A brain magnetic resonance image classification method, system and electronic device
The feature extraction and classification model built using the Fast_RNRR algorithm and convolutional neural network solves the problem of redundant feature extraction in brain MRI images and achieves higher automatic classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUBEI NEW RUIDE DEFENSE TECH CO LTD
- Filing Date
- 2023-05-16
- Publication Date
- 2026-06-02
AI Technical Summary
Existing brain MRI image feature extraction algorithms are cumbersome and lack universality, resulting in redundant feature extraction in whole brain scan images and affecting the accuracy of automatic classification results.
The Fast_RNRR algorithm is used for point cloud registration to determine feature blocks. Then, a feature extraction model and classification model based on convolutional neural networks are used to improve the rationality of feature extraction through topological feature maps, thereby realizing the automatic classification of brain MRI images.
It improves the accuracy of automatic classification of brain MRI images, overcomes the problem of redundant feature extraction in traditional methods, and enhances the reliability of classification results.
Smart Images
Figure CN116524277B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image classification technology, and in particular to a method, system, and electronic device for classifying brain magnetic resonance imaging (MRI) images. Background Technology
[0002] Automatic classification technology for medical images, especially brain MRI images, has promoted the development of medical devices and improved the automation level of the medical industry. Feature extraction algorithms for brain MRI images are crucial for automatic classification. Current feature extraction algorithms mainly include: pixel value change-based feature detection algorithms (such as edge detection and peak detection algorithms), human-computer interaction-based feature detection algorithms (i.e., manually selecting feature regions), and human visual feature detection algorithms (such as color detection and information entropy detection algorithms). These feature extraction algorithms are not only cumbersome and lack universality, but they also directly extract features from whole-brain MRI scans. However, key information in medical images only exists in local areas. Whole-brain scan feature extraction yields a large number of redundant features, thus affecting the accuracy of automatic classification results and making it difficult to apply to the design of medical devices. Summary of the Invention
[0003] The purpose of this invention is to provide a method, system, and electronic device for classifying brain MRI images, which can improve the accuracy of automatic classification of brain MRI images.
[0004] To achieve the above objectives, the present invention provides the following solution:
[0005] A method for classifying brain magnetic resonance imaging (MRI) images, comprising:
[0006] Obtain MRI images of the brain to be classified;
[0007] Using the Fast_RNRR algorithm, the brain MRI image to be classified is registered with a standard brain MRI image using point cloud, and multiple feature blocks are determined based on the point cloud registration results.
[0008] Multiple feature blocks are input into the feature extraction model to obtain multiple feature vectors;
[0009] Based on multiple feature vectors, a topological feature map of the brain MRI image to be classified is determined;
[0010] The topological feature map is input into the brain MRI image classification model to obtain the classification result of the brain MRI image to be classified; the feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model using labeled historical brain MRI images; both the initial topological feature map generation model and the initial brain MRI image classification model are based on convolutional neural networks.
[0011] Optionally, after acquiring the brain MRI image to be classified, the method further includes:
[0012] The brain MRI images to be classified are preprocessed; the preprocessing includes: head movement correction, skull dissection, B1 bias field correction, gray and white matter segmentation, reconstruction of gray and white matter boundary surfaces and reconstruction of gray and white matter pia mater surfaces, as well as marking and segmenting cortical regions and subcortical brain structures; the preprocessing tool is FreeSurfer.
[0013] Optionally, the Fast_RNRR algorithm is used to perform point cloud registration between the brain MRI image to be classified and a standard brain MRI image, and multiple feature blocks are determined based on the point cloud registration results, including:
[0014] Using the Fast_RNRR algorithm, point cloud registration was performed on classified brain MRI images and standard brain MRI images, and points with a point cloud registration difference value greater than a preset difference value were identified as feature points.
[0015] Construct a cube-shaped cut-off region centered on the feature point and with a preset length as the side length;
[0016] By using the cube-shaped regions corresponding to multiple feature points, feature blocks are extracted from the brain MRI image to be classified, resulting in multiple feature blocks.
[0017] Optionally, the initial feature extraction model includes a first processing layer, a second processing layer, a third processing layer, and a fourth processing layer connected in sequence;
[0018] The output of the first processing layer is densely connected to the output of the second processing layer;
[0019] The output of the second processing layer is densely connected to the output of the third processing layer;
[0020] The output of the first processing layer is residually connected to the output of the third processing layer;
[0021] The first processing layer includes a first convolutional module, a second convolutional module, and a first average pooling layer connected in sequence;
[0022] The second processing layer includes a third convolutional module and a second average pooling layer connected in sequence.
[0023] The third processing layer includes a fourth convolutional module and a third average pooling layer connected in sequence;
[0024] The fourth processing layer includes a fifth convolutional module and a GAP layer connected in sequence.
[0025] Optionally, determining the topological feature map of the brain MRI image to be classified based on the plurality of feature vectors includes:
[0026] Determine the Euclidean distance between any two of the feature points;
[0027] Determine the Hausdorff distance between the cortical regions containing any two of the aforementioned feature points;
[0028] Using multiple feature vectors as nodes, multiple Euclidean distances as first side lengths, and multiple Hausdorff distances as second side lengths, a topological feature map of the brain MRI image to be classified is constructed.
[0029] Optionally, the brain MRI image classification model includes a first convolutional layer, a first pooling layer, a second convolutional layer, and a second pooling layer connected in sequence.
[0030] Optionally, before acquiring the brain MRI image to be classified, the method further includes:
[0031] Acquire standard brain MRI images and multiple historical brain MRI images;
[0032] Multiple historical brain MRI images were categorized and labeled to obtain labeled historical brain MRI images; the labeling results are quasi-standard brain MRI images and non-standard brain MRI images;
[0033] Using the Fast_RNRR algorithm, multiple historical brain MRI images are registered with standard brain MRI images using point cloud registration, and multiple feature block groups are determined based on the point cloud registration results; each feature block group includes multiple feature blocks determined by registering the corresponding historical brain MRI images with standard brain MRI images using point cloud registration.
[0034] The initial topological feature map generation model is determined to be the first model at the 0th iteration;
[0035] The initial brain MRI image classification model was determined to be the second model from the 0th iteration;
[0036] Let the iteration number i = 1;
[0037] Multiple feature block groups are input into the first model in the (i-1)th iteration to obtain multiple feature vector groups; each feature vector group corresponds one-to-one with the feature block group.
[0038] Based on multiple feature vector groups, the topological feature map corresponding to each historical brain MRI image is determined;
[0039] The topological feature map corresponding to each brain MRI historical image is input into the second model in the (i-1)th iteration to obtain the classification results of multiple brain MRI historical images;
[0040] Adjust the parameters of the first model in the (i-1)th iteration to obtain the first model in the i-th iteration;
[0041] Adjust the parameters of the second model in the (i-1)th iteration to obtain the second model in the i-th iteration;
[0042] Increment the iteration number i by 1, and return to the step "input multiple feature block groups into the feature extraction model at the (i-1)th iteration to obtain multiple feature vector groups" until the loss function is less than the preset value of the loss function and the iteration number i reaches the preset value of the iteration number. Determine the first model at the i-th iteration as the feature extraction model; determine the second model at the i-th iteration as the brain MRI image classification model.
[0043] A method for classifying brain magnetic resonance imaging (MRI) images, comprising:
[0044] The brain MRI image acquisition module is used to acquire brain MRI images to be classified.
[0045] The feature block determination module is used to perform point cloud registration between the brain MRI image to be classified and the standard brain MRI image using the Fast_RNRR algorithm, and determine multiple feature blocks based on the point cloud registration results.
[0046] The feature vector determination module is used to input multiple feature blocks into the feature extraction model to obtain multiple feature vectors;
[0047] The topological feature map determination module is used to determine the topological feature map of the brain MRI image to be classified based on multiple feature vectors.
[0048] The classification module is used to input the topological feature map into the brain MRI image classification model to obtain the classification result of the brain MRI image to be classified; the feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model using labeled historical brain MRI images; both the initial topological feature map generation model and the initial brain MRI image classification model are based on convolutional neural networks.
[0049] An electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to enable the electronic device to perform the brain magnetic resonance imaging classification method.
[0050] Optionally, the memory is a readable storage medium.
[0051] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0052] This invention provides a method, system, and electronic device for classifying brain MRI images. The method involves acquiring a brain MRI image to be classified; using the Fast_RNRR algorithm, registering the brain MRI image to be classified with a standard brain MRI image using point clouds, and determining multiple feature blocks based on the point cloud registration results; inputting these feature blocks into a feature extraction model to obtain multiple feature vectors; determining the topological feature map of the brain MRI image to be classified based on these feature vectors; and inputting the topological feature map into a brain MRI image classification model to obtain the classification result of the brain MRI image to be classified. The feature extraction model and the brain MRI image classification model are obtained by simultaneously training an initial feature extraction model and an initial brain MRI image classification model using labeled historical brain MRI images. Both the initial topological feature map generation model and the initial brain MRI image classification model are constructed based on convolutional neural networks. This invention improves the rationality of feature extraction by constructing the feature extraction model and the brain MRI image classification model using convolutional neural networks, thereby improving the accuracy of automatic classification of brain MRI images. Attached Figure Description
[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0054] Figure 1This is a flowchart of the brain magnetic resonance imaging classification method in Embodiment 1 of the present invention;
[0055] Figure 2 This is a flowchart of the brain magnetic resonance imaging classification method in Embodiment 2 of the present invention;
[0056] Figure 3 This is a schematic diagram of the initial feature extraction model in Embodiment 2 of the present invention;
[0057] Figure 4 This is a schematic diagram of the model training effect in Embodiment 2 of the present invention. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] The purpose of this invention is to provide a method, system, and electronic device for classifying brain MRI images, which can improve the accuracy of automatic classification of brain MRI images.
[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0061] Example 1
[0062] like Figure 1 As shown, this embodiment provides a method for classifying brain magnetic resonance imaging (MRI) images, including:
[0063] Step 101: Obtain MRI images of the brain to be classified.
[0064] Step 102: Using the Fast_RNRR algorithm, the brain MRI image to be classified is registered with the standard brain MRI image using point cloud, and multiple feature blocks are determined based on the point cloud registration results.
[0065] Step 102 includes:
[0066] Step 1021: Using the Fast_RNRR algorithm, perform point cloud registration on the classified brain MRI images and the standard brain MRI images, and determine the points whose point cloud registration difference value is greater than the preset difference value as feature points.
[0067] Step 1022: Construct a cube-shaped cut-off region with the feature point as the center and the preset length as the side length.
[0068] Step 1023: Using the cube-shaped regions corresponding to multiple feature points, feature blocks are extracted from the MRI images of the brain to be classified, resulting in multiple feature blocks.
[0069] Step 103: Input multiple feature blocks into the feature extraction model to obtain multiple feature vectors.
[0070] Step 104: Determine the topological feature map of the brain MRI image to be classified based on multiple feature vectors.
[0071] Step 104 includes:
[0072] Step 1041: Determine the Euclidean distance between any two feature points.
[0073] Step 1042: Determine the Hausdorff distance between the cortical regions containing any two feature points.
[0074] Step 1043: Construct a topological feature map of the brain MRI image to be classified, using multiple feature vectors as nodes, multiple Euclidean distances as the first side length, and multiple Hausdorff distances as the second side length.
[0075] Step 105: Input the topological feature map into the brain MRI image classification model to obtain the classification result of the brain MRI image to be classified; the feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model with labeled historical brain MRI images; the initial topological feature map generation model and the initial brain MRI image classification model are both built based on convolutional neural networks.
[0076] The initial feature extraction model includes a first processing layer, a second processing layer, a third processing layer, and a fourth processing layer connected in sequence. The outputs of the first processing layer and the second processing layer are densely connected. The outputs of the second processing layer and the third processing layer are densely connected. The outputs of the first processing layer and the third processing layer are residually connected. The first processing layer includes a first convolutional module, a second convolutional module, and a first average pooling layer connected in sequence. The second processing layer includes a third convolutional module and a second average pooling layer connected in sequence. The third processing layer includes a fourth convolutional module and a third average pooling layer connected in sequence. The fourth processing layer includes a fifth convolutional module and a GAP layer connected in sequence.
[0077] Specifically, the brain MRI image classification model includes a first convolutional layer, a first pooling layer, a second convolutional layer, and a second pooling layer connected in sequence.
[0078] Following step 101, the following is also included:
[0079] Step 106: Preprocess the brain MRI images to be classified; preprocessing includes: head movement correction, skull dissection, B1 bias field correction, gray and white matter segmentation, reconstruction of gray and white matter boundary surfaces and reconstruction of gray and white matter pia mater surfaces, as well as labeling segmented cortical regions and subcortical brain structures; the preprocessing tool is FreeSurfer.
[0080] Before step 101, the following is also included:
[0081] Step 107: Obtain standard brain MRI images and multiple historical brain MRI images.
[0082] Step 108: Perform category labeling on multiple historical brain MRI images to obtain labeled historical brain MRI images; the labeling results are standard brain MRI images and non-standard brain MRI images.
[0083] Step 109: Using the Fast_RNRR algorithm, perform point cloud registration on multiple historical brain MRI images and standard brain MRI images, and determine multiple feature block groups based on the point cloud registration results; any feature block group includes multiple feature blocks determined by performing point cloud registration on the corresponding historical brain MRI images and standard brain MRI images.
[0084] Step 1010: Determine the initial topological feature map generation model as the first model at the 0th iteration.
[0085] Step 1011: Determine the initial brain MRI image classification model as the second model at the 0th iteration.
[0086] Step 1012: Let the iteration number i = 1.
[0087] Step 1013: Input multiple feature block groups into the first model in the (i-1)th iteration to obtain multiple feature vector groups; the feature vector groups correspond one-to-one with the feature block groups.
[0088] Step 1014: Determine the topological feature map corresponding to each brain MRI historical image based on multiple feature vector groups.
[0089] Step 1015: Input the topological feature map corresponding to each brain MRI historical image into the second model in the (i-1)th iteration to obtain the classification results of multiple brain MRI historical images.
[0090] Step 1016: Adjust the parameters of the first model in the (i-1)th iteration to obtain the first model in the i-th iteration.
[0091] Step 1017: Adjust the parameters of the second model in the (i-1)th iteration to obtain the second model in the i-th iteration.
[0092] Step 1018: Increment the value of iteration number i by 1, and return to step "Input multiple feature block groups into the feature extraction model at iteration i-1 to obtain multiple feature vector groups" until the loss function is less than the preset value of the loss function and the iteration number i reaches the preset value of the iteration number. Determine the first model at iteration i as the feature extraction model; determine the second model at iteration i as the brain MRI image classification model.
[0093] Example 2
[0094] like Figure 2 This embodiment provides a method for classifying brain MRI images: It provides a brain MRI image of the test subject, which is processed by the image processing software FreeSurfer, including head movement correction, skull dissection, B1 bias field correction, gray and white matter segmentation, reconstruction of the gray and white matter boundary surface and the pia mater surface, and labeling and segmenting cortical regions and subcortical brain structures to generate a T1-enhanced NII format brain image file and an ASEG brain region segmentation label file based on the NII format; the exported brain image file is converted into an NPY matrix file; a PCD format point cloud image of the test subject's brain is created based on the NPY format matrix; and the p... The CD image and a standard brain are registered using the Fast_RNRR algorithm. The 20 feature points with the largest registration variations are extracted, and a 32*32*32 voxel square region is taken as a feature block, centered at these points. The feature blocks are then dimensionality-reduced using a 3D convolutional neural network to obtain feature vectors of dimension 125. The Euclidean distance between the corresponding feature points and the test subject's brain, and the Hausdorff distance between the corresponding ASEG regions, are calculated. These two distances are used as edges in the topological graph structure, and the feature vectors extracted from the feature points are used as nodes in the graph structure. This graph structure is processed using a graph convolutional network to obtain the classification results. Specifically, this embodiment includes the following steps:
[0095] Step 1: Acquire and preprocess MRI images of the subject's brain as needed. The MRI images are 512*512*512 single-channel images, and the preprocessing includes the following steps:
[0096] Step 1.1 MRI images were preprocessed using FreeSurfer, including head movement correction, skull dissection, B1 bias field correction, gray and white matter segmentation, reconstruction of the gray and white matter boundary surface and the pia mater surface, and labeling of segmented cortical regions and subcortical brain structures.
[0097] Step 1.2 uses the Fast_RNRR algorithm to obtain the flexible registration process between the subject's brain and the standard brain, obtaining feature points with large corresponding point errors. Using these feature points as geometric centers, feature blocks are generated, with a feature block size of 32*32*32.
[0098] Step 2 involves using a convolutional neural network to reduce the dimensionality of the feature blocks, such as... Figure 3 As shown, the network has the following characteristics:
[0099] Step 2.1 The network takes a 64*64*64 matrix as input and contains four main layers.
[0100] The first layer processes the input twice using a 5x5x5 convolutional kernel, followed by a 2x2x2 average pooling layer. The second layer uses a 5x5x5 convolutional kernel, and the result is densely concatenated with the input of the second layer, followed by a 2x2x2 max pooling layer. The third layer uses a 3x3x3 convolutional kernel, and the result is densely concatenated with the input of the third layer and residually concatenated with the input of the second layer, followed by a 2x2x2 max pooling layer. The fourth layer uses a 3x3x3 convolutional kernel, followed by a GAP layer, resulting in a feature vector of dimension 125.
[0101] Step 2.2 Based on the feature points obtained from Fast_RNRR, construct the Euclidean distance and Hausdorff distance between each point. Then, the feature vectors together with the obtained Euclidean distance and Hausdorff distance constitute the topological feature graph. The two distance variables selected for the edges of the topological graph structure are selected through ablation experiments.
[0102] Step 3 processes the topological feature map using graph convolution. This graph convolutional network has the following characteristics:
[0103] Step 3.1 The network will use Graph Conv as the convolution kernel and contain two convolutional layers, each followed by a pooling layer.
[0104] Step 3.2: The graph convolutional network outputs a feature vector of size 25*5, which is then connected to a fully connected neural network containing three hidden layers and finally outputs a result vector of size 2*1.
[0105] Step 3.3: The network will output the classification result vector of the topological feature map. In the experiment testing this algorithm, the hardware configuration was an Intel i7-1065G7 quad-core processor with a clock speed of 2.2GHz, 32GB DDR4 3733MHz memory, and an NVIDIA GeForce RTX 2080 with Max-Q Design graphics card. The system used Ubuntu 16.04, and the neural network was built using TensorFlow 1.5. Each model used a 5x cross-validation method. The training samples and test samples were set to 70% and 30% of the original sample library, respectively. The test results are as follows: Figure 4 As shown.
[0106] To demonstrate the performance of this disclosed classifier algorithm, it was tested using data from ABIDE. The test results show that the algorithm can be applied to data from multiple sites, and its training accuracy indicates that the classifier algorithm can provide objective diagnostic guidance. The training time does not include FreeSurfer processing time; the actual processing time will vary depending on computer performance. Specific performance details are shown in Table 1.
[0107] Table 1 Comparison of Network Performance of Datasets
[0108]
[0109]
[0110] Table 1 shows the datasets sourced from the following institutions: University of Pittsburgh School of Medicine Investigators (PITT), Olin Institute of Living at Hartford Hospital (OLIN), Oregon Health and Science University (OHSU), San Diego State University (SDSU), Trinity Center for Health Sciences (TRINITY), University of Michigan (UM), University of Utah School of Medicine (USM), Yale Child Study Center (YALE), Carnegie Mellon University (CMU), KU Leuven, Kennedy Krieger Institute (KKI), NYU Langone Medical Center (NYU), and Stanford University (STANFORD). University, University of California, Los Angeles (UCLA), Ludwig Maximilians University Munich (MAX-MUN), California Institute of Technology (CALTECH), and Social Brain Lab (SBL).
[0111] In summary, this embodiment provides a classifier algorithm that is simple in structure, easy to deploy, and highly universal, overcoming the problems of poor convergence and low accuracy of traditional image classifiers under conditions of multiple residual samples.
[0112] Example 3
[0113] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a classification method for brain magnetic resonance images is provided below, including:
[0114] The module for acquiring brain MRI images to be classified is used to acquire brain MRI images to be classified.
[0115] The feature block determination module is used to perform point cloud registration between the brain MRI image to be classified and the standard brain MRI image using the Fast_RNRR algorithm, and determine multiple feature blocks based on the point cloud registration results.
[0116] The feature vector determination module is used to input multiple feature blocks into the feature extraction model to obtain multiple feature vectors.
[0117] The topology feature map determination module is used to determine the topology feature map of the brain MRI image to be classified based on multiple feature vectors.
[0118] The classification module is used to input the topological feature map into the brain MRI image classification model to obtain the classification result of the brain MRI image to be classified. The feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model using labeled historical brain MRI images. Both the initial topological feature map generation model and the initial brain MRI image classification model are built based on convolutional neural networks.
[0119] Example 4
[0120] This embodiment provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform a brain magnetic resonance imaging (MRI) image classification method as described in Embodiment 1. The memory is a readable storage medium.
[0121] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0122] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for classifying brain magnetic resonance imaging (MRI) images, characterized in that, include: Obtain MRI images of the brain to be classified; Using the Fast_RNRR algorithm, the brain MRI image to be classified is registered with a standard brain MRI image using point cloud, and multiple feature blocks are determined based on the point cloud registration results. Multiple feature blocks are input into the feature extraction model to obtain multiple feature vectors; Based on multiple feature vectors, a topological feature map of the brain MRI image to be classified is determined; The topological feature map is input into a brain MRI image classification model to obtain the classification result of the brain MRI image to be classified. The feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model using labeled historical brain MRI images. Both the initial feature extraction model and the initial brain MRI image classification model are based on convolutional neural networks. The initial feature extraction model includes a first processing layer, a second processing layer, a third processing layer, and a fourth processing layer connected in sequence. The outputs of the first processing layer and the second processing layer are densely connected. The outputs of the second processing layer and the third processing layer are densely connected. The outputs of the first processing layer and the third processing layer are residually connected. The first processing layer includes a first convolutional module, a second convolutional module, and a first average pooling layer connected in sequence. The second processing layer includes a third convolutional module and a second average pooling layer connected in sequence. The third processing layer includes a fourth convolutional module and a third average pooling layer connected in sequence. The fourth processing layer includes a fifth convolutional module and a GAP layer connected in sequence. The Fast_RNRR algorithm is used to perform point cloud registration between the brain MRI image to be classified and a standard brain MRI image, and multiple feature blocks are determined based on the point cloud registration results, including: Using the Fast_RNRR algorithm, point cloud registration was performed on classified brain MRI images and standard brain MRI images, and points with a point cloud registration difference value greater than a preset difference value were identified as feature points. Construct a cube-shaped cut-off region centered on the feature point and with a preset length as the side length; By using the cube-shaped cut-out regions corresponding to multiple feature points, feature blocks are cut out on the brain MRI image to be classified, resulting in multiple feature blocks; The step of determining the topological feature map of the brain MRI image to be classified based on multiple feature vectors includes: Determine the Euclidean distance between any two of the feature points; Determine the Hausdorff distance between the cortical regions containing any two of the aforementioned feature points; Using multiple feature vectors as nodes, multiple Euclidean distances as first side lengths, and multiple Hausdorff distances as second side lengths, a topological feature map of the brain MRI image to be classified is constructed.
2. The method for classifying brain magnetic resonance images according to claim 1, characterized in that, After acquiring the brain MRI image to be classified, the method further includes: The brain MRI images to be classified are preprocessed; the preprocessing includes: head movement correction, skull dissection, B1 bias field correction, gray and white matter segmentation, reconstruction of gray and white matter boundary surfaces and reconstruction of gray and white matter pia mater surfaces, as well as marking and segmenting cortical regions and subcortical brain structures; the preprocessing tool is FreeSurfer.
3. The method for classifying brain magnetic resonance images according to claim 1, characterized in that, The brain MRI image classification model includes a first convolutional layer, a first pooling layer, a second convolutional layer, and a second pooling layer connected in sequence.
4. The method for classifying brain magnetic resonance images according to claim 1, characterized in that, Prior to acquiring the brain MRI image to be classified, the procedure also includes: Acquire standard brain MRI images and multiple historical brain MRI images; Multiple historical brain MRI images were categorized and labeled to obtain labeled historical brain MRI images; the labeling results are quasi-standard brain MRI images and non-standard brain MRI images; Using the Fast_RNRR algorithm, multiple historical brain MRI images are registered with standard brain MRI images using point cloud registration, and multiple feature block groups are determined based on the point cloud registration results; each feature block group includes multiple feature blocks determined by registering the corresponding historical brain MRI images with standard brain MRI images using point cloud registration. The initial feature extraction model is determined to be the first model at the 0th iteration; The initial brain MRI image classification model was determined to be the second model from the 0th iteration; Let the iteration number i = 1; Multiple feature block groups are input into the first model in the (i-1)th iteration to obtain multiple feature vector groups; each feature vector group corresponds one-to-one with the feature block group. Based on multiple feature vector groups, the topological feature map corresponding to each historical brain MRI image is determined; The topological feature map corresponding to each brain MRI historical image is input into the second model in the (i-1)th iteration to obtain the classification results of multiple brain MRI historical images; Adjust the parameters of the first model in the (i-1)th iteration to obtain the first model in the i-th iteration; Adjust the parameters of the second model in the (i-1)th iteration to obtain the second model in the i-th iteration; Increment the value of iteration number i by 1, and return to the step "input multiple feature block groups into the feature extraction model at the (i-1)th iteration to obtain multiple feature vector groups" until the loss function is less than the preset value of the loss function and the iteration number i reaches the preset value of the iteration number. Determine the first model at the i-th iteration as the feature extraction model; determine the second model at the i-th iteration as the brain MRI image classification model.
5. A classification system for brain magnetic resonance imaging (MRI) images, characterized in that, include: The brain MRI image acquisition module is used to acquire brain MRI images to be classified. The feature block determination module is used to perform point cloud registration between the brain MRI image to be classified and the standard brain MRI image using the Fast_RNRR algorithm, and to determine multiple feature blocks based on the point cloud registration result. Specifically, it includes: using the Fast_RNRR algorithm to perform point cloud registration between the brain MRI image to be classified and the standard brain MRI image, and determining points whose point cloud registration difference value is greater than a preset difference value as feature points; constructing a cube-shaped truncated region with the feature points as the center and a preset length as the side length; and using the cube-shaped truncated regions corresponding to multiple feature points to respectively truncate feature blocks on the brain MRI image to be classified, thereby obtaining multiple feature blocks. The feature vector determination module is used to input multiple feature blocks into the feature extraction model to obtain multiple feature vectors; The topological feature map determination module is used to determine the topological feature map of the brain MRI image to be classified based on multiple feature vectors, specifically including: determining the Euclidean distance between any two feature points; determining the Hausdorff distance between the cortical regions where any two feature points are located; and constructing the topological feature map of the brain MRI image to be classified using multiple feature vectors as nodes, multiple Euclidean distances as first side lengths, and multiple Hausdorff distances as second side lengths. A classification module is used to input the topological feature map into a brain MRI image classification model to obtain the classification result of the brain MRI image to be classified. The feature extraction model and the brain MRI image classification model are obtained by simultaneously training the initial feature extraction model and the initial brain MRI image classification model using labeled historical brain MRI images. Both the initial feature extraction model and the initial brain MRI image classification model are based on convolutional neural networks. The initial feature extraction model includes a first processing layer, a second processing layer, a third processing layer, and a fourth processing layer connected in sequence. The output of the first processing layer is densely connected to the output of the second processing layer. The output of the second processing layer is densely connected to the output of the third processing layer. The output of the first processing layer is residually connected to the output of the third processing layer. The first processing layer includes a first convolutional module, a second convolutional module, and a first average pooling layer connected in sequence. The second processing layer includes a third convolutional module and a second average pooling layer connected in sequence. The third processing layer includes a fourth convolutional module and a third average pooling layer connected in sequence. The fourth processing layer includes a fifth convolutional module and a GAP layer connected in sequence.
6. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform a method for classifying brain magnetic resonance images according to any one of claims 1 to 4.
7. An electronic device according to claim 6, characterized in that, The memory is a readable storage medium.