentation based on a low-order statistical model, where r ∈ η(s); E1(x s ,x r ) represents the category x of the center pixel s. s Its neighboring pixel classification category x r The relationship between the two labels is such that if the labels are the same, the value is -1, otherwise it is 1; l∈{-1,1} represents the ability of the multi-scale filtering response V(r) to suppress or enhance subsequent prior probabilities. s =X V Then l=1 suppresses subsequent prior probabilities, and conversely l=-1 enhances subsequent prior probabilities; w a and w b These represent the weights, used to balance the effects of the Markov random field operator and the multi-scale filtering operator.

[0134] By summing the potential energy functions related to all neighborhoods, we obtain the energy function U(x) of the center pixel s. s The expression is as follows:

[0135]

[0136] Based on the Hammersley-Clifford theorem, the prior probability p(x) of pixel s with respect to each category is obtained. s The expression is as follows:

[0137]

[0138] Where Z represents the normalization constant, which is obtained by summing all x. s Category of molecules We obtain the prior probability p(x) s Normalize to the interval (0, 1).

[0139] Based on the likelihood and prior probabilities obtained from the aforementioned low-order and high-order statistical models, the posterior probability of pixel s is estimated using Bayes' theorem. Then, the maximum a posteriori probability algorithm is used to select the class with the highest posterior probability as the final segmentation result for pixel s. (This optimization result fully considers the multi-scale filtering response and Markov random field, which can guarantee high-precision segmentation, but its recall rate is reduced accordingly.) The expression is as follows:

[0140]

[0141] in, This indicates a direct proportional relationship; by comparing pixel intensity y s Under the condition of vessel type X V and background category X B The final segmentation result is determined by the posterior probability, with the category label corresponding to the highest posterior probability being selected. The accurate segmentation label set B is obtained by performing maximum a posteriori probability estimation on all pixels of the image.

[0142] S32. Based on the precise segmentation label set obtained in step S31 and the likelihood probability map obtained in step S2, a fuzzy segmentation label set is obtained by weighted maximum likelihood probability and Boolean difference operation.

[0143] To fully utilize the information in hard-to-segment regions, reduce background label noise, and minimize false negatives, enabling subsequent deep learning models to learn the semantic information of these regions more effectively, a weighted maximum likelihood probability estimation is performed on all pixels in the image to obtain the expanded segmentation result, i.e., the expanded foreground label set A, expressed as follows:

[0144] p(y s |X V )>λ*p(y s |X B )

[0145] Where λ represents the weighting weight, and λ < 1 expands the blood vessel category region and shrinks the background category region.

[0146] Then, a Boolean difference operation is performed between the expanded foreground label set A and the precise segmentation label set B obtained in step S31. The Boolean difference operation result (AB) is used as the fuzzy segmentation label set that needs to be re-labeled using deep learning methods in the future.

[0147] By fusing high-order and low-order statistical models, statistical segmentation results can be obtained, and the corresponding segmentation effects are shown in Table 1. The accuracy of the blood vessel labels after refinement by the high-order model is significantly improved, but the recall rate is reduced. By expanding the fuzzy region through weighted maximum likelihood probability estimation, the fuzzy region refers to the pixel position corresponding to the fuzzy segmentation label set, and the definite region refers to the pixel position corresponding to the precise segmentation label set. The effect of fusing the labels of the fuzzy region and the definite region is tested, and the recall rate is significantly improved. Therefore, it is proved that the fuzzy region contains a large number of blood vessel labels. The pseudo-labels of the fuzzy region can be updated generation by generation through the subsequent pseudo-label update algorithm. Weakly supervised learning is performed by fusing the precise segmentation label set and the pseudo-label set to further extract and confirm the semantic information of blood vessels in the fuzzy region.

[0148] Table 1

[0149]

[0150] like Figure 6 As shown in the figure, this embodiment presents the effect of the higher-order statistical model. Figure 6 (a) is the original magnetic resonance angiography (MRA) image. Figure 6 (b) is the inverse multiscale filtering response. Figure 6 (c) is the output image of the high-order statistical model, where the black area is the background label, the white area is the precise segmentation label set, and the gray area is the fuzzy segmentation label set. The overall accuracy of the optimized blood vessel label set is significantly improved, especially the fuzzy background between blood vessels in the lesion area is optimized.

[0151] The training framework for the deep learning model in this embodiment is as follows: Figure 7 As shown.

[0152] In this embodiment, step S4 is specifically as follows:

[0153] S41. Based on contrastive learning and domain adaptation algorithms, the dual-branch model is trained using a precise segmentation label set and a public dataset;

[0154] To address the underfitting or overfitting issues caused by insufficient training data, a publicly available dataset, namely magnetic resonance angiography (MRA) images of normal brain tissue, was introduced after Min-Max normalization. Using the normal brain vascular dataset as the source domain and the AVM vascular dataset as the target domain, a dual-branch network model was trained by combining contrastive learning and domain adaptation algorithms. The training framework is as follows: Figure 7 As shown.

[0155] Due to the differences between dataset domains, domain adaptation algorithms are needed to extract domain-invariant features. To improve the model's ability to extract domain-invariant features, two bi-branch models are constructed based on contrastive learning and domain adaptation algorithms to train the source and target domain datasets respectively. During training, the output features of the bi-branch network at the bottleneck layer are optimized to maximize the similarity of domain-invariant features between the two data domains and minimize the similarity of domain-specific features.

[0156] The dual-branch model adds a parallel network layer to the bottleneck layer of the U-Net skeleton network, resulting in two independent outputs at the bottleneck layer. The bottleneck layer consists of two parallel network layers: an independent layer and a shared layer. The shared layer extracts domain-invariant features between the source and target domains. Both dual-branch models share this shared layer for domain-invariant feature extraction, while the independent layers extract domain-specific features specific to their respective datasets. Each model has its own independent layer. When training the network using datasets from the source and target domains, two separate U-Net skeleton networks are used for semantic segmentation prediction. The two networks are then fused together through the added parallel bottleneck layer, yielding the outputs of the independent and shared layers, respectively. Through contrastive learning, the similarity between the shared layer outputs of the source and target domains is increased, while the similarity between the independent and shared layer outputs is reduced, thus enhancing the shared layer's ability to extract domain-invariant features and the independent layer's ability to extract domain-specific features.

[0157] In the domain adaptive learning process, a contrastive learning method is used, where the outputs of independent layers and shared layers are set to be negative pairs. Figure 7 The functions (S1, L1), (S1, L2), (S2, L1), and (S2, L2) are defined, and the output features of the shared layer in different domains are set to be positively aligned. Figure 7 The learning process is compared between (S1, S2) and (S2, S1) as follows: Figure 7 As shown, the comprehensive loss function is obtained by weighting and combining the comparison loss function, the loss function of the source domain data, and the loss function of the target domain data. The parameter update of the network model is achieved by minimizing the comprehensive loss function. After a pre-set number of training times, the optimal dual-branch model during the training process is output.

[0158] Wherein, the contrastive loss function between the source domain and the target domain Contrast loss function between target domain and source domain Comparative loss function L ct as follows:

[0159]

[0160]

[0161]

[0162] Where S and G represent the source domain and target domain datasets, respectively. and Let b represent independent samples in the source domain S and the target domain dataset G, respectively. S-sh and b G-sh N represents the feature vectors output by the source and target domains in the shared layer, respectively. i and N j Let b represent the sets of output feature vectors of the source and target domains in their respective independent layers of the dual-branch network. i and b j τ represents the feature vectors output by the source domain and the target domain in their respective independent layers of the dual-branch network, respectively, and τ represents the temperature coefficient, which is set to 0.1 in this embodiment.

[0163] This embodiment uses the Dice loss function L. Dice As a supervised loss function for the source domain, this function evaluates the consistency between the gold standard label set and the predicted source domain vessel labels. By training and reducing this loss function, the predicted vessel labels will be more accurate.

[0164]

[0165] Among them, X S and Y S X represents the gold standard label set of the source domain dataset and the predicted label set of the source domain dual-branch network, respectively. A value of 0 indicates a background label, and a value of 1 indicates a blood vessel label; S | represents the number of vascular tags in the source domain gold standard tagging cluster, |Y S | represents the number of labels predicted as blood vessels in the source domain image, |X S ∩Y S |Is X S and Y S The intersection elements between them also represent the number of accurately predicted vascular tags.

[0166] For learning the target domain, the loss function is set as a partial Dice loss function L. Dice-partial The loss function uses an uncertain mask of fuzzy regions to filter high-confidence pixel labels for model training and evaluates the consistency between the accurate segmentation label set, the pseudo label set and the predicted blood vessel label set.

[0167]

[0168] Among them, X G There are two meanings, in step S41 X G This represents the precise segmentation label set of the target domain AVM dataset, while X is used in the model fine-tuning training in step S43.G Y represents the fusion of the exact segmentation label set and the pseudo-label set of the target domain AVM dataset, where the labels of fuzzy regions are replaced by the corresponding labels in the pseudo-label set; G This represents the predicted label set of the dual-branch network in the target domain; M represents the uncertain mask for the blurred region, where a value of 0 indicates that the label at that pixel position is determined, and a value of 1 indicates that the label at that pixel position is still uncertain and cannot be used for model training; |X G *(1-M)| represents X G The number of blood vessel labels is determined by M, where the label is a blood vessel and the pixel value in M ​​is 0. |Y G *(1-M)| represents Y G The number of blood vessel labels in M ​​where the label is a blood vessel and the pixel value in M ​​is 0 is determined, |[X G *(1-M)]∩[Y G *(1-M)]| indicates that it is simultaneously affected by X G and Y G The number of true examples classified as blood vessels with pixel values ​​of 0 in M. Therefore, the overall model training loss function can be written as:

[0169]

[0170] in, and w1 and w2 represent the Dice loss function of the source domain and the partial Dice loss function of the target domain, respectively. w1 and w2 represent the influence of weight factors on the loss functions, which are set to 1 and 0.1, respectively, in this embodiment.

[0171] The training parameters in this embodiment are shown in Table 2. The learning rate gradually decreases with the number of iterations. Each training round iterates 3700 times, and pseudo-labels are updated after the training is completed. The total training rounds are 4.

[0172] Table 2

[0173] Adam 0.001 8 3700 4

[0174] After the model is trained on the accurate segmentation label set and pseudo-label set, the saved model is used to update the pseudo-labels for fuzzy regions. The continuity and reliability of historical pseudo-labels are fully considered, and high-confidence pseudo-labels are selected to fine-tune the next generation of the model, thereby improving the model's understanding of difficult-to-segment regions and the segmentation effect.

[0175] S42. In order to make full use of the pixel features of the blurred region and improve the generalization ability and segmentation effect of the model, based on the model trained in step S41, an iterative pseudo-label update method is used to select high-confidence pseudo-labels to fine-tune the next generation of the model.

[0176] The pseudocode for the iterative pseudo-label update algorithm is as follows: Figure 8 As shown.

[0177] The magnetic resonance angiography (MRA) images from the target domain AVM dataset are input into the model trained in step S41 for prediction and segmentation. First, the class probability Prob = Net(G) of each pixel in the image is calculated using the dual-branch network model of the target domain.

[0178] Here, Net represents a two-branch network model trained on the target domain dataset G.

[0179] After obtaining the current Prob, initialize the uncertain mask M in the current iteration based on the uncertain mask M and pseudo-label Ps from the previous iteration. Cur and pseudo-tags Ps Cur In the first iteration, the uncertain mask M is set as the image mask for the fuzzy segmentation label set, and the label obtained by the maximum probability estimation of Prob in the first iteration is used as the pseudo label Ps of the first iteration. In subsequent iterations, the output of the previous iteration is used as the input Prob and M.

[0180] Then, a pre-set segmentation threshold of 0.8 is used to filter high-confidence pixel labels, selecting categories that meet the criteria (Prob > 0.8) as the current pseudo-labels, and setting their uncertainty mask M. Cur A value of 0 for the corresponding pixel position indicates that the label at that pixel position is determined; otherwise, an uncertain mask M is set. Cur A pixel value of 1 indicates that the label at that pixel position is uncertain.

[0181] Finally, comparing the previous generation and the current pseudo-labels, if the pseudo-label category changes, the mask M will be uncertain. Cur Setting the corresponding pixel value to 1 indicates that the label at that pixel position is uncertain, resulting in the final pseudo-label set Ps. Cur and uncertain mask M Cur The pseudo-label set Ps in the current iteration Cur and uncertain mask M Cur Ps and M serve as input conditions for the next iteration.

[0182] S43. Based on the accurate segmentation label set and the updated pseudo-label set, a new round of model fine-tuning is carried out. After four preset model training and pseudo-label updates, the model performance can be greatly improved, and a deep learning model for segmentation of cerebral arteriovenous malformation vessels is obtained.

[0183] The model fine-tuning employs a weakly supervised learning method, which integrates both precise segmentation of the label set and pseudo-label set for model training. Both the precise segmentation of the label set and the pseudo-label set are used as the gold standard in the calculation of the loss function. In its training loss function, X... GThis needs to be obtained by fusing the precise segmentation label set and the pseudo-label set, that is, replacing the labels of blurred regions in the image with the corresponding labels in the pseudo-label set. The training parameters of the model fine-tuning process are consistent with those in step S41.

[0184] S44. Input the magnetic resonance angiography (MRA) images of AVM into the trained bi-branch model for the target domain (i.e., the deep learning model for segmenting arteriovenous malformation vessels in the brain) to complete the semantic segmentation of the images.

[0185] The segmentation result in this embodiment is as follows: Figure 9 As shown, the first row of images is the original magnetic resonance angiography (MRA) image, and the second row of images is the segmentation result corresponding to the first row of images. The final test results have DSC, ASD, and HD95 values ​​of 88.70%, 1.43 mm, and 3.88 mm, respectively.

[0186] In summary, compared with existing segmentation techniques, the method of this invention uses statistical segmentation methods to initially obtain precise segmentation label sets and fuzzy segmentation label sets, eliminating the need for professional doctors to manually label a large number of blood vessels, reducing annotation costs and requirements, and greatly alleviating the annotation burden. Simultaneously, it combines contrastive learning and domain adaptation methods to address the problem of limited case data for cerebral arteriovenous malformations, using existing normal blood vessel datasets to fully extract domain-invariant features, improving the model's generalization ability and segmentation performance. Furthermore, the weakly supervised learning method based on pseudo-label updates introduces fuzzy region information into the model training process, ensuring the historical continuity and reliability of the pseudo-labels, thus improving the model's understanding of features in difficult-to-segment regions.

[0187] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A semantic segmentation method for magnetic resonance images of cerebral arteriovenous malformations, the specific steps of which are as follows: S1. Preprocessing of magnetic resonance angiography (MRA) images of cerebral arteriovenous malformations; S2. Based on the preprocessed magnetic resonance angiography (MRA) images from step S1, a low-order statistical model is constructed to obtain the class likelihood probability of pixels, and the initial segmentation result is calculated based on the maximum likelihood probability. S3. Based on the initial segmentation results of step S2, a high-order statistical model is constructed using Markov random fields and inverse multi-scale filtering operators to obtain the precise segmentation label set and the fuzzy segmentation label set. S4. Based on the precise segmentation label set and fuzzy segmentation label set obtained in step S3, the domain adaptive method, contrastive learning method and weakly supervised learning method are integrated to train the dual-branch network model, thereby obtaining a dual-branch network model for segmentation of arteriovenous malformations in the brain, and completing the semantic segmentation of magnetic resonance images. S41. Based on contrastive learning and domain adaptation algorithms, the dual-branch model is trained using a precise segmentation label set and a public dataset; Normal brain magnetic resonance angiography (MRA) images after Min-Max normalization are introduced as the source domain in domain adaptive learning. This source domain dataset is derived from a public dataset, while the AVM dataset is used as the target domain. A domain adaptive framework is constructed by using the source domain and the target domain. Then, by combining contrastive learning and domain adaptation algorithms to extract domain-invariant features, two bi-branch models are trained using source and target domain datasets respectively. The dual-branch model adds a parallel network layer to the bottleneck layer of the U-Net skeleton network. The bottleneck layer consists of two parallel network layers: an independent layer and a shared layer. The shared layer is used to extract domain-invariant features between the source domain and the target domain. The two dual-branch models use a shared layer to extract domain-invariant features. The independent layer extracts domain-specific features for their respective datasets. The two models each have their own independent layer. In the domain adaptive learning process, a contrastive learning method is used, where the outputs of independent layers and shared layers are set to be negative pairs, and the output features of shared layers in different domains are set to be positive pairs. Through subsequent model training, the similarity between positive pairs of features is maximized, and the similarity between negative pairs of features is minimized. The comprehensive loss function is obtained by weighting and combining the comparison loss function, the loss function of the source domain data, and the loss function of the target domain data. The parameter update of the network model is achieved by minimizing the comprehensive loss function. After a pre-set number of training times, the optimal dual-branch model during the training process is output. S42. Based on the model trained in step S41, use an iterative pseudo-label update method to select high-confidence pseudo-labels to fine-tune the next generation of the model. The magnetic resonance angiography (MRA) images from the target domain AVM dataset are input into the model trained in step S41 for prediction and segmentation. First, the class probability of each pixel in the image is calculated using a dual-branch network model in the target domain. ; in, Indicates targeting the dataset of the target domain The trained dual-branch network model; In obtaining the current Then, based on the uncertain mask from the previous iteration... and pseudo-tags Initialize the uncertain mask in the current iteration and pseudo-tags ; Then, a preset segmentation threshold is used to filter high-confidence pixel labels, selecting those that meet the criteria. The category is used as the current pseudo-label, and its indeterminate mask is set. A value of 0 at the corresponding pixel position indicates that the label at that pixel position is determined; otherwise, an uncertain mask is set. A pixel value of 1 indicates that the label at that pixel position is uncertain. Finally, comparing the previous generation and the current pseudo-tags, if the pseudo-tag category changes, the mask will become uncertain. Setting the corresponding pixel value to 1 indicates that the label at that pixel position is uncertain, thus obtaining the final pseudo-label set. and uncertain masks The pseudo-label set and uncertain mask of the current iteration serve as input conditions for the next iteration. and ; S43. Based on the accurate segmentation label set and the updated pseudo-label set, a new round of model fine-tuning is carried out. After model training and pseudo-label updating for a preset number of iterations, a deep learning model for segmentation of cerebral arteriovenous malformation vessels is obtained. The model fine-tuning employs a weakly supervised learning method, which integrates the precise segmentation label set and the pseudo-label set for model training, and uses the precise segmentation label set and the pseudo-label set as the gold standard in the calculation of the loss function. S44. Input the magnetic resonance angiography (MRA) images from the AVM into the trained dual-branch model targeting the target domain to complete the semantic segmentation of the images.

2. The method for semantic segmentation of magnetic resonance images for arteriovenous malformations of the brain according to claim 1, characterized in that, The specific steps of S1 are as follows: First, gamma correction is applied to the initial magnetic resonance angiography (MRA) images to adjust the image brightness and contrast, highlighting the target vascular region. Then, median filtering is used to reduce noise while preserving image details. Next, the FSL skull removal algorithm is used to extract the brain image. Finally, Min-Max normalization is applied to the images to scale the image intensity to the range of 0 to 255, eliminating intensity differences between images.

3. The semantic segmentation method for magnetic resonance images of cerebral arteriovenous malformations according to claim 1, characterized in that, Step S2 is as follows: Based on the preprocessed magnetic resonance angiography (MRA) images obtained in step S1, a low-order statistical model is constructed to perform preliminary statistical segmentation and obtain initial segmentation results. The low-order statistical model is a finite mixture model based on the image pixel intensity histogram, combining Rayleigh and Gaussian distributions. First, using prior knowledge of magnetic resonance angiography (MRA) images, the pixel intensity histogram of the MRA images is divided into three parts according to pixel intensity: low pixel intensity regions correspond to cerebrospinal fluid, bones, and background air; medium intensity regions correspond to brain tissue, including gray and white matter, and part of the eye; and high intensity regions correspond to blood vessels and subcutaneous tissue. Then, based on the finite mixture model, the EM algorithm is used to perform high-precision fitting of the pixel intensity distribution; The high-intensity region distribution function of the finite mixture model is used to calculate the vessel category likelihood function, while the other distribution functions of the finite mixture model are combined to calculate the background category likelihood function. The finite mixture model is given by a Rayleigh distribution. and three Gaussian distributions Its structure, and its function expression are as follows: ； in, and These represent the weights that control the fit of each distribution function, and satisfy the weight normalization constraint. , Represents pixels The intensity value, whose Rayleigh and Gaussian distributions are expressed as follows: ； Among them, the weight parameters and , and The distribution parameters are initialized using histogram distribution features and K-Means clustering algorithm, and the pixel distribution is fitted using the expectation-maximization (EM) algorithm. After a set number of iterations of the EM algorithm, the parameters are updated based on the previous iteration in each iteration, and finally an accurate approximation of the pixel distribution is obtained. Then, based on the distribution function of the fitted finite mixture model, the blood vessel category of each pixel is calculated. and background categories Likelihood probability The expression is as follows: ； in, Represents pixels The possible classification categories, the likelihood probability of all pixels in the image constitutes the likelihood probability map, and the value of each pixel in the likelihood probability map is the likelihood probability of that pixel; Finally, the pixel points are obtained through maximum likelihood probability estimation. The categories are used as the classification results of the low-order statistical model, i.e., the initial segmentation results. The expression is as follows: 。 4. The semantic segmentation method for magnetic resonance images of cerebral arteriovenous malformations according to claim 1, characterized in that, Step S3 is as follows: S31. Construct a higher-order statistical model using Markov random field and inverse multi-scale filtering operator. Input the initial segmentation results and likelihood probability map obtained in step S2 into the higher-order statistical model to obtain the accurate segmentation label set. First, based on the inverse multi-scale filtering operator, the intensity response of the blood vessel edge is calculated. Then, the image is flipped pixel by pixel intensity, setting the background as a high-intensity region and the blood vessel as a low-intensity region. ； in, Indicates the original image in The pixel intensity value at the location. This indicates the transpose operation. This represents the pixel intensity value after flipping. Multi-scale filtering is applied to the flipped image to enhance the target edges, and the Hessian matrix of the filtered response is used to obtain the enhancement. The expression is as follows: ； ； Where * denotes convolution operation, Describes the scale of the Gaussian operator; The scale is represented as Gaussian kernel; For each pixel, the Hessian matrix It can be decomposed into two eigenvalues: The eigenvalue relationship is used to distinguish different structural features; then the filter operator response value The expression is as follows: ； in, The regularization term used to reduce noise sensitivity is expressed as follows: ； in, This represents a pre-defined parameter used to control the regularization strength; its value is between 0 and 1. When applying multiple filtering operators of different scales, the pixel with the largest response is selected. The final filtered response value , Represents the set of positive real numbers; Then, the inverse multi-scale filter response value is fused with the Markov random field to obtain the pixel-level potential function, as shown in the following expression: ； in, Represents the center pixel Possible classification categories; This indicates that in the initial segmentation results obtained based on a low-order statistical model, 26-neighborhood of a pixel pixel The classification results at the location, ; Represents the center pixel Classification categories Classification of neighboring pixels The relationship between the two labels is such that if the two labels have the same category, the value is -1, otherwise it is 1. Indicates the control of multi-scale filtering response The ability to suppress and enhance subsequent prior probabilities, if but To suppress subsequent prior probabilities, and vice versa. This enhances the probability of subsequent prior knowledge. and These represent the weights, used to weigh the effects of the Markov random field operator and the multi-scale filtering operator; The center pixel is obtained by summing the potential functions related to all neighborhoods. Energy function The expression is as follows: ； Then, based on the Hammersley-Clifford theorem, the pixel points are obtained. Prior probabilities for each category The expression is as follows: ； in, Represents the normalization constant, obtained by summing all Category of molecules The result is used to obtain the prior probability. Normalize to the (0, 1) interval; Based on the likelihood and prior probabilities obtained from the aforementioned low-order and high-order statistical models, the center pixel is then calculated using Bayes' theorem. The posterior probability value is then used, and the class with the highest posterior probability is selected as the pixel using the maximum a posteriori probability algorithm. The final segmentation result The expression is as follows: ； in, Indicates a direct proportional relationship; by comparing pixel intensity Blood vessel categories under conditions and background categories The final segmentation result is determined by the posterior probability, with the category label corresponding to the highest posterior probability being selected. The accurate segmentation label set is obtained by performing maximum a posteriori probability estimation on all pixels of the image. ; S32. Based on the precise segmentation label set obtained in step S31 and the likelihood probability map obtained in step S2, a fuzzy segmentation label set is obtained through weighted maximum likelihood probability estimation and Boolean difference operation. The expanded segmentation result, i.e., the expanded foreground label set, is obtained by performing weighted maximum likelihood probability estimation on all pixels in the image. The expression is as follows: ； in, Indicates the weighted weight. This expands the blood vessel category area and shrinks the background category area; Then expand the foreground label set Compared with the precise segmentation tag set obtained in step S31 Perform Boolean difference operation and convert the result of the Boolean difference operation. This serves as a fuzzy segmentation label set that requires subsequent relabeling using deep learning methods.

Images