A brain magnetic resonance image segmentation method based on scattering map neural network

Abstract

The present invention discloses a brain magnetic resonance image segmentation method based on a scattering graph neural network. Firstly, in order to reduce the computational complexity of the model, a supervoxel is used as a basic unit, and a corresponding number of supervoxels are generated for a reference image and an image to be segmented. Meanwhile, the supervoxel's own information, surrounding neighbor information and spatial position information are considered, and its grayscale features, tensor features and key point spatial prior features are pre-extracted. Secondly, since certain topological structure information is implicit between each supervoxel of the brain, a topological map is constructed with supervoxels as nodes, and a scattering graph neural network is used to learn global topological information and update node features. Finally, the supervoxels of the image to be segmented are directly feature matched with the supervoxels of the labeled reference image to obtain a semantic segmentation result. The present invention can be well applied to brain magnetic resonance images and effectively segment the tissue structure of the brain magnetic resonance images.

Description

Technical Field

[0001] The invention relates to a brain magnetic resonance image segmentation method based on a scattering graph neural network, and belongs to the field of digital images. Background Art

[0002] In real life, there are unstructured data, such as social networks, molecular structures and other graph-structured data. For this kind of graph-structured data, each node has a different topological structure (the number of neighbor nodes, the distribution of neighbor nodes, etc. are different). Therefore, unlike the convolutional neural network (CNN), the graph convolutional neural network (GCN) does not have translation invariance and cannot use parameter-sharing convolution kernels. GCN obtains a new node representation by aggregating the features of each node with its neighbor nodes, and embeds the topological structure information between nodes while each node is passing messages. However, when integrating graph structure information, GCN will strengthen the similarity between adjacent nodes, essentially achieving local smoothing on the graph, without distinction between nodes, and when the network structure deepens, the problem of over-smoothing will occur. The scattering graph neural network introduces the geometric scattering transformation on the basis of GCN, so that the learned node-level features contain geometric features in addition to the smooth activation signal, effectively alleviating the over-smoothing problem in GCN. Recently, the scattering graph neural network has achieved excellent results in node classification and graph classification tasks, but there has been no research on its application in brain magnetic resonance segmentation.

[0003] The segmentation of magnetic resonance images is an important step in medical diagnosis and research. It can help analyze tissues and is the key to disease detection and diagnosis. The tissue structure in magnetic resonance images is complex, and experts with professional knowledge are required to perform correct segmentation. In addition, magnetic resonance images are easily affected by noise, and accurate segmentation is very difficult. Manual tissue segmentation of a three-dimensional image usually takes an expert several hours. Therefore, the cost of obtaining segmentation and annotation of magnetic resonance images is very high. In order to achieve accurate automatic segmentation, deep learning has become the most popular image semantic segmentation method in recent years and has been widely used in the research of medical image segmentation. Deep learning methods can learn task-related features from data, and greatly exceed traditional segmentation algorithms in segmentation effects, but deep learning models require a large amount of pixel-level annotated data to achieve ideal results. Considering that magnetic resonance images of the same tissue of different individuals have high structural similarities, some scholars have proposed the use of image matching methods, which only requires the reference image to be annotated, and the image to be segmented is matched with the reference image to obtain the segmentation result, which can greatly reduce the annotation cost.

[0004] At present, traditional image matching algorithms, such as grayscale-based template matching algorithms and feature-based matching algorithms such as SIFT and KLT, have been applied to many fields of computer vision and have achieved quite good results. However, this kind of matching based on pixels has a high time complexity and is not applicable to large three-dimensional magnetic resonance images.

[0005] In summary, deep learning models require a large amount of pixel-level labeled data to obtain ideal results, and the cost of obtaining segmentation and annotation of magnetic resonance images is very high. Therefore, image matching methods can be considered to reduce the annotation cost, and supervoxels can be used as the basic unit to reduce the matching complexity. At the same time, current research has not yet disclosed how to use the topological structure information between supervoxels in magnetic resonance images. Consider using scatter graph neural networks to learn the topological structure between supervoxels, so as to more comprehensively describe the supervoxel characteristics. Summary of the invention

[0006] The present invention is aimed at the problems in the prior art and provides a brain magnetic resonance image segmentation method based on a scattering graph neural network. First, in order to reduce the computational complexity of the model, supervoxels are used as the basic unit, and a corresponding number of supervoxels are generated for the reference image and the image to be segmented; at the same time, the supervoxel's own information, surrounding neighbor information and spatial position information are considered, and its grayscale features, tensor features and key point spatial prior features are pre-extracted; thirdly, since certain topological structure information is implicit between each supervoxel of the brain, a topological map is constructed with supervoxels as nodes, and a scattering graph neural network is used to learn global topological information and update node features; finally, the supervoxels of the image to be segmented are directly matched with the supervoxels of the labeled reference image to obtain a semantic segmentation result. The present invention can be well applied to brain magnetic resonance images and effectively segment the tissue structure of brain magnetic resonance images.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is: a brain magnetic resonance image segmentation method based on a scattering map neural network, comprising the following steps

[0008] S1, extracting supervoxels: extracting supervoxels from the reference image and the image to be segmented respectively, wherein the extraction method adopts a fuzzy iterative clustering method, uniformly sampling seed points in the brain region of the magnetic resonance image, fuzzily associating each voxel with the seed point closest to it in space, calculating the fuzzy membership of each voxel and the seed, updating the spatial coordinates and grayscale value of the seed point according to the fuzzy membership, iteratively updating, and after reducing the error, assigning each voxel to the seed point with the maximum fuzzy membership, that is, generating supervoxels of the magnetic resonance image;

[0009] S2, extracting characteristic values: extracting characteristic values ​​of the supervoxels respectively according to the supervoxels obtained in S1, wherein the characteristic values ​​at least include a grayscale histogram, a structured tensor feature, and a key point space prior feature of the supervoxel;

[0010] S3, constructing a topological map and learning global topological features: using the reference image obtained in step S1 and the supervoxels of the image to be segmented as nodes, splicing the grayscale histogram, structured tensor features and key point spatial prior features of the supervoxels as node features to construct a topological map, and using a scattering map neural network to learn global topological features;

[0011] S4, get the semantic segmentation result: construct the feature distance matrix D between the supervoxel of the image to be segmented and the supervoxel of the reference image, find the column with the smallest value for each row of the matrix D, and use it as the supervoxel T in the reference image that best matches the supervoxel of the image to be segmented. i , and the supervoxel category of the image to be segmented is recorded as T i Finally, the labeled supervoxel categories are mapped to voxels to obtain the final semantic segmentation result.

[0012] As an improvement of the present invention, the step S1 comprises:

[0013] S11, uniformly sample N seed points in the brain region of the magnetic resonance image;

[0014] S12, calculating the distance between each voxel in the magnetic resonance image and various seed points, wherein the distance D(i, j) between the i-th voxel and the j-th seed point can be expressed as: D(i, j)=d I (i,j)+λd s (i, j), where d I (i, j) is the grayscale distance between the i-th voxel and the j-th seed point, d s (i, j) is the spatial distance between the i-th voxel and the j-th seed point, and λ is the weight of the spatial distance;

[0015] S13, fuzzy association is performed between each voxel in the magnetic resonance image and the K seed points closest to it in space, that is, the fuzzy membership of each voxel with these K seeds is calculated.

[0016]

[0017] Among them, D(i, S j ) represents the relationship between the i-th voxel and the seed point S j The distance between them, m represents the fuzzy weighted index of fuzzy membership, S j , S t ∈S, S={s1,s2,…,s K}, S represents the set of K seed points closest to the i-th voxel;

[0018] S14, according to the fuzzy membership calculated in step S13, update the spatial coordinates and grayscale value of the seed point:

[0019]

[0020] in, Represents the seed point S in the magnetic resonance image space j The number of fuzzy-connected voxels, Represents the seed point S j The fuzzy membership of the rth voxel of the fuzzy association; for v r =[x r ,y r , z r , I r ] T ,(x r ,y r , z r ) represents the seed point S j The coordinates of the rth voxel of the fuzzy association, I r Represents the seed point S j The gray value of the rth voxel of the fuzzy association;

[0021] S15, iteratively updating the seed points of the magnetic resonance image according to steps S13 and S14, and defining an update error. When the update error is less than a set threshold, the iteration is stopped, and each voxel is assigned to the seed point with the maximum fuzzy membership, so as to generate a supervoxel of the brain magnetic resonance image.

[0022] As another improvement of the present invention, the grayscale histogram extraction method of the supervoxel in step S2 is: the grayscale values ​​of 0 to 255 are evenly divided into 16 intervals P[1:16], for each supervoxel, the number of voxels whose grayscale values ​​fall within each interval is counted, mapped to the 16-dimensional interval P, and then the value of each dimension is divided by the total number of voxels for normalization;

[0023] The method for extracting the structured tensor features of the supervoxel in step S2 is: using the Pearson correlation coefficient to calculate the correlation C between the supervoxel n and its adjacent supervoxel i ni , using the center coordinate vector (x ni ,y ni , z ni ), and obtain the spatial relationship vector M between the two ni , and then calculate the tensor feature T n , where the correlation C ni It can be expressed as:

[0024]

[0025] Among them, His is the grayscale feature;

[0026] The spatial relationship vector M between supervoxel n and its adjacent supervoxel i ni It can be expressed as:

[0027]

[0028] Tensor feature T n It can be expressed as:

[0029]

[0030] The method for extracting the spatial prior features of the key points in step S2 is: using the SIFT3D algorithm to match the key points of the reference image and the image to be segmented, and then calculating the spatial coordinate distance d between the supervoxel n of the image to be segmented and each key point i in the image ni ,

[0031]

[0032] As a further improvement of the present invention, step S3 further comprises:

[0033] S31, for the reference image, randomly select K similar supervoxels from the adjacent supervoxels of each supervoxel for connection;

[0034] S32, for the image to be segmented, connecting each supervoxel with K supervoxels with the smallest feature distance in the image;

[0035] S33, for the reference image and the image to be segmented, the supervoxel of the image to be segmented is connected with K supervoxels of the reference image with the smallest feature distance;

[0036] S34, taking the supervoxel as a node, the connection relationship constructed in steps S31, S32, and S33 as an edge, and the three features in step S2 as a node feature, a scattering graph neural network is used to learn the topological relationship of the supervoxels in the image to be segmented and between the image to be segmented and the reference image, and the feature of each supervoxel is updated;

[0037] First, define the random walk matrix R = AD -1 , then the lazy random walk matrix P can be expressed as On this basis, the wavelet matrix is ​​defined as:

[0038]

[0039] (Ψ k X)[v i ] is for the i-th node to collect 2k The information of the order neighbors is obtained through a tuple J containing the scale parameter J = (k1, k2, ..., k m ), stacked wavelet transform:

[0040]

[0041] According to the wavelet matrix, the process of scattering propagation is (H l represents the node features output by the lth layer, are the learnable parameters of the model):

[0042]

[0043] Among them, A represents the node adjacency matrix, and D represents the node degree matrix

[0044] Compared with the prior art: the present invention uses feature matching technology to achieve semantic segmentation of magnetic resonance images, that is, the unlabeled image obtains the segmentation result by matching the features with the labeled image, which effectively reduces the annotation cost of medical images; on this basis, it overcomes the matching based on pixels and uses supervoxels as the basic unit, which reduces the time complexity of the matching process and improves the model segmentation efficiency; in addition, the scattering map neural network is innovatively introduced into the segmentation task of magnetic resonance images. On the basis of considering the grayscale features, neighborhood features and spatial position features of supervoxels, the scattering map neural network is used to learn the topological structure between each supervoxel, so as to more comprehensively describe the supervoxel features and effectively retain the features with high discrimination. Compared with the prior art, the present invention can be better applied to brain magnetic resonance images and effectively segment the tissue structure of brain magnetic resonance images. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is a schematic diagram of the implementation process of the present invention;

[0047] Figure 3 The brain MRI segmentation results of the method proposed in this invention, where (a) is the original MRBrainS18 brain MRI image, and (b) is the real annotation result.

[0048] This is followed by a series of comparison methods: (c) FSL, (d) SegNet, (e) Unet, (f) SuperPatchMatch, (g) SRBD and (h) the segmentation results obtained by the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0050] Example 1

[0051] A brain magnetic resonance image segmentation method based on scatter graph neural network, such as Figure 1 and Figure 2 As shown, the following steps are included:

[0052] First, step S1: the method of extracting supervoxels is as follows: for the reference image and the image to be segmented, the fuzzy iterative clustering method is used to extract their supervoxels respectively, including the following five steps:

[0053] (1-1) N seed points are uniformly sampled in the brain region of the brain magnetic resonance image;

[0054] (1-2) Calculate the distance between each voxel and various seed points in the brain magnetic resonance image, where the distance D(i, j) between the i-th voxel and the j-th seed point can be expressed as: D(i, j) = d I (i,j)+λd s (i, j), where d I (i, j) is the grayscale distance between the i-th voxel and the j-th seed point, which can be expressed as:

[0055] d I (i, j) = |I i -I j |

[0056] d s (i, j) is the spatial distance between the i-th voxel and the j-th seed point, which can be expressed as:

[0057]

[0058] λ is the weight of spatial distance and grayscale distance;

[0059] (1-3) Fuzzily associate each voxel in the magnetic resonance image with the K seed points closest to it in space, and calculate the fuzzy membership of each voxel with these K seeds.

[0060]

[0061] Among them, D(i, S j ) represents the relationship between the i-th voxel and the seed point S j The distance between them, m represents the fuzzy weighted index of fuzzy membership, S j , S t ∈S, S={s1,s2,…,sK}, S represents the set of K nearest seed points of the i-th voxel;

[0062] (1-4) Use the fuzzy membership calculated in step 1-3 to update the seed point and obtain:

[0063]

[0064] in, Represents the seed point S in the magnetic resonance image space j The number of fuzzy-connected voxels, Represents the seed point S j The fuzzy membership of the rth voxel of the fuzzy association. r =[x r ,y r , z r , I r ] T ,(x r ,y r , z r ) represents the seed point S j The coordinates of the rth voxel of the fuzzy association, I r Represents the seed point S j The gray value of the rth voxel of the fuzzy association;

[0065] (1-5) Iteratively update the seed points of the magnetic resonance image according to steps 1-3 and 1-4, and define an update error. When the update error is less than a set threshold, stop the iteration, and assign each supervoxel to the seed point with the maximum fuzzy membership as the supervoxel of the brain magnetic resonance image. In the present invention, the update error is set to 0.01.

[0066] Secondly, step S2, accurate supervoxel matching puts forward higher requirements on the features of supervoxels. The present invention further considers the supervoxel itself, local neighbor information and its global spatial position, extracts the grayscale histogram, structured tensor features and key point spatial prior features of the supervoxel, including the following three steps:

[0067] (2-1) For the super-voxel grayscale feature, the grayscale values ​​0 to 255 are evenly divided into 16 intervals P[1:16], the number of voxels in each interval is counted, mapped to the 16-dimensional interval P, and then the value of each dimension is divided by the total number of voxels for normalization:

[0068]

[0069] (2-2) Based on the structured tensor features of the supervoxel, the Pearson correlation coefficient is used to calculate the correlation C between the supervoxel and its adjacent supervoxels. ni, using the center coordinate vector (x ni ,y ni , z ni ), and obtain the spatial relationship vector M between the two ni , combined with the correlation C ni Then calculate the tensor feature T n The correlation C ni It can be expressed as:

[0070]

[0071] The Pearson correlation coefficient can be specifically expressed as two supervoxel features f n , f i The covariance cov(f n , f i ) multiplied by their respective standard deviations Business:

[0072]

[0073] Assume (x n ,y n , z n ) is the coordinate of supervoxel n, (x i ,y i , z i ) is the coordinate of supervoxel i, then the center coordinate vector (x ni ,y ni , z ni ) can be expressed as:

[0074]

[0075]

[0076] The spatial relationship vector M between supervoxel n and its adjacent supervoxel i ni It can be expressed as:

[0077]

[0078] Tensor feature T n It can be expressed as:

[0079]

[0080] (2-3) Based on the spatial prior features of key points, the SIFT3D algorithm is used to match the key points of the reference image and the image to be segmented, and the spatial coordinate distance d between the supervoxel n of the image to be segmented and each key point i in the image is calculated. ni , This describes the global position of the supervoxel in the entire image.

[0081] Then, in step S3, in order to better learn the topological information of the supervoxel, the supervoxels of the reference image and the image to be segmented are used as nodes, the three features of step 2 are spliced ​​as node features, and a topological map is constructed, which includes the following four steps:

[0082] (3-1) For the reference image (known label), randomly select K similar supervoxels from the neighboring supervoxels of each supervoxel for connection;

[0083] (3-2) For the image to be segmented, each supervoxel is connected to the K supervoxels with the smallest feature distance in the image;

[0084] (3-3) For the reference image and the image to be segmented, the supervoxel of the image to be segmented is connected with the K supervoxels of the reference image with the smallest feature distance;

[0085] (3-4) Taking supervoxels as nodes, the connection relationships constructed in steps 3-1, 3-2, and 3-3 as edges, and the three features of step 2 as node features, a scattering graph neural network is used to learn the topological relationship of supervoxels in the image to be segmented and between the image to be segmented and the reference image, and update the features of each supervoxel. The scattering graph neural network can learn bandpass features to avoid the over-smoothing problem of graph convolutional neural networks. First, define the random walk matrix R = AD -1 (A is the node adjacency matrix, D is the node degree matrix), then the lazy random walk matrix P can be expressed as On this basis, the wavelet matrix is ​​defined as:

[0086]

[0087] (Ψ k X)[v i ] is for the i-th node to collect 2 k The information of the order neighbors is obtained through a tuple J containing the scale parameter J = (k1, k2, ..., k m ), stacked wavelet transform:

[0088]

[0089] According to the wavelet matrix, the feature update process based on scattering propagation is (H l represents the node features output by the lth layer, are the learnable parameters of the model):

[0090]

[0091] Finally, step S4 constructs the supervoxel f of the image to be segmented refWith the reference image supervoxel f tar The characteristic distance matrix D is:

[0092]

[0093] Where N ref represents the number of supervoxels in the reference image, N tar represents the number of supervoxels in the image to be segmented, d ij represents the feature distance between supervoxels i and j. The feature distance is calculated based on the Euclidean distance between features. For each row of D, find the column with the smallest value as the supervoxel T in the reference image that best matches the supervoxel in the image to be segmented. i , and the supervoxel category of the image to be segmented is recorded as T i The supervoxel categories are mapped to voxels to obtain the final semantic segmentation result.

[0094] In the present invention, firstly, supervoxels are used as the basic unit, and a corresponding number of supervoxels are generated for the reference image and the image to be segmented. Secondly, in order to better extract supervoxel features, the supervoxel's own information, surrounding neighbor information and spatial position information are considered at the same time, and grayscale features, tensor features and key point spatial prior features are pre-extracted as supervoxel features. Afterwards, considering that certain topological information is implicit between each supervoxel of the brain, the supervoxels of the reference image and the image to be segmented are used as nodes, and the connection relationship between the nodes is determined according to their characteristics, and a topological graph is constructed. The graph structure and node features are used as inputs of the scattering graph neural network, and the topological structure information is learned to update the features of the supervoxels. Finally, the feature distance of each pair of supervoxels between the image to be segmented and the reference image is calculated, and the supervoxels with the smallest feature distance are matched with each other, and the two are recorded as the same category. The annotations at the supervoxel level are mapped to the pixel level to obtain the semantic segmentation result.

[0095] Example:

[0096] The following takes the MRBrainS18 data set as an example to illustrate a brain magnetic resonance image segmentation method based on a scattering map neural network of the present invention.

[0097] Experimental conditions: Python and Matlab are used to implement the method in this paper, and the Tensorflow framework is used to implement the scattering graph neural network. The specific software and hardware environment is as follows:

[0098] (1) Software development environment: Windows 10 operating system, Visual Studio Code, Python 3.6, Matlab 2015b, Tensorflow 1.12

[0099] (2) Hardware equipment: Nvidia 1080Ti GPU, i9-9900K CPU, 64G memory

[0100] Experimental data: Taking the brain magnetic resonance images of the MRBrainS18 dataset as an example, each MRI image includes 240×240×48 voxels of size 1mm×1mm×1mm. The experimental parameters are set as follows: supervoxel setting (the number of supervoxels is 8000, K=6; m=2, λ=0.7, sigma=1), the number of supervoxel neighbors is 5, and the graph neural network adopts the scattering graph neural network. Figure 3 From left to right, they are (a) the original MRBrainS18 brain MRI image, (b) the real annotation result GroundTruth, and the comparison methods: (c) FSL, (d) SegNet, (e) Unet, (f) SuperPatchMatch, (g) SRBD, and (h) the segmentation results obtained by the present invention. It can be seen from the figure that compared with the prior art, the magnetic resonance segmentation method based on the scattering graph neural network proposed in the present invention can obtain the brain contour more accurately and segment the brain tissue more cleanly and effectively while improving the segmentation efficiency and reducing the number of sample annotations.

[0101] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A magnetic resonance image segmentation method based on a scattering map neural network, characterized in that: The following steps are involved: S1, extract supervoxels: extract supervoxels from the reference image and the image to be segmented respectively. The extraction method adopts fuzzy iterative clustering method, uniformly samples seed points in the brain area of ​​the magnetic resonance image, fuzzily associates each voxel with its nearest seed point in space, calculates the fuzzy membership of each voxel and the seed, updates the spatial coordinates and grayscale value of the seed point according to the fuzzy membership, iteratively updates, and after reducing the error, assigns each voxel to the seed point with the maximum fuzzy membership, that is, generates the supervoxel of the magnetic resonance image; S2, extracting characteristic values: extracting characteristic values ​​of the supervoxels respectively according to the supervoxels obtained in S1, wherein the characteristic values ​​at least include a grayscale histogram, a structured tensor feature, and a key point space prior feature of the supervoxel; S3, constructing a topological map and learning global topological features: taking the reference image obtained in step S1 and the supervoxels of the image to be segmented as nodes, splicing the grayscale histogram, structured tensor features and key point spatial prior features of the supervoxels as node features to construct a topological map, and using a scattering map neural network to learn global topological features; S4, get the semantic segmentation result: construct the feature distance matrix D between the supervoxel of the image to be segmented and the supervoxel of the reference image, find the column with the smallest value for each row of the matrix D, and use it as the supervoxel T in the reference image that best matches the supervoxel of the image to be segmented. i , and the supervoxel category of the image to be segmented is recorded as T i Finally, the labeled supervoxel categories are mapped to the corresponding voxels to obtain the final semantic segmentation result.

2. A magnetic resonance image segmentation method based on a scattering map neural network as claimed in claim 1, characterized in that: The step S1 comprises: S11, uniformly sample N seed points in the brain region of the magnetic resonance image; S12, calculating the distance between each voxel in the magnetic resonance image and various seed points, wherein the distance D(i, j) between the i-th voxel and the j-th seed point can be expressed as: D(i, j)=d I (i, j)+λd s (i, j), where d I (i, j) is the grayscale distance between the i-th voxel and the j-th seed point, d s (i, j) is the spatial distance between the i-th voxel and the j-th seed point, and λ is the weight of the spatial distance; S13, fuzzy association is performed between each voxel in the magnetic resonance image and the K seed points closest to it in space, that is, the fuzzy membership of each voxel with these K seeds is calculated. Among them, D(i, S j ) represents the relationship between the i-th voxel and the seed point S j The distance between them, m represents the fuzzy weighted index of fuzzy membership, S j , S t ∈S, S={s1,s2,…,s K }, S represents the set of K seed points closest to the i-th voxel; S14, according to the fuzzy membership calculated in step S13, update the spatial coordinates and grayscale value of the seed point: in, Represents the seed point S in the magnetic resonance image space j The number of fuzzy-connected voxels, Represents the seed point S j The fuzzy membership of the rth voxel of the fuzzy association; for v r =[x r ,y r , z r , I r ] T ,(x r ,y r , z r ) represents the seed point S j The coordinates of the rth voxel of the fuzzy association, I r Represents the seed point S j The gray value of the rth voxel of the fuzzy association; S15, iteratively updating the seed points of the magnetic resonance image according to steps S13 and S14, and defining an update error. When the update error is less than a set threshold, the iteration is stopped, and each voxel is assigned to the seed point with the maximum fuzzy membership, so as to generate a supervoxel of the brain magnetic resonance image.

3. A magnetic resonance image segmentation method based on a scattering map neural network as claimed in claim 2, characterized in that: The grayscale histogram extraction method of the supervoxel in step S2 is: the grayscale values ​​of 0 to 255 are evenly divided into 16 intervals P[1:16], for each supervoxel, the number of voxels whose grayscale values ​​fall within each interval is counted, mapped to the 16-dimensional interval P, and then normalized by dividing the value of each dimension by the total number of voxels.

4. The magnetic resonance image segmentation method based on scattering map neural network according to claim 2, characterized in that: The method for extracting the structured tensor features of the supervoxel in step S2 is: using the Pearson correlation coefficient to calculate the correlation C between the supervoxel n and its adjacent supervoxel i ni , using the center coordinate vector (x ni ,y ni , z ni ), and obtain the spatial relationship vector M between the two ni , and then calculate the tensor feature T n , where the correlation C ni It can be expressed as: Among them, His is the grayscale feature; The spatial relationship vector M between supervoxel n and its adjacent supervoxel i ni It can be expressed as: Tensor feature T n It can be expressed as:

5. The magnetic resonance image segmentation method based on scattering map neural network according to claim 2, characterized in that: The method for extracting the spatial prior features of the key points in step S2 is: using the SIFT3D algorithm to match the key points of the reference image and the image to be segmented, and then calculating the spatial coordinate distance d between the supervoxel n of the image to be segmented and each key point i in the image ni , 6. A magnetic resonance image segmentation method based on a scattering map neural network as claimed in claim 3, 4 or 5, characterized in that: The step S3 further comprises: S31, for the reference image, randomly select K similar supervoxels from the adjacent supervoxels of each supervoxel for connection; S32, for the image to be segmented, connecting each supervoxel with K supervoxels with the smallest feature distance in the image; S33, for the reference image and the image to be segmented, the supervoxel of the image to be segmented is connected with K supervoxels of the reference image with the smallest feature distance; S34, taking the supervoxel as a node, the connection relationship constructed in steps S31, S32, and S33 as an edge, and the three features in step S2 as a node feature, a scattering graph neural network is used to learn the topological relationship of the supervoxels in the image to be segmented and between the image to be segmented and the reference image, and the feature of each supervoxel is updated; First, define the random walk matrix R = AD -1 , then the lazy random walk matrix P can be expressed as On this basis, the wavelet matrix is ​​defined as: (Ψ k X)[v i ] is for the i-th node to collect 2 k The information of the order neighbors is obtained through a tuple J containing the scale parameter J = (k1, k2, ..., k m ), stacked wavelet transform: According to the wavelet matrix, the process of scattering propagation is H l represents the node features output by the lth layer, are the learnable parameters of the model: Among them, A represents the node adjacency matrix, and D represents the node degree matrix.

7. The magnetic resonance image segmentation method based on scattering map neural network according to claim 2, characterized in that: In step S12, the grayscale distance d between the i-th voxel and the j-th seed point I (i, j) = |I i -I j |, where I i and I j Represents the pixel intensity of the i-th seed point and the j-th voxel respectively; the spatial distance between the i-th voxel and the j-th seed point Where (x i ,y i , z i ) and (x j ,y j , z j ) represent the coordinates of the i-th seed point and the j-th voxel respectively.

8. The magnetic resonance image segmentation method based on scattering map neural network according to claim 2, characterized in that: The update error defined in step S15 is the difference between the Euclidean distance of the seed point coordinates after the update and before the update. The error threshold error_threshold is set to 0.

01. When the change in the coordinates before and after the update is less than the threshold, the iteration is stopped.

9. The magnetic resonance image segmentation method based on scattering map neural network according to claim 5, characterized in that: The number of key points is controlled at 20.

10. The magnetic resonance image segmentation method based on scattering map neural network according to claim 6, characterized in that: The feature distance matrix D in step S4 can be expressed as: Where N ref represents the number of supervoxels in the reference image, N tar represents the number of supervoxels in the image to be segmented, d ij It represents the feature distance between supervoxel i and supervoxel j, and the feature distance is calculated according to the Euclidean distance between features.

Images