Magnetic Resonance Imaging-Based Classification Method, Device, and Medium for Cerebral Small Vessel Lesion Images

Abstract

The present invention relates to a method, device and medium for classifying brain small vessel disease images based on magnetic resonance imaging. The method first constructs and trains an image classification model based on ensemble learning, then obtains the brain MRI image to be classified, and applies the image classification model to obtain the lesion category corresponding to the brain MRI image to be classified. Specifically, training the image classification model based on ensemble learning includes the following steps: S101, obtaining a brain MRI data set, preprocessing the images in the brain MRI data set to obtain preprocessed images; S102, performing computational analysis on the preprocessed images to obtain corresponding multiple functional metrics, and screening a number of metric features for classification based on the multiple functional metrics; S103, training the image classification model based on the metric features. Compared with the prior art, the present invention has the advantages of high accuracy and easy operation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and relates to an image automatic classification method, and in particular to a method, device and medium for classifying small cerebral vessel disease images based on magnetic resonance imaging. Background Art

[0002] Small cerebral vessel disease (CSVD) is a clinical, pathological and imaging syndrome that can affect small blood vessels such as small arteries, small veins, arterioles, venules and capillaries in the brain due to various causes and pathogenesis, and is the main cause of cognitive impairment and dementia such as vascular dementia, AD, etc. The earliest and most obvious manifestation of CSVD on MRI images is white matter hyperintensities (WMH), which are very common in the brain MRI of the elderly. However, the potential pathological mechanisms of WMH are very diverse, such as incomplete subcortical infarcts, gliosis and axonal loss or mild demyelination of the white matter adjacent to the perivascular space, etc., and its clinical impact is also very complex, including cognitive impairment, increased risk of stroke recurrence, dementia and death.

[0003] CSVD has an insidious onset, progressive development and poor prognosis. At present, there is still a lack of mature and targeted treatment strategies. Therefore, it is urgent to correctly understand the mechanism of action and clinical manifestations of CSVD in the aging brain. At present, most of the imaging intelligent applications based on CSVD are dedicated to the segmentation and localization of lesions, while CSVD is a whole-brain disease, not a focal lesion. The specific clinical manifestations of CSVD are the result of the interconnection between neurons, and then the formation of complex brain networks and their interaction. Existing methods are difficult to classify CSVD lesion images conveniently and accurately. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects existing in the prior art and provide a method, device and medium for classifying small cerebral vessel disease images based on magnetic resonance imaging with high accuracy and easy operation.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for classifying small cerebral vessel disease images based on magnetic resonance imaging, the method first constructs and trains an image classification model based on ensemble learning, and then obtains a brain MRI image to be classified, and uses the image classification model to obtain the lesion category corresponding to the brain MRI image to be classified;

[0007] Training the image classification model based on ensemble learning specifically includes the following steps:

[0008] S101. Obtain a brain MRI data set, preprocess the images in the brain MRI data set, and obtain preprocessed images;

[0009] S102. Perform computational analysis on the preprocessed image to obtain a plurality of corresponding functional metrics, and screen a number of metric features for classification based on the plurality of functional metrics;

[0010] S103. Train the image classification model based on the metric features.

[0011] Further, in step S101, the preprocessing performed on the images of the brain MRI dataset includes one or more of temporal slice correction, head motion correction, spatial normalization, smoothing, band-pass filtering, detrending, and regression of nuisance parameters.

[0012] Further, in step S102, the functional metrics include ReHo map, ALFF map, FC pattern, and brain functional network, and the brain functional network has local network attributes and global network attributes.

[0013] Further, two-sample T-test is used for feature selection to obtain representative brain voxel clusters from the ReHo map and ALFF map as metric features for classification.

[0014] Further, the L0minCV method is used for feature selection of the FC pattern and the local network attributes in the brain functional network.

[0015] Further, the brain functional network is a weighted brain network with sparsity, and the construction process includes:

[0016] Set the connectivity degree between every two brain network nodes u and v, use the correlation as the weight of the edge between the two nodes, and only take the connections of the first s% of the nodes when all the weight values are sorted in descending order to obtain the weighted brain network with sparsity.

[0017] Further, the local network attributes include betweenness centrality, degree centrality, clustering coefficient, local efficiency, and node efficiency;

[0018] The global network attributes include assortativity, hierarchical degree, local brain network efficiency, global brain network efficiency, and small-world network attributes.

[0019] Further, the image classification model based on ensemble learning includes an SVM model for each metric feature, and the final classification result is obtained by weighting the preliminary classification results of each SVM model.

[0020] The present invention also provides an electronic device, including:

[0021] One or more processors;

[0022] A memory; and

[0023] One or more programs stored in a memory, the one or more programs including instructions for performing the image classification method as described above.

[0024] The present invention also provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, the one or more programs including instructions for performing the image classification method as described above.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention automatically classifies the input image to be classified by using an ensemble learning method, and has high classification accuracy.

[0027] 2. The present invention analyzes the functional metrics corresponding to the images, and uses representative features as the metric features for classification, effectively improving the classification accuracy.

[0028] 3. The present invention realizes the automatic classification of brain images by a computer program, is easy to promote, and does not require doctors to perform cumbersome visual assessments on the severity of WHM. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a schematic flowchart of the present invention;

[0030] Figure 2 is a schematic diagram of an image classification model based on ensemble learning of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0032] Embodiment 1

[0033] This embodiment provides a method for classifying small cerebral vessel disease images based on magnetic resonance imaging. The method first constructs and trains an image classification model based on ensemble learning, then obtains the brain MRI image to be classified, and applies the image classification model to obtain the lesion category corresponding to the brain MRI image to be classified. Training the image classification model based on ensemble learning specifically includes the following steps:

[0034] S101. Obtain a brain MRI data set, preprocess the images in the brain MRI data set, and obtain preprocessed images.

[0035] The images of the obtained brain MRI dataset include T1 structural images and rs-fMRI images. The preprocessing performed on the images of the brain MRI dataset includes one or more of temporal slice correction, head motion correction, spatial normalization, smoothing, band-pass filtering, removing linear drift, and regressing nuisance parameters.

[0036] S102. Perform computational analysis on the preprocessed images to obtain corresponding multiple functional metrics, and screen a number of metric features for classification based on the multiple functional metrics.

[0037] In this embodiment, the functional metrics include ReHo maps, ALFF maps, FC patterns, and brain functional networks. The specific methods for obtaining each functional metric are as follows:

[0038] Steps for constructing the ALFF map: For the preprocessed rs-fMRI images, first use a 0.01 - 0.08 Hz band-pass filter to retain the signals in the low-frequency band, then perform Fourier transform on all signal intensities to obtain the power spectrum, take the square root of it to obtain the amplitude of the BOLD signal, and finally sum and average the amplitude values at all frequency points. This average value is the final ALFF value, as shown in formula (1):

[0039]

[0040] In formula (1), N is the total number of voxels, a k , b k are the corresponding coefficients at different frequencies.

[0041] Steps for constructing the ReHo map: For the preprocessed rs-fMRI images, calculate the Kendall's coefficient of concordance for each brain voxel to evaluate the consistency of its spontaneous neural activity in the local region. The calculation formula is as follows:

[0042]

[0043] In formula (2), R i is the sum of ranks at the i-th time point; the number of time points is represented by n; K is the time series of the calculated voxel and its nearest neighbor voxels (here, K is 27, a given voxel plus its 26 neighbors); represents the average value of R i , where W is the ReHo value of the given voxel, and its range is from 0 to 1.

[0044] FC mode construction steps: Use the Brainnetome Atlas to divide each processed brain image into 246 brain regions, that is, 123 nodes in each of the left and right hemispheres, and extract the average time series signal of each brain region; calculate the correlation coefficient between the time series signals of any two brain region nodes, and use the Pearson similarity measure to obtain the whole-brain functional connectivity matrix Maps. The element (u, v) represents the functional connectivity strength from the u-th brain node to the v-th brain node. Each row of this matrix Maps describes the connection pattern of this brain region with all brain nodes in the resting state. The calculation formula of the Pearson similarity measure is:

[0045]

[0046] In formula (3), u and v represent any two nodes, I(·, t) represents the voxel value at time t, represents the average signal value of the node, T represents the length of the time series, and S I (·) represents the standard deviation of the node signal value.

[0047] Apply the Fisher Z-transform to Maps to obtain the FC mode of each image, that is, the FC matrix of size 246×246. Since the FC matrix is symmetric about the main diagonal, therefore, take the lower triangular matrix and stretch it into a row. Finally, the FC of each image is a one-dimensional vector of 1×30135.

[0048] Brain functional network construction steps: Set the connectivity degree between every two brain network nodes u and v: Use the correlation as the weight of the edge between the two nodes, and only take the top s% of the edges with larger weights when all weight values are sorted in descending order to obtain a weighted brain network with a certain sparsity.

[0049] Analyze the above brain functional network through graph theory methods to extract brain network attributes, including local network attributes and global network attributes. Local network attributes include betweenness centrality, degree centrality, clustering coefficient, local efficiency, and node efficiency; global network attributes include assortativity, hierarchical degree, local brain network efficiency, global brain network efficiency, and small-world network attributes.

[0050] Among them, when constructing a weighted brain network with sparsity, the value of the sparsity s% is uncertain, with large individual differences, and the selection of a specific sparsity threshold is cumbersome and error-prone. At this time, calculate the area under the curve of each node attribute at different sparsities, and combine the attribute values at different sparsity threshold levels. The combined result is used as the brain network attribute that can truly and accurately describe the original image.

[0051] With respect to the above-mentioned functional metrics such as the ReHo map, ALFF map, FC pattern, brain functional network, etc., this embodiment uses different methods to analyze and determine the metric features that are ultimately used for classification.

[0052] The steps for feature screening of ALFF graph and ReHo graph are as follows:

[0053] The two-sample T test was used to perform inter-group statistical analysis on the ALFF and ReHo maps of the samples. The effects of age, gender, and education level were regressed out, and the ALFF and ReHo maps were corrected by AlphaSim. The statistical threshold was p<0.01, and representative brain voxel clusters in the ALFF and ReHo maps were obtained, respectively.

[0054] The steps for screening FC patterns and local brain functional network features are as follows: use the L0minCV method to perform feature selection on FC patterns and local network attributes respectively.

[0055] The L0minCV method specifically includes:

[0056] When using the L0min feature selection method, it is important to determine the number of features r that need to be selected in the end, which directly affects the combination of the optimal feature subset. In order to balance the accuracy and robustness of subsequent experimental results, a stratified cross-validation L0min method, namely L0minCV, is proposed on this basis to find the optimal feature subset that can enable the classification model (preliminarily using linear kernel SVM) to obtain the highest classification accuracy. The specific steps are as follows:

[0057] (1) Using stratified five-fold cross validation, the data to be used for feature selection algorithm is re-divided into training and test data sets;

[0058] (2) All features of the training data are regarded as initial features with a dimension of M. According to the feature ranking obtained by the L0min algorithm, the last D features are discarded to form a new feature subset with a dimension of M-D.

[0059] (3) Extract the features of the test set by index. The feature dimension of the updated test set is also M-D. Use the SVM algorithm to calculate its classification accuracy on the test set.

[0060] (4) The second time, the last D features are discarded to obtain M-2×D dimensional features and the new classification accuracy is calculated;

[0061] (5) When the nth feature is discarded, M-n×D dimensional features are obtained and the classification accuracy is calculated; the feature dimension cannot be reduced any further and the discarding is stopped;

[0062] (6) The feature dimension with the highest classification accuracy and the smallest number of features is determined as the final feature selection number r.

[0063] S103. Train the image classification model based on the metric features.

[0064] In this method, the image classification model based on ensemble learning uses SVM as the base classifier, including SVM models for each metric feature, and the final classification result is obtained by weighting the preliminary classification results of each SVM model.

[0065] In this embodiment, 9 different feature subsets based on rs-fMRI data are obtained based on the selected metric features, including 2 feature subsets of ALFF and ReHo obtained by two-sample T-test; 6 feature subsets of FC patterns and local network attributes obtained by L0minCV feature selection, and a global network attribute feature subset without feature selection. The final output of the integrated SVM is jointly determined by multiple SVM classifiers, and the final automatic classification result is obtained by using a weighted voting strategy.

[0066] If the above method is implemented in the form of software functional units and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0067] Embodiment 2

[0068] This embodiment provides a brain small vessel disease image classification system based on magnetic resonance imaging, including a model construction module and a classification module. The model construction module includes:

[0069] A dataset acquisition unit for acquiring a brain MRI dataset, preprocessing the images of the brain MRI dataset, and obtaining preprocessed images;

[0070] A functional metric construction unit for performing computational analysis on the preprocessed images to obtain corresponding multiple functional metrics;

[0071] A screening unit for screening and obtaining several metric features for classification based on the multiple functional metrics;

[0072] A training unit for training an image classification model based on ensemble learning according to the selected metric features.

[0073] The rest is the same as in Embodiment 1.

[0074] Embodiment 3

[0075] This embodiment provides an electronic device, including one or more processors, a memory, and one or more programs stored in the memory. The one or more programs include instructions for executing the image classification method as described in Embodiment 1.

[0076] This electronic device can be applied to the classification of CSVD lesion images. Different severities of WMH are used as different lesion categories, divided into mild and moderate-to-severe. The use of this electronic device can effectively improve the diagnostic efficiency and early screening efficiency of CSVD, further assist doctors in diagnosing and early intervening in CSVD patients, prevent WMH-related brain damage in the elderly, and give them corresponding health tips.

[0077] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A method for classifying images of cerebral small vessel lesions based on magnetic resonance imaging, characterized in that, The method first constructs and trains an image classification model based on ensemble learning, then obtains the brain MRI image to be classified, and applies the image classification model to obtain the lesion category corresponding to the brain MRI image to be classified; Training the image classification model based on ensemble learning specifically includes the following steps: S101. Obtain a brain MRI dataset, preprocess the images in the brain MRI dataset to obtain preprocessed images. The images in the obtained brain MRI dataset include T1 structural images and rs-fMRI images; S102. Perform computational analysis on the preprocessed images to obtain corresponding multiple functional metrics, and screen a number of metric features for classification based on the multiple functional metrics; S103. Train the image classification model based on the metric features; In step S102, the functional metrics include ReHo map, ALFF map, FC pattern, and brain functional network. The brain functional network has local network attributes and global network attributes; The construction steps of the ALFF map include: for the preprocessed rs-fMRI image, first use a 0.01 - 0.08 Hz band-pass filter to retain the signals in the low-frequency band, then perform Fourier transform on all signal intensities to obtain the power spectrum, take the square root of it to obtain the amplitude of the BOLD signal, and finally sum and average the amplitude values at all frequency points. This average value is the final ALFF value, and then construct the ALFF map; The construction steps of the ReHo map include: for the preprocessed rs-fMRI image, calculate the Kendall's coefficient of concordance of each brain voxel to evaluate the consistency of its spontaneous neural activity in the local area, and then construct the ReHo map; The brain functional network is a weighted brain network with sparsity. The construction process includes: Set the connectivity degree between every two brain network nodes u and v, use the correlation as the weight of the edge between the two nodes, and only take the connections of the first s% of the nodes when all the weight values are sorted in descending order to obtain the weighted brain network with sparsity; The local network attributes include betweenness centrality, degree centrality, clustering coefficient, local efficiency, and node efficiency; The global network attributes include assortativity, hierarchical degree, local brain network efficiency, global brain network efficiency, and small-world network attributes.

2. The method for classifying cerebral small vessel disease images based on magnetic resonance imaging according to claim 1, wherein In step S101, the preprocessing performed on the images in the brain MRI dataset includes one or more of temporal slice correction, head motion correction, spatial normalization, smoothing, band-pass filtering, detrending, and regression of nuisance parameters.

3. The method for classifying cerebral small vessel disease images based on magnetic resonance imaging according to claim 1, characterized in that Adopt a two-sample T-test for feature selection to obtain representative brain voxel clusters from the ReHo map and the ALFF map as the metric features for classification.

4. The method for classifying cerebral small vessel disease images based on magnetic resonance imaging according to claim 1, wherein Use the L0minCV method for feature selection of the FC pattern and the local network attributes in the brain functional network.

5. The method for classifying cerebral small vessel disease images based on magnetic resonance imaging according to claim 1, wherein The image classification model based on ensemble learning includes an SVM model for each metric feature, and the final classification result is obtained by weighting the preliminary classification results of each SVM model.

6. An electronic device, characterized in that, Including: One or more processors; A memory; And One or more programs stored in a memory, the one or more programs including instructions for performing the image classification method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, One or more programs for execution by one or more processors of an electronic device, the one or more programs including instructions for performing the image classification method according to any one of claims 1-5.

Images