Construction method and device of VCI recognition algorithm for diffusion tensor imaging
Through the convolutional network in deep learning technology, the image features of diffusion tensor imaging data are automatically extracted, and the vascular cognitive impairment recognition algorithm is constructed, which solves the problem of traditional machine learning methods relying on artificial feature engineering, and achieves a more accurate and objective VCI Identification, suitable for clinical diagnosis and research.
Patent Information
- Application Number
- CN202410963815.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-12
AI Technical Summary
The existing recognition algorithm for vascular cognitive impairment relies on traditional machine learning and requires manual extraction of image features, which is highly subjective, and simple numerical features cannot effectively reflect brain function and structural changes, limiting recognition performance.
Using convolutional networks in deep learning technology, we automatically extract image features based on diffusion tensor imaging data, build VCI recognition algorithms to achieve early recognition.
It achieves more accurate and objective VCI recognition, which is suitable for clinical diagnosis and research, and provides auxiliary basis for early diagnosis and treatment and research.
Smart Images

Figure CN118942127A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of neuroimaging markers. Specifically, the present application provides a method and device for constructing a VCI recognition algorithm for diffusion tensor imaging. Background Art
[0002] Vascular cognitive impairment (VCI) is a stroke or subclinical vascular brain injury caused by cerebrovascular lesions and their risk factors. VCI includes all disease stages from mild cognitive impairment originating from cerebrovascular diseases to vascular dementia. It is divided into three subtypes according to its clinical manifestations: 1) non-dementia vascular cognitive impairment; 2) vascular dementia; 3) mixed dementia, among which non-dementia vascular cognitive impairment is the most common subtype of VCI. If vascular cognitive impairment can be identified early, it will help to intervene in VCI patients early, which has certain clinical significance.
[0003] At present, the diagnosis of VCI is mainly based on clinical manifestations and neuropsychological scales, which are very subjective. Existing studies have reported that the brain structure, brain function and brain network of patients with vascular cognitive impairment are different from those of normal subjects. Based on this, some studies have manually extracted the features of magnetic resonance imaging and established traditional machine learning models to realize the identification of VCI. However, traditional machine learning methods require subjective feature engineering. In addition, simple imaging index values may not reflect changes in brain function and brain structure, limiting the performance of the model. Deep learning methods can automatically extract image features from images without the need for manual feature selection based on experience, capturing brain changes between VCI and normal people in a more objective way, and achieving more accurate diagnosis and treatment of VCI.
[0004] Most existing vascular cognitive impairment recognition algorithms are based on traditional machine learning methods, which require manual extraction of features from medical images that can reflect changes in brain structure. This feature engineering is highly subjective, and performance is often determined by the extracted features. In addition, the brain changes of vascular cognitive impairment and normal subjects are more complex, and simple numerical features extracted from medical images alone often cannot effectively reflect brain changes, thus limiting the performance of vascular cognitive impairment recognition. Summary of the invention
[0005] In order to solve the above problems, the present invention innovatively uses the convolutional network in deep learning technology to develop a method for constructing a vascular cognitive impairment recognition algorithm for diffusion tensor imaging. The purpose of the present invention is to propose a method for constructing a VCI recognition algorithm based on diffusion tensor imaging data, using a convolutional neural network to automatically extract the image features of diffusion tensor imaging data, rather than numerical features manually extracted in feature engineering, to achieve early recognition of VCI. This method and system provide assistance and basis for the diagnosis and treatment of VCI and clinical research, and has the potential for clinical application.
[0006] On the one hand, the present application provides a method for constructing a VCI identification algorithm for diffusion tensor imaging, the method comprising:
[0007] 1) Obtain diffusion tensor imaging data of normal subjects and subjects with vascular cognitive impairment;
[0008] 2) Preprocessing of diffusion tensor imaging data;
[0009] 3) Calculate the diffusion anisotropy parameter map for diffusion tensor imaging;
[0010] 4) Establish a model to automatically extract the imaging features of diffusion anisotropy parameter maps and realize the identification of vascular cognitive impairment;
[0011] The method is for non-diagnostic purposes.
[0012] Further, step 1) includes: selecting appropriate vascular cognitive impairment subjects according to the inclusion and exclusion criteria of vascular cognitive impairment; scanning the subjects using a magnetic resonance imaging system to obtain diffusion tensor imaging data;
[0013] Furthermore, the scanning parameters of diffusion tensor imaging in step 1) are as follows: diffusion tensor imaging is acquired using a diffusion-weighted dual spin echo-echo planar imaging sequence with a repetition time of 8000 milliseconds; an echo time of 96 milliseconds; 64 diffusion weighted directions; and a b value of 1000 s / mm 2 ; Deflection angle is 90°; Field of view is 224×224mm 2 ; Matrix is 128×128; Plane resolution is 1.75×1.75mm 2 / voxel.
[0014] Furthermore, step 2) includes: under the Linux system, applying the brain diffusion imaging analysis process toolbox in Matlab software to analyze the diffusion tensor imaging data; the analysis uses the standard pipeline in the brain diffusion imaging analysis process toolbox for preprocessing, including: image format conversion, resampling, removal of non-brain tissue, image cropping, turbine correction, head motion correction and intensity normalization.
[0015] The brain diffusion imaging analysis pipeline toolbox is also known as PANDA (Pipeline for Analyzing brain diffusion imaging Ages).
[0016] Further, step 3) comprises:
[0017] 3-1) Using a tensor model to model and analyze the diffusion characteristics of tissue water molecules to obtain a tensor matrix, thereby calculating a diffusion anisotropy parameter map;
[0018] The diffusion tensor model is used to quantify and decompose the signal data of diffusion anisotropy. The formula is as follows:
[0019]
[0020] Among them, D xx , D yy , D zz is the diffusion coefficient applied along the x, y, and z axes of the spatial rectangular coordinate system. The diffusion tensor model contains three eigenvalues λ1λ2λ3 and eigenvectors of related quantities, where the eigenvector reflects the diffusion direction of water molecules and the eigenvalue reflects the degree of diffusion of water molecules in each direction. In a homogeneous medium, water molecules are isotropic, that is, the diffusion rate in each direction in space is the same. At this time, the diffusion tensor D is a spherical model. In white matter fibers, water molecules are restricted in their diffusion mode due to the influence of myelin sheaths and have high anisotropy. At this time, the model is ellipsoidal, and the size of the eigenvalue can be expressed as λ1>λ2>λ3, where the maximum characteristic λ1 corresponds to the direction v1 of the fiber running through the voxel, and the corresponding directions of λ2 and λ3 are radial directions perpendicular to it.
[0021] The anisotropy analysis is performed using fractional anisotropy (FA), which reflects the ratio of the anisotropic component to the diffusion tensor and is between 0 and 1. The formula is as follows:
[0022]
[0023] The mean diffusivity (MD) is used to reflect the ability of water molecules to diffuse per unit time in a certain voxel. The formula is as follows:
[0024]
[0025] In addition, when calculating the diffusion index, the axial diffusion coefficient (AD) and radial diffusion coefficient (RD) are used to analyze the diffusion levels in different directions. The formula is as follows:
[0026] AD=λ1#(4)
[0027]
[0028] 3-2) The diffusion anisotropy parameter map is aligned from the individual space to the standard space to obtain the diffusion anisotropy parameter map in the standard MNI space.
[0029] Furthermore, in step 4), the five-fold cross-validation method is used to adjust the model parameters during the construction process; the data set is randomly divided into five non-overlapping data subsets, and then four data subsets are used as training sets to train the model, and the remaining one data set is used as a validation set; repeat five times, each data set is used as a validation set, and five models are obtained; according to the average accuracy results of the five models, a suitable and stable parameter combination is selected to determine the final model of the algorithm, the trained model is saved, and the final recognition algorithm is constructed; including:
[0030] 4-1) Preprocessing of diffusion anisotropy parameter map
[0031] The anisotropic parameter maps are shuffled, and the training set and validation set data are processed separately: for the training set data, online enhancement is performed to increase the diversity of the images, including normalizing the parameter maps; for the test set, only the parameter maps are normalized without online enhancement.
[0032] 4-2) Construction and parameter adjustment of vascular cognitive impairment recognition algorithm
[0033] DenseNet-169 is used as the algorithm model to automatically extract the features of the diffusion anisotropy parameter map and realize the recognition of VCI. The algorithm is optimized by Adam optimizer. The loss function adopts the cross entropy loss function, which is weighted according to the ratio of VCI subjects to normal subjects. The formula of the cross entropy loss function is:
[0034]
[0035] in is the cross entropy loss function, y i and are the actual labels of the i samples and the probability of VCI predicted by the model, λ1 and λ2 are the weights of VCI and normal subjects, respectively, and N is the total number of samples; λ1 and λ2 are 0.4 and 1; the diffusion anisotropy parameter map is input into the model to obtain the prediction of the algorithm, the prediction is compared with the actual label, and the loss function is obtained through cross entropy; the algorithm is iterated and the parameters are updated according to the optimization rule of the Adam optimizer to realize automatic feature extraction and VCI prediction; the learning rate of Adam is 1e-5, 12 regularization is 1e-3, betas is (0.9, 0.999), and eps is 1e-8. The accuracy is selected as the indicator for evaluating the algorithm, which is used to judge the accuracy of the algorithm in identifying cognitive impairment;
[0036] The training and validation data were divided into five groups using the five-fold cross-validation method. The training was conducted five times in total. During the training, the diffusion tensor image anisotropy parameter map was input into the model. The loss between the model prediction result and the true label was calculated by the loss function. The optimizer used the gradient descent algorithm to optimize and update the model parameters to minimize the training loss and improve the accuracy of the model in identifying vascular cognitive impairment.
[0037] The formula for the gradient descent algorithm is as follows:
[0038]
[0039] Where θ is the parameter of the model, α is the learning rate, which controls the step size of each parameter update, and Loss is the loss function. It is the gradient of the loss function with respect to the parameter θ, which indicates the rate of change of the loss function at the current parameter;
[0040] After all training set data are input into the model to complete a round of training, the effect of model training is verified on the validation set data, the accuracy of the model in identifying VCI is calculated, and the optimal model is saved.
[0041] On the other hand, the present application provides application of the above method in the study of brain mechanisms of vascular cognitive impairment.
[0042] On the other hand, the present application provides the application of the above method in macro statistics or research on the health status of a population, and the application does not include diagnostic purposes.
[0043] Due to the accuracy and convenience of the present application, the applications include but are not limited to the study of the possibility of developing vascular cognitive dysfunction in a specific population, the study of the association between vascular cognitive dysfunction and other diseases, etc.
[0044] The application objects / patients of the present application include but are not limited to humans or other animals, such as model animals.
[0045] On the other hand, the present application provides an apparatus for executing the above method, wherein the apparatus includes a module part for executing step 1), step 2), step 3), and step 4).
[0046] The above module steps include, but are not limited to, a magnetic resonance imaging system (for step 1), a calculator that executes each step of the processing (such as a local processor, a remote processor or a cloud processor of X86 or ARM architecture), a memory that stores each step of the algorithm (such as a local and online memory that stores the corresponding software), and an output device that outputs the results (such as a display, print or storage read-write device).
[0047] Combined with DenseNet, a method for constructing a vascular cognitive impairment recognition algorithm for diffusion tensor imaging is proposed. This method can automatically and objectively extract image features in diffusion tensor imaging to achieve accurate VCI recognition. The recognition algorithm established by this method has a fast and objective recognition process, which is suitable for use in primary hospitals and scientific research processes. At the same time, this method can provide services for early accurate diagnosis and treatment and early screening, and provide assistance and basis for clinical VCI diagnosis and research, which has promotion significance and value. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 The technical roadmap for this application;
[0049] Figure 2 This is the recognition effect result of the algorithm constructed in the embodiment. DETAILED DESCRIPTION
[0050] The present invention is described in detail below in conjunction with specific embodiments, but the protection scope of the present invention is not limited to the following embodiments.
[0051] The VCI recognition method proposed in this application has nothing to do with the age, gender, occupation, education level, region, and whether or not the subject suffers from underlying diseases, but is only related to the diffusion tensor imaging data of the subject. The VCI recognition method proposed in this application is based on deep learning technology, models a convolutional neural network to process high-dimensional diffusion tensor imaging data, automatically extracts specific features related to vascular cognitive impairment in diffusion tensor images, and automatically recognizes the features. The recognition method can process images objectively and quickly, and is suitable for scientific research related to clinical VCI diagnosis and VCI recognition.
[0052] The specific process of implementing the VCI identification method proposed in this application is as follows:
[0053] The basic process of VCI identification method is as follows Figure 1 As shown:
[0054] 1) Obtain diffusion tensor imaging data:
[0055] According to the inclusion and exclusion criteria of vascular cognitive impairment, subjects who met the inclusion criteria were selected to establish the vascular cognitive impairment group and the normal group (a total of 181 subjects, including 124 in the vascular cognitive impairment group and 57 in the normal group). Siemens 3T magnetic resonance imaging equipment equipped with a 32-channel head coil was used to acquire the head magnetic resonance images of the subjects, and a diffusion-weighted dual spin echo planar imaging sequence was set to obtain magnetic resonance diffusion tensor imaging data. During the acquisition process, the subjects lay flat on the examination bed, and fillers were filled in the gaps around their heads to prevent head movement from affecting the quality of the acquired images.
[0056] 2) MRI Diffusion Tensor Imaging Data Preprocessing:
[0057] The MRI diffusion tensor image data were preprocessed using the standard preprocessing process of the PANDA toolkit in Matlab under Linux. The specific operations included image format conversion, image resampling, brain tissue segmentation, image cropping, eddy current correction, and head motion correction.
[0058] 3) Calculate the diffusion anisotropy parameter image:
[0059] The diffusion characteristics of tissue water molecules were analyzed by tensor modeling. The diffusion tensor parameter index was calculated to obtain the diffusion tensor anisotropy parameter image. The diffusion tensor parameter image was registered from the individual space to the MNI standard space. The diffusion anisotropy parameter map in the standard MNI space was obtained.
[0060] 4) Five-fold cross validation:
[0061] The order of all anisotropic parameter images was shuffled, and the test set was first divided from the data. The remaining data set was randomly divided into 5 non-overlapping data subsets using 5-fold cross validation. Then 4 data subsets were used as training sets to train the model, and the remaining 1 data set was used as the validation set. This was repeated 5 times to obtain five combinations of training sets and validation sets.
[0062] 4-1) Perform image standardization and offline image enhancement operations on the data in the training set to increase the diversity of the image. Offline image enhancement operation refers to saving the processed image after enhancing the image. During training, the enhanced image and the original image are used as the training set for training, which expands the amount of training set data. Specific enhancement operations include randomly flipping the image and randomly scaling the image, and the probability of each operation occurring is 0.2. For the validation set data and the test set data, only image standardization is performed.
[0063] The standardized formula is:
[0064]
[0065] where μ xis the image mean, std x is the image standard deviation.
[0066] 4-2) Model construction:
[0067] A three-dimensional DenseNet-169 network model was established under the pytorch framework. The number of input feature channels of the model is 1, the number of initial features is 128, the number of features added in each layer of the network is 64, the expansion coefficient in the bottleneck layer is 4, and the model has a total of 4 modules, and the number of layers in each module is configured as [6, 12, 32, 32]. The Xavier initialization method is used to initialize the parameters of the model to avoid the problem of gradient explosion or gradient disappearance in the early stage of training, help accelerate model convergence, and improve model performance.
[0068] 4-3) Model training:
[0069] The training and validation data were divided into five groups using the five-fold cross-validation method. A total of five training times were required. During the training, the diffusion tensor image anisotropy parameter map was input into the model. The loss between the model prediction result and the true label was calculated by the loss function. The optimizer used the gradient descent algorithm to optimize and update the model parameters to minimize the training loss and improve the accuracy of the model in identifying vascular cognitive impairment.
[0070] The formula for the gradient descent algorithm is as follows:
[0071]
[0072] Where θ is the parameter of the model, α is the learning rate, which controls the step size of each parameter update, and Loss is the loss function. It is the gradient of the loss function with respect to the parameter θ, which indicates the rate of change of the loss function at the current parameter.
[0073] After all training set data are input into the model to complete a round of training, the effect of model training is verified on the validation set data, the accuracy of the model in identifying VCI is calculated, and the optimal model is saved.
[0074] Each set of training parameters is kept consistent. The batch size of the data input into the model training is 4 each time. One round of training is completed after all the training data are input into the model for training. To prevent the model from overfitting, the model is required to stop training in advance when the accuracy of identifying VCI in the validation set data does not improve after 30 consecutive rounds of training. The maximum number of training rounds allowed is 200. The cross entropy loss function is used as the loss function. The cross entropy loss function is weighted according to the ratio of the number of VCI objects and normal objects in the training data set. For the detailed definition of the cross entropy loss function, refer to Formula 6. The learning rate of the Adam optimizer is set to 1e-5, the L2 regularization is set to 1e-3, the betas is set to (0.9, 0.999), and the eps is set to 1e-8.
[0075] 5) Model testing:
[0076] The data in the test set were sequentially input into the five optimal models obtained through training. The model prediction results of the data and the true labels were used to calculate the accuracy, sensitivity, specificity, F1 score, and AUC of the model recognition. The evaluation index results obtained by the five models were averaged to obtain the final vascular cognitive impairment recognition performance evaluation results. Figure 2 As shown, the VCI recognition method proposed in this application can achieve a VCI recognition accuracy of 80%, a sensitivity of 0.92, a specificity of 0.74, an F1 score, and an AUC of 0.85 on the test data.
Claims
1. A method for constructing a VCI recognition algorithm for diffusion tensor imaging, characterized in that: The method comprises: 1) Obtain diffusion tensor imaging data of normal subjects and subjects with vascular cognitive impairment; 2) Preprocessing of diffusion tensor imaging data; 3) Calculate the diffusion anisotropy parameter map for diffusion tensor imaging; 4) Establish a model to automatically extract the imaging features of diffusion anisotropy parameter maps and realize the identification of vascular cognitive impairment; The method is for non-diagnostic purposes.
2. The method according to claim 1, wherein step 1) comprises: Appropriate subjects with vascular cognitive impairment were selected according to the inclusion and exclusion criteria of vascular cognitive impairment; the subjects were scanned using a magnetic resonance imaging system to obtain diffusion tensor imaging data.
3. The method according to claim 2, wherein the scanning parameters of the diffusion tensor imaging in step 1) are as follows: the diffusion tensor imaging is acquired using a diffusion-weighted double spin echo-echo planar imaging sequence, with a repetition time of 8000 milliseconds; an echo time of 96 milliseconds; 64 diffusion weighted directions; and a b value of 1000 s / mm 2 ; Deflection angle is 90°; Field of view is 224×224mm 2 ; Matrix is 128×128; Plane resolution is 1.75×1.75mm 2 / voxel.
4. The method according to any one of claims 1 to 3, wherein step 2) comprises: Under Linux system, the brain diffusion imaging analysis process toolbox in Matlab software was used to analyze the diffusion tensor imaging data; The analysis used the standard pipeline in the brain diffusion imaging analysis workflow toolbox for preprocessing, including: image format conversion, resampling, removal of non-brain tissue, image cropping, turbine correction, head motion correction, and intensity normalization.
5. The method according to any one of claims 1 to 4, wherein step 3) comprises: 3-1) The diffusion characteristics of tissue water molecules are modeled and analyzed using a tensor model to obtain a tensor matrix, thereby calculating a diffusion anisotropy parameter map; the diffusion tensor model is used to quantify and decompose the signal data of the diffusion anisotropy, and the formula is as follows: Among them, D xx , D yy , D zz is the diffusion coefficient applied along the x, y, and z axes of the spatial rectangular coordinate system. The diffusion tensor model contains three eigenvalues λ1λ2λ3 and eigenvectors of related quantities. The eigenvector reflects the diffusion direction of water molecules, and the eigenvalue reflects the degree of diffusion of water molecules in each direction. The anisotropy fraction FA is used for anisotropy analysis, which reflects the ratio of the anisotropic component to the diffusion tensor, ranging from 0 to 1. The formula is as follows: The average diffusion MD is used to reflect the ability of water molecules in a certain voxel to diffuse per unit time. The formula is as follows: The axial diffusion coefficient AD and radial diffusion coefficient RD are used to analyze the diffusion levels in different directions. The formula is as follows: AD=λ1#(4) 3-2) The diffusion anisotropy parameter map is aligned from the individual space to the standard space to obtain the diffusion anisotropy parameter map in the standard MNI space.
6. The method according to any one of claims 1 to 5, wherein step 4) preprocesses the extracted diffusion anisotropy parameter map and constructs a VCI identification algorithm, comprising: 4-1) Preprocessing of diffusion anisotropy parameter map The anisotropic parameter maps are shuffled, and the training set and validation set data are processed separately: for the training set data, online enhancement is performed to increase the diversity of the images, including standardization of the parameter maps; for the test set, only the parameter maps are standardized without online enhancement; 4-2) Construction and parameter adjustment of vascular cognitive impairment recognition algorithm DenseNet-169 is used as the algorithm model to automatically extract the features of the diffusion anisotropy parameter map and realize the identification of VCI; the algorithm is optimized using the Adam optimizer; The loss function adopts the cross entropy loss function, which is weighted according to the ratio of VCI subjects to normal subjects. The formula of the cross entropy loss function is: in is the cross entropy loss function, y i and are the actual labels of the i-th sample and the probability that the model predicts that it is VCI, λ1 and λ2 are the weights of VCI and normal subjects, respectively, and N is the total number of samples; λ1 and λ2 are 0.4 and 1; the diffusion anisotropy parameter map is input into the model to obtain the prediction of the algorithm, and the prediction is compared with the actual label, and the loss function is obtained through cross entropy; the algorithm is iterated and the parameters are updated according to the optimization rule of the Adam optimizer to realize automatic feature extraction and VCI prediction; the learning rate of Adam is 1e-5, 12 regularization is 1e-3, betas is (0.9, 0.999), and eps is 1e-8; the accuracy rate is selected as the indicator for evaluating the algorithm, which is used to judge the accuracy of the algorithm in identifying cognitive impairment; During training, the diffusion tensor image anisotropy parameter map is input into the model, and the loss between the model prediction result and the true label is calculated through the loss function. The optimizer uses the gradient descent algorithm to optimize and update the model parameters to minimize the training loss and improve the accuracy of the model in identifying vascular cognitive impairment. The formula for the gradient descent algorithm is as follows: Where θ is the parameter of the model, α is the learning rate, which controls the step size of each parameter update, and Loss is the loss function. It is the gradient of the loss function with respect to the parameter θ, which indicates the rate of change of the loss function at the current parameter.
7. The method according to claim 6, wherein 4-2) uses a five-fold cross-validation method to adjust the model parameters; randomly divide the data set into 5 non-overlapping data subsets, and then use 4 data subsets as training sets to train the model, and the remaining 1 data set is used as a validation set; repeat 5 times, each data set is used as a validation set, and 5 models are obtained; according to the average accuracy results of the 5 models, a suitable and stable parameter combination is selected to determine the final model of the algorithm, the trained model is saved, and the final recognition algorithm is constructed.
8. The method according to claim 1, wherein the subject is a human or an animal.
9. Use of the method according to any one of claims 1 to 8 in the study of brain mechanisms of vascular cognitive impairment.
10. Application of the method according to any one of claims 1 to 8 in macro statistics or research on the health status of a population, wherein the application does not include diagnostic purposes.
11. The use according to claim 10, wherein the use is a study on the possibility of vascular cognitive dysfunction in a specific population, or a study on the association between vascular cognitive dysfunction and other diseases.
12. A device for executing the method according to any one of claims 1 to 8, characterized in that: The device includes a module part for executing step 1), step 2), step 3), and step 4).
Citation Information
Patent Citations
Magnetic resonance diffusion tensor brain image analysis method and system for newborns
CN111311585A
Construction method and device for auxiliary diagnosis model of vascular cognitive impairment
CN115131415A
Image convolutional neural network disease prediction system based on multi-modal magnetic resonance imaging
CN115359045A
Traditional Chinese medicinal material identification method based on improved Resnet neural network
CN115761709A
Extraction method of key image marker of vascular cognitive impairment
CN116309336A
Cited By
Serum marker combination as well as screening method and application thereof
CN119198972A
Vascular cognitive impairment assessment method and device, electronic equipment and storage medium
CN121730761A
Vascular cognitive impairment assessment method and device, electronic equipment and storage medium
CN121730761B
