A method and system for classifying perivascular spaces based on magnetic resonance imaging
Through the application of deep learning methods and three -dimensional convolutional neural networks in the MRI image, a gap classification model is established to solve the problem of high difficulty and low efficiency of surrounding vascular gap rating, and the automatic and accurate rating of the gaps around the vascular vascular.
Patent Information
- Application Number
- CN202210305047.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-25
- Publication Date
- 2025-01-24
- Estimated Expiration
- 2042-03-25
AI Technical Summary
In the prior art, the labeling of perivascular gap gaps is difficult, time-consuming, and there are differences between and internal scorers in manual interpretation, resulting in inaccurate judgment results and low efficiency.
A deep learning method based on MRI is used to establish a gap classification model through a three-dimensional convolutional neural network, and T1- and T2-weighted MRI images are used to automatically perform two-class and five-class classification of perivascular gaps.
The automatic and accurate rating of the gap around the vascular vascular reducing the time and errors of manual labeling, and improving the accuracy and efficiency of rating.
Smart Images
Figure CN114723937B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing and deep learning, and in particular to a method and system for classifying perivascular spaces based on nuclear magnetic resonance images. Background Art
[0002] Cerebral small-vessel disease (CSVD) is a common disease that seriously endangers the health of the Chinese people. Due to its hidden onset, it is easily overlooked by patients and even clinicians. Enlarged perivascular spaces (EPVS) are one of the important imaging manifestations of CSVD. They are an indicator of the presence and severity of CSVD, which is related to aging and cognitive decline, and is also a marker of neuroinflammation.
[0003] Expanding the perivascular space rating can objectively reflect the severity of the perivascular space, but it is difficult and time-consuming to mark. The currently widely used perivascular space rating and marking method is mainly for doctors to manually interpret magnetic resonance images. However, the manual interpretation method has differences between and within raters, resulting in inaccurate and inefficient judgment results.
[0004] Therefore, an automatic and accurate perivascular space rating and classification method is urgently needed. Summary of the invention
[0005] The purpose of the present invention is to provide a method and system for classifying perivascular spaces based on nuclear magnetic resonance images, which can improve the accuracy of perivascular space classification.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A method for classifying perivascular spaces based on nuclear magnetic resonance images, the method comprising:
[0008] Acquiring magnetic resonance imaging (MRI) images to be classified; the magnetic resonance imaging (MRI) images to be classified include T1-weighted MRI images and T2-weighted MRI images;
[0009] According to the magnetic resonance image to be classified, based on the gap classification model, determining a first classification result and a second classification result corresponding to the magnetic resonance image to be classified; the first classification result is no perivascular gap or perivascular gap; the second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps;
[0010] The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set; the training sample set includes multiple groups of sample magnetic resonance images and categories corresponding to each group of sample magnetic resonance images; the sample magnetic resonance images include T1-weighted magnetic resonance image samples and T2-weighted magnetic resonance image samples; the categories characterize the levels of perivascular gaps, including binary and quinary categories; the binary categories include no perivascular gaps and perivascular gaps; the quinary categories include no perivascular gaps, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, and more than 40 perivascular gaps.
[0011] Optionally, the method for establishing the gap classification model includes:
[0012] Extracting a region of interest of the T1-weighted nuclear magnetic resonance image sample to obtain a first region of interest;
[0013] Registering the first region of interest and the T1-weighted nuclear magnetic resonance image sample to the T2-weighted nuclear magnetic resonance image sample space to obtain a registered first region of interest and a registered T1-weighted image;
[0014] Extracting the region of interest of the T2-weighted magnetic resonance image according to the registered first region of interest to obtain a second region of interest;
[0015] According to the registered first region of interest, the second region of interest and the corresponding categories, based on the loss function, the three-dimensional convolutional neural network is iteratively trained using back propagation and gradient descent algorithms until the loss function converges to obtain an optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model.
[0016] Optionally, extracting the region of interest of the T1-weighted magnetic resonance image sample to obtain a first region of interest specifically includes:
[0017] Extracting the basal ganglia region and the centrum semiovale region of the T1-weighted MRI image sample;
[0018] Performing morphological binary expansion on the basal ganglia region to obtain an expanded basal ganglia region;
[0019] The expanded basal ganglia region and the semioval center region are respectively normalized to obtain a first region of interest.
[0020] Optionally, the three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively;
[0021] The iterative training of the three-dimensional convolutional neural network according to the categories corresponding to the first ROI and the second ROI after the registration and based on the loss function and using the back propagation and gradient descent algorithms specifically includes:
[0022] Transmitting the registered first region of interest and the second region of interest to a hidden layer through an input layer;
[0023] Extracting features of the first region of interest and the second region of interest after the registration through a hidden layer to obtain initial feature information;
[0024] Performing feature selection on the initial feature information through the global average pooling layer to obtain optimal feature information;
[0025] Classifying the optimal feature information through the first linear layer to obtain a first classification result; the first classification result is no perivascular space or a perivascular space;
[0026] Classifying the optimal feature information through the second linear layer to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces;
[0027] Determine a loss value of a loss function according to the first classification result, the second classification result, the category corresponding to the first region of interest after the registration, and the category corresponding to the second region of interest;
[0028] Based on the loss value of the loss function, back propagation and gradient descent algorithms are used to iteratively train the hidden layer, the global average pooling layer, the first linear layer and the second linear layer.
[0029] Optionally, the loss function is:
[0030] L PVS =0.5*L CE +0.5*L ICC ;
[0031]
[0032] Among them, L PVS is the loss function value, L CEis the cross entropy loss, L ICC is the intra-class correlation coefficient loss, MSR is the mean square of the ratings of different samples by the same rater, MSC is the mean square of the ratings of the same sample by different raters, MSE is the mean square error of the ratings of all samples by all raters, and n is the number of samples.
[0033] To achieve the above object, the present invention also provides the following solution:
[0034] A perivascular space classification system based on nuclear magnetic resonance imaging, the perivascular space classification system based on nuclear magnetic resonance imaging comprising:
[0035] An image acquisition unit, used for acquiring nuclear magnetic resonance images to be classified; the nuclear magnetic resonance images to be classified include T1-weighted nuclear magnetic resonance images and T2-weighted nuclear magnetic resonance images;
[0036] a classification unit connected to the image acquisition unit, and configured to determine, according to the magnetic resonance image to be classified, a first classification result and a second classification result corresponding to the magnetic resonance image to be classified based on a gap classification model; the first classification result is no perivascular gap or perivascular gap; the second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps;
[0037] The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set; the training sample set includes multiple groups of sample magnetic resonance images and categories corresponding to each group of sample magnetic resonance images; the sample magnetic resonance images include T1-weighted magnetic resonance image samples and T2-weighted magnetic resonance image samples; the categories characterize the levels of perivascular gaps, including binary and quinary categories; the binary categories include no perivascular gaps and perivascular gaps; the quinary categories include no perivascular gaps, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, and more than 40 perivascular gaps.
[0038] Optionally, the magnetic resonance imaging-based perivascular space classification system further includes:
[0039] A first extraction unit, configured to extract a region of interest of the T1-weighted nuclear magnetic resonance image sample to obtain a first region of interest;
[0040] a registration unit connected to the first extraction unit, and configured to register the first region of interest and the T1-weighted nuclear magnetic resonance image sample to a T2-weighted nuclear magnetic resonance image sample space, to obtain a registered first region of interest and a registered T1-weighted image;
[0041] A second extraction unit, connected to the registration unit, is used to extract the region of interest of the T2-weighted magnetic resonance image according to the registered first region of interest to obtain a second region of interest;
[0042] A training unit is respectively connected to the registration unit, the second extraction unit and the classification unit, and is used to iteratively train the three-dimensional convolutional neural network according to the registered first region of interest, the second region of interest and the corresponding categories, based on the loss function, using back propagation and gradient descent algorithms until the loss function converges to obtain an optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model.
[0043] Optionally, the first extraction unit includes:
[0044] A region extraction module, used to extract the basal ganglia region and the semiovale centrum region of the T1-weighted MRI image sample;
[0045] An expansion module, connected to the region extraction module, for performing morphological binary expansion on the basal ganglia region to obtain an expanded basal ganglia region;
[0046] The normalization module is connected to the expansion module and the region extraction module respectively, and is used to perform normalization processing on the expanded basal ganglia region and the semioval center region respectively to obtain a first region of interest.
[0047] Optionally, the three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively;
[0048] The input layer is used to transmit the registered first region of interest and the second region of interest to the hidden layer;
[0049] The hidden layer is used to extract features from the registered first region of interest and the second region of interest to obtain initial feature information;
[0050] The global average pooling layer is connected to the hidden layer, and the global average pooling layer is used to perform feature selection on the initial feature information to obtain optimal feature information;
[0051] The first linear layer is connected to the global average pooling layer, and the first linear layer is used to classify the optimal feature information to obtain a first classification result; the first classification result is no perivascular space or perivascular space;
[0052] The second linear layer is connected to the global average pooling layer, and the second linear layer is used to classify the optimal feature information to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces.
[0053] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0054] Based on T1- and T2-weighted MRI images, a gap classification model was established through deep learning methods, so that the perivascular spaces corresponding to the binary and quinary classifications can be automatically determined using T1- and T2-weighted MRI images. At the same time, the two rating tasks of perivascular space binary classification and quinary classification were combined into one model, which improved the accuracy of perivascular space classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0056] Figure 1 It is a flow chart of the method for classifying perivascular spaces based on magnetic resonance imaging of the present invention;
[0057] Figure 2 It is a schematic diagram of the structure of the gap classification model;
[0058] Figure 3 The figure is a schematic diagram of the module structure of the perivascular space classification system based on magnetic resonance imaging of the present invention.
[0059] Explanation of symbols:
[0060] Image acquisition unit-1, classification unit-2, first extraction unit-3, registration unit-4, second extraction unit-5, training unit-6. DETAILED DESCRIPTION
[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0062] The purpose of the present invention is to provide a method and system for classifying perivascular spaces based on magnetic resonance imaging. By establishing a space classification model based on T1 and T2 weighted magnetic resonance images through a deep learning method, the perivascular spaces corresponding to binary and quinary classifications can be automatically determined using T1 and T2 weighted magnetic resonance images. At the same time, the two rating tasks of binary and quinary classification of perivascular spaces are combined into one model, thereby improving the accuracy of perivascular space classification.
[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0064] In magnetic resonance imaging, contrast depends on the magnetism and number of hydrogen nuclei in the imaged area. By running different sequences with different weights, different contrasts can be selected in the area to be imaged. If the magnetic resonance image mainly reflects the difference in T1 values between tissues, it is a T1-weighted image (T1weighted image, T1WI). T1 images have a higher resolution in the vertical cross-sectional direction and can better reflect brain structural information; if it mainly reflects the difference in T2 values between tissues, it is a T2-weighted image (T2weighted image, T2WI).
[0065] like Figure 1 As shown, the perivascular space classification method based on nuclear magnetic resonance imaging of the present invention includes:
[0066] S1: Acquire magnetic resonance images to be classified; the magnetic resonance images to be classified include T1-weighted magnetic resonance images and T2-weighted magnetic resonance images.
[0067] S2: According to the magnetic resonance image to be classified, based on the gap classification model, determine the first classification result and the second classification result corresponding to the magnetic resonance image to be classified. The first classification result is no perivascular gap or perivascular gap. The second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps.
[0068] The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set. The training sample set includes multiple groups of sample nuclear magnetic resonance images and categories corresponding to each group of the sample nuclear magnetic resonance images. The sample nuclear magnetic resonance images include T1-weighted nuclear magnetic resonance image samples and T2-weighted nuclear magnetic resonance image samples. The categories characterize the levels of perivascular spaces, including two categories and five categories; the two categories include no perivascular spaces and perivascular spaces. The five categories include no perivascular spaces, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, and more than 40 perivascular spaces.
[0069] The present invention uses a deep learning method to complete the automatic classification of the rating of enlarged perivascular spaces through T1 and T2 weighted MRI imaging data. The classification results are completely objective, avoiding the evaluation differences caused by the observation process of medical staff. With the help of the powerful computing power of computers, it can quickly and accurately predict the patient's future disability, eliminating tedious evaluation time and effectively reducing the doctor's annotation pressure. In addition, a multi-task learning method is used to combine the binary and five-class rating tasks to alleviate the overfitting of the gap classification model and improve the generalization ability of the gap classification model.
[0070] Furthermore, the method for establishing the gap classification model specifically includes:
[0071] S101: Extract the region of interest of the T1-weighted MRI image sample to obtain a first region of interest. The first region of interest includes the basal ganglia region and the semioval centrum region. Since the enlarged perivascular space mainly appears in the basal ganglia and the semioval centrum, in order to reduce the redundant information input into the model, the two ROIs are extracted by the region of interest (ROI) extraction method. Since the T1-weighted MRI image contains rich structural information, the mask of the basal ganglia and the semioval centrum region is extracted using the T1-weighted MRI image using the FreeSurfer software.
[0072] S102: registering the first region of interest and the T1-weighted MRI image sample to a T2-weighted MRI image sample space to obtain a registered first region of interest and a registered T1-weighted image.
[0073] S103: Extract the region of interest of the T2-weighted MRI image according to the registered first region of interest to obtain a second region of interest. Specifically, firstly, the basal ganglia region, the semiovale center region and the T1-weighted MRI image sample are registered to the T2-weighted MRI image sample space, and then the registered mask is applied to the registered T1-weighted image and T2-weighted MRI image of the corresponding sample to obtain the ROI of the T1-weighted and T2-weighted MRI images.
[0074] S104: According to the registered first region of interest, the second region of interest and the corresponding categories, based on the loss function, the three-dimensional convolutional neural network is iteratively trained using back propagation and gradient descent algorithms until the loss function converges to obtain an optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model.
[0075] Specifically, step S101 includes:
[0076] The basal ganglia region and the centrum semiovale region of the T1-weighted MRI image sample are extracted.
[0077] The basal ganglia region is morphologically binary expanded to obtain an expanded basal ganglia region. Since the enlarged perivascular space is usually distributed on the boundary of the basal ganglia, the present invention performs morphological binary expansion on the basal ganglia region to make the perivascular space clearer.
[0078] The expanded basal ganglia region and the semioval center region are normalized to obtain a first region of interest. In this embodiment, the edges are first calculated and the maximum value of all samples is taken for cropping, so that the size of the magnetic resonance imaging samples in the sample set is unified, which is convenient for batch processing, while reducing the amount of calculation and saving computing resources. By normalizing the expanded basal ganglia region and the semioval center region, background voxel information is prevented from interfering with data distribution, thereby improving the accuracy of the gap classification model.
[0079] Furthermore, the three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively. In this embodiment, the first linear layer and the second linear layer are independent of each other to respectively realize two-classification and five-classification. The first linear layer includes 1 neuron, and the second linear layer includes 5 neurons.
[0080] Step S104 specifically includes:
[0081] The registered first region of interest and the second region of interest are transmitted to the hidden layer through the input layer.
[0082] The registered first region of interest and the second region of interest are respectively subjected to feature extraction through a hidden layer to obtain initial feature information.
[0083] The initial feature information is subjected to feature selection through the global average pooling layer to obtain optimal feature information.
[0084] The optimal feature information is classified through the first linear layer to obtain a first classification result; the first classification result is that there is no perivascular space or there is a perivascular space.
[0085] The optimal feature information is classified through the second linear layer to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces.
[0086] A loss value of a loss function is determined according to the first classification result, the second classification result, the category corresponding to the first region of interest, and the category corresponding to the second region of interest.
[0087] Based on the loss value of the loss function, back propagation and gradient descent algorithms are used to iteratively train the hidden layer, the global average pooling layer, the first linear layer and the second linear layer.
[0088] The enlarged perivascular space is usually tubular in the direction perpendicular to the cross section, so its visual rating needs to be compared between multiple cross sections, and the most serious layer is taken as the rating of the entire area, and contextual information needs to be considered. The present invention uses a 3D convolutional neural network based on the DenseNet idea to construct a convolutional neural network model PVS-DenseNet. The 3D convolutional neural network has the ability to effectively extract contextual information, and the neural network model is built using the Pytorch deep learning framework. During the training process, the hyperparameters of the neural network are adjusted, including the learning rate, the batch training size, and the number of training rounds. During training, the parameters of each convolution are optimized at each iteration.
[0089] Since the binary and quinary classification tasks are highly correlated, the present invention applies a multi-task learning method to the gap classification model. The multi-task learning adopts hard parameter sharing, and only in the last linear layer the parameters of each task are independently updated. Therefore, the gap classification model can simultaneously and accurately classify the enlarged perivascular space into two levels, namely, level 2 and level 5 ratings.
[0090] The enlarged perivascular space is a low signal on T1-weighted MRI images and a high signal on T2-weighted MRI images. The space classification model of the present invention uses fixed-size three-dimensional brain T1 and T2 weighted MRI image data as input. In order to alleviate the gradient vanishing problem and improve the efficiency of feature utilization, inspired by DenseNet, the feature maps learned by different convolution blocks are connected in series, the variables of subsequent layer input are increased, and the feature utilization efficiency and learning effect of the network are improved. Each convolution block will integrate the information of the feature maps output by all previous convolution blocks as input.
[0091] like Figure 2 As shown, the hidden layer has 5 repeated convolutional blocks, each of which contains two identical convolutional units and a 2×2×1 maximum pooling layer with a stride of 2. The convolutional unit contains a 3×3×3 convolutional layer with a stride of 1, a ReLU activation, and a 3D batch normalization layer. In the first convolutional block, the number of feature channels is set to 8, and it is doubled after entering the next convolutional block to infer a sufficiently rich representation of brain information.
[0092] Specifically, the loss function is:
[0093] L PVS =0.5*L CE +0.5*L ICC ;
[0094]
[0095] Among them, L PVS is the loss function value, L CE is the cross entropy loss, L ICC is the intraclass correlation coefficient loss, MSR is the mean square of the ratings of different samples by the same rater, MSC is the mean square of the ratings of the same sample by different raters, MSE is the mean square error of the ratings of all samples by all raters, and n is the number of samples. CE The traditional cross entropy loss is used.
[0096] ICC (intraclass correlation coefficient) is one of the reliability coefficient indicators for measuring and evaluating inter-observer reliability and test-retest reliability, and is defined as follows:
[0097]
[0098] Introducing L ICCThe loss can effectively improve the classification effect of the model, and significantly improve the classification evaluation index ICC (increased by about 37%).
[0099] In addition, after the model training is completed, the present invention uses a cross-validation method to verify the model, and uses AUC and ICC indicators to verify the model.
[0100] like Figure 3 As shown, the perivascular space classification system based on nuclear magnetic resonance imaging of the present invention includes: an image acquisition unit 1 and a classification unit 2.
[0101] The image acquisition unit 1 is used to acquire the to-be-classified MRI images, which include T1-weighted MRI images and T2-weighted MRI images.
[0102] The classification unit 2 is connected to the image acquisition unit 1, and the classification unit 2 is used to determine the first classification result and the second classification result corresponding to the magnetic resonance image to be classified based on the gap classification model according to the magnetic resonance image to be classified. The first classification result is no perivascular gap or perivascular gap. The second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps.
[0103] The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set; the training sample set includes multiple groups of sample magnetic resonance images and categories corresponding to each group of sample magnetic resonance images; the sample magnetic resonance images include T1-weighted magnetic resonance image samples and T2-weighted magnetic resonance image samples; the categories characterize the levels of perivascular gaps, including binary and quinary categories; the binary categories include no perivascular gaps and perivascular gaps; the quinary categories include no perivascular gaps, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, and more than 40 perivascular gaps.
[0104] Furthermore, the perivascular space classification system based on nuclear magnetic resonance imaging also includes: a first extraction unit 3 , a registration unit 4 , a second extraction unit 5 and a training unit 6 .
[0105] The first extraction unit 3 is used to extract the region of interest of the T1-weighted MRI image sample to obtain a first region of interest, wherein the region of interest includes the basal ganglia region and the centrum semiovale region.
[0106] The registration unit 4 is connected to the first extraction unit 3, and is used to register the first region of interest and the T1-weighted MRI image sample to the T2-weighted MRI image sample space to obtain the registered first region of interest and the registered T1-weighted image.
[0107] The second extraction unit 5 is connected to the registration unit 4, and is used for extracting the region of interest of the T2-weighted magnetic resonance image according to the registered first region of interest to obtain a second region of interest.
[0108] The training unit 6 is connected to the alignment unit 4 and the second extraction unit 5 respectively. The training unit 6 is used to iteratively train the three-dimensional convolutional neural network according to the aligned first region of interest, the second region of interest and the corresponding categories, based on the loss function, using back propagation and gradient descent algorithms until the loss function converges to obtain the optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model.
[0109] Specifically, the first extraction unit 3 includes: a region extraction module, a dilation module and a normalization module.
[0110] The region extraction module is used to extract the basal ganglia region and the centrum semiovale region of the T1-weighted MRI image sample.
[0111] The expansion module is connected to the region extraction module, and is used to perform morphological binary expansion on the basal ganglia region to obtain an expanded basal ganglia region.
[0112] The normalization module is connected to the expansion module and the region extraction module respectively, and is used to perform normalization processing on the expanded basal ganglia region and the semioval center region respectively to obtain a first region of interest.
[0113] Furthermore, the three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively;
[0114] The input layer is used to transmit the registered first region of interest and the second region of interest to the hidden layer.
[0115] The hidden layer is used to extract features from the registered first region of interest and the second region of interest to obtain initial feature information.
[0116] The global average pooling layer is connected to the hidden layer, and the global average pooling layer is used to perform feature selection on the initial feature information to obtain optimal feature information.
[0117] The first linear layer is connected to the global average pooling layer, and the first linear layer is used to classify the optimal feature information to obtain a first classification result; the first classification result is no perivascular space or perivascular space.
[0118] The second linear layer is connected to the global average pooling layer, and the second linear layer is used to classify the optimal feature information to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces.
[0119] Compared with the prior art, the perivascular space classification system based on nuclear magnetic resonance imaging of the present invention has the same beneficial effects as the above-mentioned perivascular space classification method based on nuclear magnetic resonance imaging, which will not be described in detail here.
[0120] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0121] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A method for classifying perivascular spaces based on magnetic resonance imaging, characterized in that: The perivascular space classification method based on nuclear magnetic resonance imaging includes: Acquiring magnetic resonance imaging (MRI) images to be classified; the magnetic resonance imaging (MRI) images to be classified include T1-weighted MRI images and T2-weighted MRI images; According to the magnetic resonance image to be classified, based on the gap classification model, determining a first classification result and a second classification result corresponding to the magnetic resonance image to be classified; the first classification result is no perivascular gap or perivascular gap; the second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps; The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set; the training sample set includes multiple groups of sample nuclear magnetic resonance images and categories corresponding to each group of the sample nuclear magnetic resonance images; the sample nuclear magnetic resonance images include T1-weighted nuclear magnetic resonance image samples and T2-weighted nuclear magnetic resonance image samples; the categories represent the levels of perivascular gaps, including two categories and five categories; the two categories include no perivascular gaps and perivascular gaps; the five categories include no perivascular gaps, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, and more than 40 perivascular gaps; The method for establishing the gap classification model comprises: Extracting a region of interest of the T1-weighted nuclear magnetic resonance image sample to obtain a first region of interest; Registering the first region of interest and the T1-weighted nuclear magnetic resonance image sample to the T2-weighted nuclear magnetic resonance image sample space to obtain a registered first region of interest and a registered T1-weighted image; Extracting the region of interest of the T2-weighted magnetic resonance image according to the registered first region of interest to obtain a second region of interest; According to the registered first region of interest, the second region of interest and the corresponding categories, based on the loss function, the three-dimensional convolutional neural network is iteratively trained using back propagation and gradient descent algorithms until the loss function converges to obtain an optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model; The loss function is: L PVS =0.5*L CE +0.5*L ICC ; Among them, L PVS is the loss function value, L CE is the cross entropy loss, L ICC is the intra-class correlation coefficient loss, MSR is the mean square of the ratings of different samples by the same rater, MSC is the mean square of the ratings of the same sample by different raters, MSE is the mean square error of the ratings of all samples by all raters, and n is the number of samples.
2. The method for classifying perivascular spaces based on magnetic resonance imaging according to claim 1, characterized in that: The extracting the region of interest of the T1-weighted MRI image sample to obtain the first region of interest specifically includes: Extracting the basal ganglia region and the centrum semiovale region of the T1-weighted MRI image sample; Performing morphological binary expansion on the basal ganglia region to obtain an expanded basal ganglia region; The expanded basal ganglia region and the semioval center region are respectively normalized to obtain a first region of interest.
3. The method for classifying perivascular spaces based on magnetic resonance imaging according to claim 1, characterized in that: The three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively; The iterative training of the three-dimensional convolutional neural network according to the categories corresponding to the first ROI and the second ROI after the registration and based on the loss function and using the back propagation and gradient descent algorithms specifically includes: Transmitting the registered first region of interest and the second region of interest to a hidden layer through an input layer; Extracting features of the first region of interest and the second region of interest after the registration through a hidden layer to obtain initial feature information; Performing feature selection on the initial feature information through the global average pooling layer to obtain optimal feature information; Classifying the optimal feature information through the first linear layer to obtain a first classification result; the first classification result is no perivascular space or a perivascular space; Classifying the optimal feature information through the second linear layer to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces; Determining a loss value of a loss function according to the first classification result, the second classification result, the category corresponding to the first region of interest, and the category corresponding to the second region of interest; Based on the loss value of the loss function, back propagation and gradient descent algorithms are used to iteratively train the hidden layer, the global average pooling layer, the first linear layer and the second linear layer.
4. A perivascular space classification system based on magnetic resonance imaging, characterized in that: The perivascular space classification system based on magnetic resonance imaging includes: An image acquisition unit, used for acquiring nuclear magnetic resonance images to be classified; the nuclear magnetic resonance images to be classified include T1-weighted nuclear magnetic resonance images and T2-weighted nuclear magnetic resonance images; a classification unit connected to the image acquisition unit, and configured to determine, according to the magnetic resonance image to be classified, a first classification result and a second classification result corresponding to the magnetic resonance image to be classified based on a gap classification model; the first classification result is no perivascular gap or perivascular gap; the second classification result is no perivascular gap, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, or more than 40 perivascular gaps; The gap classification model is obtained by pre-training a three-dimensional convolutional neural network using a training sample set; the training sample set includes multiple groups of sample nuclear magnetic resonance images and categories corresponding to each group of the sample nuclear magnetic resonance images; the sample nuclear magnetic resonance images include T1-weighted nuclear magnetic resonance image samples and T2-weighted nuclear magnetic resonance image samples; the categories represent the levels of perivascular gaps, including two categories and five categories; the two categories include no perivascular gaps and perivascular gaps; the five categories include no perivascular gaps, 1-10 perivascular gaps, 11-20 perivascular gaps, 21-40 perivascular gaps, and more than 40 perivascular gaps; The perivascular space classification system based on nuclear magnetic resonance imaging also includes: A first extraction unit, configured to extract a region of interest of the T1-weighted nuclear magnetic resonance image sample to obtain a first region of interest; a registration unit connected to the first extraction unit, and configured to register the first region of interest and the T1-weighted nuclear magnetic resonance image sample to a T2-weighted nuclear magnetic resonance image sample space, to obtain a registered first region of interest and a registered T1-weighted image; A second extraction unit, connected to the registration unit, is used to extract the region of interest of the T2-weighted magnetic resonance image according to the registered first region of interest to obtain a second region of interest; A training unit is connected to the registration unit, the second extraction unit and the classification unit respectively, and is used to iteratively train the three-dimensional convolutional neural network according to the registered first region of interest, the second region of interest and the corresponding categories, based on a loss function, using back propagation and gradient descent algorithms until the loss function converges, so as to obtain an optimal three-dimensional convolutional neural network; the optimal three-dimensional convolutional neural network is a gap classification model; The loss function is: L PVS =0.5*L CE +0.5*L ICC ; Among them, L PVS is the loss function value, L CE is the cross entropy loss, L ICC is the intra-class correlation coefficient loss, MSR is the mean square of the ratings of different samples by the same rater, MSC is the mean square of the ratings of the same sample by different raters, MSE is the mean square error of the ratings of all samples by all raters, and n is the number of samples.
5. The perivascular space classification system based on magnetic resonance imaging according to claim 4, characterized in that: The first extraction unit comprises: A region extraction module, used to extract the basal ganglia region and the semiovale centrum region of the T1-weighted MRI image sample; An expansion module, connected to the region extraction module, for performing morphological binary expansion on the basal ganglia region to obtain an expanded basal ganglia region; The normalization module is connected to the expansion module and the region extraction module respectively, and is used to perform normalization processing on the expanded basal ganglia region and the semioval center region respectively to obtain a first region of interest.
6. The perivascular space classification system based on magnetic resonance imaging according to claim 4, characterized in that: The three-dimensional convolutional neural network includes an input layer, a hidden layer and an output layer connected in sequence; the output layer includes a global average pooling layer, a first linear layer and a second linear layer; the global average pooling layer is connected to the hidden layer, the first linear layer and the second linear layer respectively; The input layer is used to transmit the registered first region of interest and the second region of interest to the hidden layer; The hidden layer is used to extract features from the registered first region of interest and the second region of interest to obtain initial feature information; The global average pooling layer is connected to the hidden layer, and the global average pooling layer is used to perform feature selection on the initial feature information to obtain optimal feature information; The first linear layer is connected to the global average pooling layer, and the first linear layer is used to classify the optimal feature information to obtain a first classification result; the first classification result is no perivascular space or perivascular space; The second linear layer is connected to the global average pooling layer, and the second linear layer is used to classify the optimal feature information to obtain a second classification result; the second classification result is no perivascular space, 1-10 perivascular spaces, 11-20 perivascular spaces, 21-40 perivascular spaces, or more than 40 perivascular spaces.
Citation Information
Patent Citations
Intracranial artery segmentation method and system
CN113313728A
Method and system for identification of cerebrovascular abnormalities
US20200020435A1