Skull stripping method, device, equipment and medium
Through a method based on spatial prior and position-related convolutional neural networks, the problem of insufficient accuracy of existing skull stripping methods under different modalities and contrasts is solved, and a more precise and stable skull stripping effect is achieved.
Patent Information
- Application Number
- CN202410293808.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-14
- Publication Date
- 2025-09-16
AI Technical Summary
Existing skull stripping methods lack accuracy in different modal images and contrasts, and are not precise enough in brain boundary segmentation, making them difficult to apply to large-scale datasets.
A method based on spatial prior and position-related convolutional neural network is adopted to extract brain tissue intensity information through the subspace model, pre-segmentation is performed in combination with the multi-resolution multi-atlas method, and a multi-resolution position-related network is constructed for training. Finally, the skull stripping results are output through the fusion network.
The accuracy and stability of skull stripping results are improved, accurate skull stripping can be achieved in brain images of different contrasts and modalities, and the complexity of network learning features is reduced.
Smart Images
Figure CN120655906A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of digital image processing, and in particular to a skull stripping method, device, equipment and medium based on spatial prior and position-related convolutional neural networks. Background Art
[0002] Skull stripping is a crucial step in brain image preprocessing. It preserves the brain parenchyma while removing non-brain components such as the skull, subcutaneous fat, skin, and muscle. This process, also known as brain extraction, is crucial for applications such as brain image segmentation, registration, and analysis.
[0003] The most popular skull stripping tool is the Brain Extraction Tool (BET). This tool constructs a dynamic function by assuming the local intensity distribution and smoothness of the brain edges to pull the initial brain surface closer to the brain boundary. Its performance is relatively stable across different modal images and its processing speed is relatively fast, making it widely used. However, achieving accurate skull stripping results requires careful hyperparameter tuning, which limits its applicability in large-scale datasets. Furthermore, BET performs poorly in some less smooth brain regions, such as those near the eyeballs.
[0004] To address this problem, BEAST, a multi-atlas approach, and ANTs, a probabilistic atlas approach, were proposed. Unlike surface-based methods such as BET, BEAST and ANTs assign a label to each voxel to determine whether the voxel is brain parenchyma. BEAST performs multi-atlas fusion based on the similarity between image blocks and atlas image blocks, and improves robustness using a multi-scale approach, while ANTs transforms the brain extraction problem into a tissue segmentation problem, that is, by segmenting all components of the brain parenchyma (gray matter, white matter, cerebrospinal fluid) and combining them to achieve the purpose of brain extraction. By introducing prior information, these two methods have achieved very stable performance, but due to the complex intensity distribution and low contrast of brain boundaries, the boundary segmentation is not accurate enough.
[0005] With the widespread application of deep learning methods in segmentation, a growing number of segmentation networks have achieved promising results in skull stripping applications, such as 3D CNN, auto-contextual U-Net, and O-Net, a combination of GAN and U-Net. These methods demonstrate excellent accuracy in skull stripping applications, but due to limited network generalization, they struggle to achieve good results on test images acquired with varying acquisition parameters or equipment. To address this issue, SynStrip uses existing brain image labels to synthesize training images of varying contrast and modality for training a U-Net-based network, enhancing the network's generalization performance. This allows skull stripping for brain images of any modality and contrast, demonstrating strong generalization. However, due to the relatively simple training data synthesis method, the generated images differ significantly from the real brain images, resulting in inferior performance compared to supervised network-based methods. ASM-CNN, on the other hand, uses the ASM method to obtain an initial brain contour and then refines the corresponding segmentation near the contour using the network to reduce the complexity of the features the network must learn. While achieving good segmentation performance, its high initial contour accuracy requirement makes it difficult to apply to poor-quality images. Summary of the Invention
[0006] The purpose of the present invention is to provide a skull stripping method, device, equipment and medium based on spatial prior and position-related convolutional neural network, using a position-related network guided by subspace prior to learn the intensity distribution of different positions of the brain boundary, and using a multi-resolution method to improve stability and improve the accuracy of the segmentation results.
[0007] The purpose of the present invention can be achieved by the following technical solutions:
[0008] A skull stripping method based on spatial prior and position-dependent convolutional neural network includes the following steps:
[0009] S1, extracting brain tissue spatial prior intensity information from public brain magnetic resonance imaging datasets based on subspace model;
[0010] S2, pre-segmentation of the brain MRI based on the multi-resolution multi-atlas method;
[0011] S3, constructing a multi-resolution position-dependent network, guiding the position-dependent network training based on the subspace prior of step S1 and the pre-segmentation result of step S2, and obtaining the segmentation result;
[0012] S4, constructing a fusion network, taking the subspace prior obtained in step S1 and the segmentation results of the multi-resolution position correlation network obtained in step S3 as input, and fusing them to output the final skull stripping result.
[0013] Said S1 comprises the following steps:
[0014] S11, obtain a public dataset of brain magnetic resonance images and perform initial brain parenchyma segmentation on the images using manual labeling;
[0015] S12, performing singular value decomposition on the initial brain parenchyma segmentation result, and determining the rank of the subspace through the singular value decay curve to construct the subspace;
[0016] S13, subspace projection: The initial brain parenchyma segmentation result is projected onto the constructed subspace using the least squares method to obtain the corresponding spatial prior intensity information.
[0017] The S3 includes the following steps:
[0018] S301, performing singular value decomposition on the pre-segmented images obtained in step S2, constructing a subspace, and projecting the pre-segmented results onto the subspace to obtain a brain parenchyma probability map;
[0019] S302: On the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0020] The S3 includes the following steps:
[0021] S311, performing multi-resolution downsampling based on the pre-segmentation result of step S2 to obtain pre-segmented images of different resolutions;
[0022] S312, performing singular value decomposition on the pre-segmented images of different resolutions, constructing subspaces, and projecting the pre-segmented results onto the subspaces to obtain a brain parenchyma probability map;
[0023] S313, on the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0024] A skull stripping device based on spatial prior and position-dependent convolutional neural network, comprising:
[0025] Subspace prior construction module: extracts brain tissue spatial prior intensity information from public brain MRI datasets based on the subspace model;
[0026] Pre-segmentation module: pre-segment the brain MRI images based on the multi-resolution multi-atlas method;
[0027] Position-dependent segmentation module: Builds a multi-resolution position-dependent network. Based on the subspace prior and pre-segmentation results of the subspace prior construction module, it guides the position-dependent network training to obtain the segmentation results.
[0028] Fusion module: Construct a fusion network, take the subspace prior obtained by the subspace prior construction module and the segmentation results of the multi-resolution position-related network obtained by the position-related segmentation module as input, and fuse them to output the final skull stripping result.
[0029] The subspace prior construction module performs the following steps:
[0030] S11, obtain a public dataset of brain magnetic resonance images and perform initial brain parenchyma segmentation on the images using manual labeling;
[0031] S12, performing singular value decomposition on the initial brain parenchyma segmentation result, and determining the rank of the subspace through the singular value decay curve to construct the subspace;
[0032] S13, subspace projection: The initial brain parenchyma segmentation result is projected onto the constructed subspace using the least squares method to obtain the corresponding spatial prior intensity information.
[0033] The position-dependent segmentation module performs the following steps:
[0034] S301, performing singular value decomposition on the pre-segmented images, constructing a subspace, and projecting the pre-segmented results onto the subspace to obtain a brain parenchyma probability map;
[0035] S302: On the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0036] The position-dependent segmentation module performs the following steps:
[0037] S311, performing multi-resolution downsampling based on the pre-segmentation result to obtain pre-segmented images of different resolutions;
[0038] S312, performing singular value decomposition on the pre-segmented images of different resolutions, constructing subspaces, and projecting the pre-segmented results onto the subspaces to obtain a brain parenchyma probability map;
[0039] S313, on the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0040] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described above when executing the program.
[0041] A computer-readable storage medium stores a computer program, which implements the method described above when executed by a processor.
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] 1. Compared with existing deep learning algorithms, the present invention greatly reduces the complexity of the learning features required by the network by using different networks to learn the intensity distribution of brain parenchyma and non-brain tissue at different locations. Therefore, it can better capture the intensity distribution at the brain boundary, greatly improve the accuracy of boundary definition, and make the skull stripping results more accurate.
[0044] 2. The present invention utilizes prior information extracted from a large amount of data as spatial constraints, combined with a multi-resolution framework, which greatly improves the stability of the network and its robustness to various interferences. At the same time, it can achieve accurate skull stripping for brain images of different contrasts. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 is a flow chart of the method of the present invention;
[0046] Figure 2 A diagram showing the skull stripping results in one embodiment. DETAILED DESCRIPTION
[0047] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0048] This embodiment provides a skull stripping method based on spatial prior and position-dependent convolutional neural network, the purpose of which is to extract brain substance from brain magnetic resonance images and remove non-brain tissues such as skull. Specifically, Figure 1 As shown, the method includes the following steps:
[0049] S1, Extracting spatial prior intensity information of brain tissue from public brain magnetic resonance imaging datasets based on subspace model.
[0050] Specifically, S1 includes the following steps:
[0051] S11, obtain a public dataset of brain magnetic resonance images and perform initial brain parenchyma segmentation on the images using manual labeling based on the BEaST software;
[0052] S12, performing singular value decomposition on the initial brain parenchyma segmentation result, and determining the rank of the subspace through the singular value decay curve to construct the subspace;
[0053] S13, subspace projection: The initial brain parenchyma segmentation result is projected onto the constructed subspace using the least squares method to obtain the corresponding spatial prior intensity information.
[0054] S2, pre-segmentation of the brain magnetic resonance image to be segmented based on the multi-resolution multi-atlas method.
[0055] The pre-segmentation in this step is implemented by BEaST software. The specific implementation method belongs to the conventional technical means used by those skilled in the art and will not be described in detail in this embodiment.
[0056] S3, construct a multi-resolution position-related network, guide the position-related network training based on the subspace prior of step S1 and the pre-segmentation result of step S2, and obtain the segmentation result.
[0057] In one embodiment, S3 includes the following steps:
[0058] S301, performing singular value decomposition on the pre-segmented images obtained in step S2, constructing a subspace, and projecting the pre-segmented results onto the subspace to obtain a brain parenchyma probability map;
[0059] S302: On the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0060] In another preferred embodiment, in order to improve the stability and smoothness of the result, S3 includes the following steps:
[0061] S311, performing multi-resolution downsampling based on the pre-segmentation result of step S2 to obtain pre-segmented images of different resolutions;
[0062] S312, performing singular value decomposition on the pre-segmented images of different resolutions, constructing subspaces, and projecting the pre-segmented results onto the subspaces to obtain a brain parenchyma probability map;
[0063] S313, on the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
[0064] S4, constructing a fusion network, taking the subspace prior obtained in step S1 and the segmentation results of the multi-resolution position correlation network obtained in step S3 as input, and fusing them to output the final skull stripping result.
[0065] like Figure 2 The results of skull stripping performed by the method of the present invention were compared with those of the prior art. It was found that the method of the present invention achieved more accurate results on image datasets with different contrasts than existing commonly used brain extraction tools, demonstrating the accuracy and stability of the present invention.
[0066] The above is an introduction to a method embodiment. The following further illustrates the solution of the present invention through an apparatus embodiment.
[0067] This embodiment also provides a skull stripping device based on a spatial prior and position-dependent convolutional neural network, comprising:
[0068] Subspace prior construction module: extracts brain tissue spatial prior intensity information from public brain MRI datasets based on the subspace model;
[0069] Pre-segmentation module: pre-segment the brain MRI images based on the multi-resolution multi-atlas method;
[0070] Position-dependent segmentation module: Builds a multi-resolution position-dependent network. Based on the subspace prior and pre-segmentation results of the subspace prior construction module, it guides the position-dependent network training to obtain the segmentation results.
[0071] Fusion module: Construct a fusion network, take the subspace prior obtained by the subspace prior construction module and the segmentation results of the multi-resolution position-related network obtained by the position-related segmentation module as input, and fuse them to output the final skull stripping result.
[0072] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0073] In one embodiment, the electronic device includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). Various programs and data required for device operation can also be stored in the RAM. The computing unit, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.
[0074] Many components in an electronic device are connected to the I / O interface, including: input units, such as a keyboard and mouse; output units, such as various types of displays and speakers; storage units, such as magnetic disks and optical disks; and communication units, such as network cards, modems, and wireless communication transceivers. The communication unit allows the device to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunication networks.
[0075] The computing unit can be a variety of general and / or special processing components with processing and computing capabilities. Some examples of computing units include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit performs the various methods and processes described above, such as image processing methods and / or model training methods. For example, in some embodiments, the image processing methods and / or model training methods can be implemented as computer software programs that are tangibly contained in machine-readable media, such as storage units. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of the image processing methods and / or model training methods described above can be performed. Alternatively, in other embodiments, the computing unit can be configured to perform image processing methods and / or model training methods in any other appropriate manner (e.g., by means of firmware).
[0076] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.
[0077] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0078] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.
[0079] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.
Claims
1. A skull stripping method based on spatial prior and position-dependent convolutional neural network, characterized in that: The following steps are involved: S1, extracting brain tissue spatial prior intensity information from public brain magnetic resonance imaging datasets based on subspace model; S2, pre-segmentation of the brain MRI based on the multi-resolution multi-atlas method; S3, constructing a multi-resolution position-dependent network, guiding the position-dependent network training based on the subspace prior of step S1 and the pre-segmentation result of step S2, and obtaining the segmentation result; S4, constructing a fusion network, taking the subspace prior obtained in step S1 and the segmentation results of the multi-resolution position correlation network obtained in step S3 as input, and fusing them to output the final skull stripping result.
2. The skull stripping method based on spatial prior and position-dependent convolutional neural network according to claim 1, characterized in that: Said S1 comprises the following steps: S11, obtain a public dataset of brain magnetic resonance images and perform initial brain parenchyma segmentation on the images using manual labeling; S12, performing singular value decomposition on the initial brain parenchyma segmentation result, and determining the rank of the subspace through the singular value decay curve to construct the subspace; S13, subspace projection: The initial brain parenchyma segmentation result is projected onto the constructed subspace using the least squares method to obtain the corresponding spatial prior intensity information.
3. The skull stripping method based on spatial prior and position-dependent convolutional neural network according to claim 1, characterized in that: The S3 includes the following steps: S301, performing singular value decomposition on the pre-segmented images obtained in step S2, constructing a subspace, and projecting the pre-segmented results onto the subspace to obtain a brain parenchyma probability map; S302: On the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
4. The skull stripping method based on spatial prior and position-dependent convolutional neural network according to claim 1, characterized in that: The S3 includes the following steps: S311, performing multi-resolution downsampling based on the pre-segmentation result of step S2 to obtain pre-segmented images of different resolutions; S312, performing singular value decomposition on the pre-segmented images of different resolutions, constructing subspaces, and projecting the pre-segmented results onto the subspaces to obtain a brain parenchyma probability map; S313, on the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
5. A skull stripping device based on spatial prior and position-dependent convolutional neural network, characterized in that: include: Subspace prior construction module: extracts brain tissue spatial prior intensity information from public brain MRI datasets based on the subspace model; Pre-segmentation module: pre-segment the brain MRI images based on the multi-resolution multi-atlas method; Position-dependent segmentation module: Builds a multi-resolution position-dependent network. Based on the subspace prior and pre-segmentation results of the subspace prior construction module, it guides the position-dependent network training to obtain the segmentation results. Fusion module: Construct a fusion network, take the subspace prior obtained by the subspace prior construction module and the segmentation results of the multi-resolution position-related network obtained by the position-related segmentation module as input, and fuse them to output the final skull stripping result.
6. The skull stripping device based on spatial prior and position-dependent convolutional neural network according to claim 5, characterized in that: The subspace prior construction module performs the following steps: S11, obtain a public dataset of brain magnetic resonance images and perform initial brain parenchyma segmentation on the images using manual labeling; S12, performing singular value decomposition on the initial brain parenchyma segmentation result, and determining the rank of the subspace through the singular value decay curve to construct the subspace; S13, subspace projection: The initial brain parenchyma segmentation result is projected onto the constructed subspace using the least squares method to obtain the corresponding spatial prior intensity information.
7. The skull stripping device based on spatial prior and position-dependent convolutional neural network according to claim 5, characterized in that: The position-dependent segmentation module performs the following steps: S301, performing singular value decomposition on the pre-segmented images, constructing a subspace, and projecting the pre-segmented results onto the subspace to obtain a brain parenchyma probability map; S302: On the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
8. The skull stripping device based on spatial prior and position-dependent convolutional neural network according to claim 5, characterized in that: The position-dependent segmentation module performs the following steps: S311, performing multi-resolution downsampling based on the pre-segmentation result to obtain pre-segmented images of different resolutions; S312, performing singular value decomposition on the pre-segmented images of different resolutions, constructing subspaces, and projecting the pre-segmented results onto the subspaces to obtain a brain parenchyma probability map; S313, on the obtained brain parenchyma probability map, the voxels are divided into different groups based on the brain parenchyma probability corresponding to each voxel, and position-related networks are trained separately to classify the image blocks centered on the voxels belonging to each group, thereby obtaining a brain parenchyma segmentation result.
9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 4 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.