A full-band DiffusionNet deep learning model construction method for cerebral cortex segmentation
Patent Information
- Application Number
- CN202311524313.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-15
- Publication Date
- 2026-08-21
- Estimated Expiration
- 2043-11-15
AI Technical Summary
[0006]然而，第一种依赖于球面映射方法的缺陷在于：其首先需要将原始皮层表面映射至球面，并将球面上的分割结果再映射回原始皮层表面，球面映射的过程非常耗时，需要1～2小时
本发明不依赖于球面映射，能够直接在原始皮层表面上开展分割任务；本发明额外利用了高频信息，使DiffusionNet在皮层分割任务上的性能有了显著提升；本发明的分割模型对于皮层表面mesh的采样方式是不敏感的，不会因为mesh的拓扑连接发生变化而导致性能下降。
Smart Images
Figure CN117496150B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computational neuroscience technology in medical image processing, and relates to a method for constructing a full-band DiffusionNet deep learning model for cerebral cortex segmentation. Background Technology
[0002] The surface of the cerebral cortex is the highest level of the central nervous system, and research on many neurodegenerative and psychological diseases relies on the analysis of its surface. In clinical medicine, doctors primarily use MRI (Magnetic Resonance Imaging) images acquired by imaging equipment for diagnosis. For sequential MRI scans, experienced neurosurgeons often need to analyze and interpret the images layer by layer to locate lesions and determine appropriate treatment plans. To aid in diagnosis, MRI-based three-dimensional reconstruction and cerebral cortex segmentation have become popular research topics in recent years.
[0003] In existing technologies, deep learning models for cortical segmentation mainly fall into two categories. One category relies on spherical mapping, while the other does not rely on spherical mapping and can perform segmentation directly on the original cortical surface.
[0004] In the first method relying on spherical mapping, to simplify computation when processing the cerebral cortex, the irregular original cortical surface needs to be mapped onto a standard sphere. After atlas-based registration, segmentation is performed, and then the segmentation markers on the sphere are mapped back to the original cortical surface. Based on this, a deep neural network was designed according to the characteristics of the sphere. Taking cortical attributes such as curvature, sulcus depth, and cortical thickness as inputs, the cerebral cortex can be segmented on the sphere. The overall process is as follows: Figure 1 As shown.
[0005] The second method, which does not rely on spherical mapping and can directly segment the surface of the raw cortex, utilizes a 3D triangular mesh. Meshes and graphs share a similar structure; the vertices of a mesh resemble the nodes of a graph, and edges connect the vertices of a mesh and the nodes of a graph. Therefore, the cerebral cortex is treated as graph-structured data, allowing graph neural networks to be directly applied for segmentation of the raw cortical surface. Furthermore, deep learning methods specifically designed for mesh data can also be used for segmentation on the raw cortical surface. These methods borrow from traditional image convolutional neural networks (CNNs), designing corresponding convolution, pooling, and upsampling operators around the basic elements of a mesh—vertices, faces, and edges—and then applying mature backbone network structures from the image domain for raw cortical surface segmentation. The overall process is as follows: Figure 2 As shown.
[0006] However, the first method, which relies on spherical mapping, has a drawback: it first requires mapping the original cortical surface to a sphere, and then mapping the segmentation results on the sphere back to the original cortical surface. This spherical mapping process is very time-consuming, taking 1-2 hours. Moreover, the spherical mapping process involves multiple steps, which can introduce unnecessary errors. The second method, which does not rely on spherical mapping and can be directly applied to the original cortical surface, has a drawback: whether it directly applies graph neural networks or borrows from CNN mesh deep learning methods, it is very sensitive to changes in the mesh structure. Once transferred to cortical surface data with different sampling methods, significant performance degradation will occur due to changes in the mesh's resolution, density, and other properties.
[0007] Therefore, a deep learning model for cortical segmentation that is time-efficient, has low error, and stable performance is needed to solve the above-mentioned technical problems. Summary of the Invention
[0008] The technical solution adopted by this invention to solve the technical problem is: a method for constructing a full-band DiffusionNet deep learning model for cerebral cortex segmentation, comprising the following steps: Step S1: Use the DiffusionNet deep learning model for 3D surfaces as the backbone network; Step S2: DiffusionNet takes the vertex coordinates of the mesh as input, extracts features, and uses them for segmentation. Furthermore, it learns the intrinsic, essential features of the mesh and is insensitive to the mesh sampling method. Specifically, DiffusionNet exhibits consistent performance when dealing with mesh variants of the same object at different resolutions and densities. Step S3: Specifically, regarding the calculation method of spatial diffusion, the efficient spectral acceleration method is selected, using the Laplace transform; for a closed surface S, equation (1) is satisfied: (1) In equation (1), It is a Laplace-Beltrami differentiable operator, and λ is The eigenvalues of λ are such that the solution Φ is the eigenvector corresponding to λ. The eigenvector Φ serves as the basis for the Laplace transform. According to equation (2), the eigenvalues f defined on S are mapped to the Laplace frequency domain to obtain the corresponding coefficients c. (2) At this point, the spectral acceleration method will use the eigenvalue λ to scale the coefficient c with respect to time t, thereby approximating the diffusion with respect to time t, as shown in equation (3): (3) According to equation (4), the feature f is reconstructed using c and λ: (4).
[0009] Preferably, in step S3, the high-frequency information is separated according to equation (4), that is, the high-frequency information is obtained by subtracting the sum of the first k terms on the right side of equation (4) from f on the left side; a parallel MLP is added to the spatial diffusion module, and the high-frequency information is handed over to the MLP for calculation, while the low-frequency information corresponding to the sum of the first k terms on the right side of equation (4) is calculated in the spatial diffusion module as usual; S can have infinitely many eigenvalues and eigenvectors, spectral The acceleration method can only calculate the diffusion of the first k terms of Equation (4), which results in the loss of all remaining information. Although the value of k can be increased to calculate the diffusion of as much information as possible, a larger k means a longer computation time for the Laplace spectral transform and a larger amount of data storage. Based on the above analysis, when k is at a moderate value, high-frequency information can be utilized in the computation process of DiffusionNet, thereby improving the performance of DiffusionNet. By adding an MLP specifically for modeling high-frequency information, the performance of DiffusionNet on the cortical segmentation task has been significantly improved. At the same time, the model performance will not decrease significantly due to changes in the topological connections of the mesh.
[0010] Preferably, to facilitate the model training process, a BatchNormalization layer is added between the layers of the original MLP model.
[0011] Preferably, in step S1, the original DiffusionNet basic module DiffusionBlock consists of three parts: a spatial diffusion module for information transmission, a spatial gradient features module for capturing anisotropic features, and an MLP multilayer perceptron for modeling point-by-point features.
[0012] The beneficial effects of this invention are: This invention does not rely on spherical mapping and can perform segmentation tasks directly on the original cortical surface. This invention also makes additional use of high-frequency information, which significantly improves the performance of DiffusionNet in cortical segmentation tasks. The segmentation model of this invention is insensitive to the sampling method of the cortical surface mesh and will not cause performance degradation due to changes in the topological connections of the mesh. Attached Figure Description
[0013] Figure 1 This is a flowchart of a prior art cortical segmentation method that relies on spherical mapping, which is a method for constructing a full-band DiffusionNet deep learning model for cortical segmentation. Figure 2 This is a flowchart of existing segmentation methods on the original cortex; Figure 3 These are the original DiffusionBlock structure diagram a and the improved DiffusionBlock structure diagram b of the present invention. Detailed Implementation
[0014] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0015] refer to Figure 3 This paper presents a method for constructing a full-band DiffusionNet deep learning model for cerebral cortex segmentation. First, DiffusionNet, a deep learning model for 3D surfaces, is used as the backbone network. Figure 3 Figure a shows the basic module of the original DiffusionNet, DiffusionBlock, which consists of three parts: a spatial diffusion module for information transmission; a spatial gradient module for capturing anisotropy; and an MLP (Multilayer Perceptron) for modeling point-by-point features. DiffusionNet only needs the coordinates of each vertex of the mesh as input to extract features and use them for segmentation tasks. Moreover, it learns the intrinsic and essential features of the mesh and is insensitive to the sampling method of the mesh. Specifically, DiffusionNet has consistent performance when dealing with mesh variants of the same object with different resolutions and densities.
[0016] Then, regarding the specific calculation method of spatial diffusion, the efficient spectral acceleration method is chosen, which utilizes the Laplace spectral transform.
[0017] For a closed surface S, it satisfies the following equation (1): (1) In equation (1), It is a Laplace-Beltrami differentiable operator, and λ is The eigenvalues of λ are given by the solution Φ, which is the eigenvector corresponding to λ. The eigenvector Φ can be used as the basis for the Laplace spectral transform. According to the following equation (2), the eigenvalue f defined on S is mapped to the Laplace frequency domain to obtain the corresponding coefficient c: (2) At this point, the spectral acceleration method will use the eigenvalue λ to scale the coefficient c with respect to time t, thereby approximating the diffusion with respect to time t, as shown in equation (3): (3) Finally, according to the following equation (4), we reconstruct f using c and λ: (4) However, S can have an infinite number of eigenvalues and eigenvectors, and the spectral acceleration method can only calculate the diffusion for the first k terms of a finite number of equations (4), resulting in the loss of all remaining information. Although the value of k can be increased to calculate the diffusion for as much information as possible, a larger k means a longer computation time for the Laplace spectral transform and a larger amount of data storage, as shown in Table 1 below:
[0018] Table 1. Computation time and data storage requirements for the Laplace spectral transform for different values of k. Based on the above analysis, the goal is to improve the performance of DiffusionNet by utilizing high-frequency information during its computation when k is at a suitable value. First, the high-frequency information is separated according to equation (4), i.e., by subtracting the sum of the first k terms on the right side from f on the left side of equation (4). Then, according to... Figure 3As shown in b, an MLP is added to compute the high-frequency information, while the low-frequency information corresponding to the sum of the first k terms on the right side of equation (4) is computed in the spatial diffusion module as usual. In addition, to facilitate the training process of the model, a Batch Normalization layer is added between the layers of the MLP in the original model.
[0019] By adding an MLP specifically designed to model high-frequency information, DiffusionNet significantly improved its performance on cortical segmentation tasks. The relevant experimental results are shown in Table 2.
[0020] Table 2. Dice coefficients (a metric for segmentation accuracy) of DiffusionNet before and after improvement. Meanwhile, the model performance does not significantly decrease due to changes in the mesh topology. The relevant experimental results are shown in Table 3.
[0021] Table 3. Dice coefficient results of the improved DiffusionNet on data of the same individual at different resolutions. (A / B refers to training on data at resolution A and testing on data at resolution B.) Therefore, this implementation does not rely on spherical mapping and can perform segmentation tasks directly on the original cortical surface. This implementation also utilizes high-frequency information, significantly improving the performance of DiffusionNet on cortical segmentation tasks. The segmentation model in this implementation is insensitive to the sampling method of the cortical surface mesh and will not experience performance degradation due to changes in the mesh's topological connectivity.
[0022] Abbreviations used in this embodiment: MLP stands for Multilayer Perceptron. CNN stands for Convolutional Neural Networks. MRI stands for Magnetic Resonance Imaging. In summary, this invention does not rely on spherical mapping and can perform segmentation tasks directly on the original cortical surface. Furthermore, this invention utilizes high-frequency information, significantly improving the performance of DiffusionNet on cortical segmentation tasks. The segmentation model of this invention is insensitive to the sampling method of the cortical surface mesh and will not experience performance degradation due to changes in the mesh's topological connections. Therefore, this invention has broad application prospects.
[0023] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.
Claims
1. A method for constructing a full-band DiffusionNet deep learning model for cerebral cortex segmentation, characterized in that, Includes the following steps: Step S1: Use the DiffusionNet deep learning model for 3D surfaces as the backbone network; Step S2: DiffusionNet takes the coordinates of each vertex of the mesh as input, extracts features, and uses them for the segmentation task; Step S3: Specifically, regarding the calculation method of the spatial diffusion module, the efficient spectral acceleration method using Laplace spectral transform is selected; for a closed surface S, equation (1) is satisfied: （1） In equation (1), It is a Laplace-Beltrami differentiable operator, and λ is The eigenvalues of λ are such that the solution Φ is the eigenvector corresponding to λ. The eigenvector Φ serves as the basis for the Laplace spectral transform. According to equation (2), the eigenvalues f defined on S are mapped to the Laplace frequency domain to obtain the corresponding coefficients c. （2） The spectral acceleration method uses the eigenvalue λ to scale the coefficient c with respect to time t, thereby approximating the diffusion with respect to time t, as shown in equation (3): （3） According to equation (4), the feature f is reconstructed using c and λ: （4）； In step S3, high-frequency information is separated according to equation (4), that is, high-frequency information is obtained by subtracting the sum of the first k terms on the right side from f on the left side of equation (4); a parallel MLP is added to the spatial diffusion module, and the high-frequency information is handed over to the MLP for calculation, while the low-frequency information corresponding to the sum of the first k terms on the right side of equation (4) is calculated in the spatial diffusion module as usual. To facilitate the training process of the model, a Batch Normalization layer was added between the layers of the original MLP model.
2. The method for constructing a full-band DiffusionNet deep learning model for cerebral cortex segmentation according to claim 1, characterized in that, In step S1, the original DiffusionNet basic module DiffusionBlock consists of three parts: a spatial diffusion module for information transmission, a spatial gradient features module for capturing anisotropic features, and an MLP multilayer perceptron for modeling point-by-point features.