An autism detection device for a specific age group

Abstract

The present invention belongs to the technical field of medical image processing, and proposes an autism detection device for a specific age group. The autism detection device for a specific age group includes a model training module and an autism detection module; the magnetic resonance structural image features are fused based on the multiple kernel learning method to assist in the autism detection of a specific age group. The multiple kernel learning is used to combine two structural image features, so that the features representing the information within the region and the features representing the relationship between regions are more effectively fused with appropriate weights; in addition, considering the high heterogeneity among ASD patients, it is proposed to divide ASD patients into three age groups, and a classification model suitable for each age group is used within each age group, so as to better complete the classification task.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an autism detection device for a specific age group. Background Art

[0002] Autism Spectrum Disorders (ASD) is a highly heterogeneous neurodevelopmental disorder, and its clinical manifestations are mainly speech development disorders, interpersonal communication disorders, repetitive stereotyped behaviors, intellectual disabilities, etc. It usually develops in childhood, and due to defects in social interaction, it is difficult for patients to live and work like normal people, requiring families and society to invest a lot of extra money and energy, which causes great mental pressure and economic burden on families and society. Intervention in the early stages of the disease can alleviate the social interaction disorders of ASD patients and help ASD patients improve their social interaction skills. Therefore, early diagnosis is crucial for timely intervention of ASD. However, since the pathogenesis of autism is still unclear, there is still no objective biomarker for the diagnosis of ASD. The current diagnosis of autism is mainly based on clinical interviews and behavioral observations, and the diagnostic results are highly subjective, which may lead to missed diagnosis and miss the best time to intervene in ASD patients. Therefore, it is very necessary to establish an objective ASD diagnostic system based on imaging.

[0003] Structural magnetic resonance imaging (sMRI), as a non-invasive technique, has been widely used in the study of the brain morphology of autistic patients due to its high spatial resolution and easy accessibility of data. Univariate analysis methods such as Voxel-Based Morphometry (VBM) and Surface-Based Morphometry (SBM) are commonly used in brain imaging research to detect brain differences at the group level. In previous studies "Khundrakpam B S, Lewis J D, Kostopoulos P, et al. Cortical thickness abnormalities in autism spectrum disorders through late childhood, adolescence, and adulthood: a large-scale MRI study[J]. Cerebral Cortex, 2017, 27(3): 1721-1731.", it was reported that the differences between autistic patients and healthy subjects were concentrated in some areas of the prefrontal, parietal, temporal, and occipital lobes. However, the reported difference regions were inconsistent in different studies. "Braden B B, Riecken C. Thinning faster? Age-related cortical thickness differences in adults with autism spectrum disorder[J]. Research in autism spectrum disorders, 2019, 64: 31-38." showed that there was a problem with the reproducibility of the results. In addition, the univariate analysis method based on the group level limits the study of the relationships between variables and cannot provide subject-specific information that is crucial for personalized diagnosis.

[0004] Machine learning-based methods are increasingly being applied to train classifiers for reliably classifying the diagnostic categories of subjects at the individual level. Among them, the Support Vector Machine (SVM) is the most commonly used classifier in ASD classification research. SVM is a supervised machine learning algorithm with good generalization and robustness, and its purpose is to find a decision boundary that maximizes the boundary between two classes in a high-dimensional space. Among them, the Multiple Kernel Learning (MKL) method is a commonly used kernel matrix-based fusion method in multimodal data fusion. It can effectively integrate heterogeneous information by assigning specified kernel functions and appropriate weights to different features, and has achieved good results in the diagnostic prediction of some mental diseases (such as Alzheimer's disease, attention deficit hyperactivity disorder, etc.). ―Ecker C, Marquand A, J, et al. Describing the brain in autism in five dimensions—magnetic resonance imaging-assisted diagnosis of autism spectrum disorder using a multiparameter classification approach[J]. Journal of Neuroscience, 2010, 30(32): 10612-10623.” and “Ecker C, Rocha-Rego V, Johnston P, et al. Investigating the predictive value of whole-brain structural MR scans in autism: a pattern classification approach[J]. Neuroimage, 2010, 49(1): 44-56.” used SVM to classify ASD and normal people based on structural images, achieving a classification accuracy of about 90%, which proved the value of brain structural features for ASD diagnosis. However, the sample sizes of these studies were small, and the research results obtained may not be reproducible in other cohorts of subjects, and the prediction models may not be universal. The establishment of the large-sample multi-center database ABIDE has solved this problem to a certain extent.However, when using a large sample database and based on structural images, the classification accuracy is greatly reduced. The classification accuracies in "Uddin L Q, Menon V, Young CB, et al. Multivariate searchlight classification of structural magnetic resonance imaging in children and adolescents with autism[J]. Biological psychiatry, 2011, 70(9): 833-841.", "Haar S, Berman S, Behrmann M, et al. Anatomical abnormalities in autism[J]. Cerebral cortex, 2016, 26(4): 1440-1452." and "Sabuncu MR, Konukoglu E. Clinical Prediction from Structural Brain MRI Scans: A Large-Scale Empirical Study. Neuroinformatics. 2014; 13: 31–46." are all around 60%, and it is impossible to distinguish autism patients from normal people. This may be due to the higher heterogeneity of ASD patients in the large sample data and the low discriminability of the selected single feature of the structural image between the two groups. Summary of the Invention

[0005] In view of the existing problems, the present invention provides an autism detection device for a specific age group, which effectively fuses heterogeneous structural features containing complementary information to obtain better classification performance. The device is based on a multi-kernel learning method to fuse magnetic resonance structural image features to assist in the detection of autism for a specific age group. By using multi-kernel learning to combine two structural image features, the features representing the information within the region and the features representing the relationship between regions are more effectively fused with appropriate weights; in addition, considering the high heterogeneity among ASD patients, we propose to divide ASD patients into three age groups and use a classification model suitable for each age group within each age group to better complete the classification task.

[0006] Technical solution of the present invention:

[0007] An autism detection device for a specific age group, comprising a model training module and an autism detection module;

[0008] The model training module is trained based on the large sample multi-center database ABIDE Ⅱ, and includes a feature extraction and feature construction module, three feature selection modules and three classifier modules;

[0009] The specific steps are as follows:

[0010] Use the processing flow provided by Freesurfer (http: / / surfer.nmr.mgh.harvard.edu) to preprocess the brain regions, and use the Desikan-Killiany parcellation template to divide the brain regions in the individual space in detail. The Desikan-Killiany parcellation template divides the cortical surfaces of the bilateral cerebral hemispheres into 68 brain regions, which are used as regions of interest (ROIs).

[0011] (1) Extract the morphological features of the structural images within the regions of interest;

[0012] Extract the morphological features within each region of interest, including cortical thickness, cortical surface area, gray matter volume, cortical folding index, and mean curvature;

[0013] (2) Construct features representing the relationships between the regions of interest;

[0014] Considering that calculating the correlation of feature vectors within regions pairwise will ignore the potential influence of other regions, the method of calculating Pearson correlation is not used to quantify the relationships between regions. Instead, the elastic net is used to quantify the relationships between the target brain region and multiple other brain regions; use the 5 morphological features extracted within the regions of interest in step (1) to form the feature matrix of each subject in the ABIDE II database, and use the feature matrix to construct a multi-feature network (MFN) representing the relationships between the target brain region and the other brain regions; the multi-feature network MFN is obtained through a multiple regression process. The feature vector of the selected target brain region and the feature vectors of the other brain regions are used as the target variable and the predictor variables in the regression respectively. Define the linear regression model y = Aw, where y is the target vector, A is the feature matrix composed of the other brain regions, and w is the regression coefficient vector; obtain the sparse solution by solving the regularized optimization problem. In the regression process, the L1 norm regularization can exclude irrelevant predictors, thus obtaining a sparse solution; however, the L1 norm regularization can only identify the predictor variables equal to the number of observations and only select one from the highly correlated predictor variables, while the L2 norm regularization can make up for the deficiency of the L1 norm and prevent the obtained solution from being too sparse; the formula is:

[0015]

[0016] where, λ 1 is a parameter controlling the sparsity degree, and the larger λ 1 is, the sparser the regression coefficient vector is; λ 2It is a control parameter to prevent overfitting and plays a constraining role; non-zero values in the regression coefficient vector are set to 1 to obtain a binarized MFN matrix; to find the optimal solution of MFN suitable for ASD diagnosis, λ is traversed within a range. 1 , λ 2 values, where λ 1 ={2 -1 , 2 -2 , …, 2 -10}, λ 2 ={0.1, 0.2, …, 1}, select the parameters λ 1 , λ 2 corresponding MFN features to obtain the optimal solution of MFN; the multi-feature network MFN is an asymmetric matrix, where a value of 1 represents a strong relationship between the target brain region and other brain regions, and a value of 0 represents a weak relationship.

[0017] Complete the construction of the feature extraction and construction module.

[0018] (3) Perform feature selection on the morphological features and the features of the relationships between regions of interest within three age groups respectively.

[0019] Since both types of features are high-dimensional, inputting all of them into the classifier may cause data redundancy and affect classification performance. Therefore, a two-step feature selection is performed using the filtering method and the recursive feature elimination method before classification; considering that the differences between autistic patients and normal people are different among different age groups, the subjects are divided into three groups according to age, under 12 years old, 12 - 18 years old, and over 18 years old. Select the features specific to each age group within each age group to form three feature selection modules respectively.

[0020] The steps of feature selection are as follows:

[0021] First, use the two-sample t-test and the chi-square test to screen out the features most relevant to the between-group differences, and exclude the features with a statistical test significance p-value greater than 0.05.

[0022] For the first-step feature selection, for morphological features, use the two-sample t-test to obtain the features with significant differences between the two groups of people for the second-step feature selection; for the multi-feature network MFN, since it is a binarized feature matrix, use the chi-square test to obtain the features with significant differences between the two groups of people for the second-step feature selection.

[0023] After the first feature selection, to further reduce the number of features used for classification and improve the classification effect, the second step of feature selection is carried out. The support vector machine recursive feature elimination algorithm SVM-RFE is used for the second step of feature selection; the SVM-RFE algorithm constructs a sorting coefficient according to the weight vector generated during SVM training, removes the feature with the smallest sorting coefficient in each iteration, and finally obtains the decreasing order sorting of all features, representing their contributions to classification;

[0024] (4) Use multi-kernel learning to separately select weights and features within three age groups according to the results of step (3) and construct three classifiers;

[0025] Suppose X is the input space and H is the Hilbert space. When there is a mapping from X to H, satisfying all x i , x j ∈X, the function is called the kernel function; for the kernel matrix K of the m-th feature, i , x j ), the calculation method is as follows: m

[0026]

[0027] Among them, the linear kernel function is: k(x i , x j ) = x i T x j ;

[0028] The kernel matrix K after fusing multiple features is ∑ m β m K m , where β m m represents the weight of each feature;

[0029] For each feature, use the linear kernel function and adopt nested five-fold cross-validation; search for the optimal parameter C on the training set in each cross-validation process, and the search range is {2 -8 , 2 -7 , …, 2 7 , 2 8}, and finally use this optimal parameter C and the selected features to construct the kernel matrix; when fusing two kernel matrices, the sum of the coefficients is 1, and finally use the kernel matrix fused with the weights corresponding to the highest accuracy to construct the classifier;

[0030] The autism detection module includes the same feature extraction and feature construction modules as the model construction module, a feature selection module for the data to be measured, and a classifier module for the data to be measured;

[0031] The specific steps are as follows:

[0032] Input each subject, and extract the morphological features of the subject and the parameter λ selected based on step (2) 1 , λ 2 , and construct MFN features under λ; through the determination of the subject's age, use one of the three feature selection modules obtained in the model training module to form a feature selection module for the data to be measured; use one of the three classifier modules obtained in the model training module to form a classifier module for the data to be measured, and finally obtain the classification information of the subject according to the classifier.

[0033] Advantages of the present invention: The present invention provides an autism detection device for a specific age group. Compared with the research on all age groups, the present invention aims to use the specific feature information of each age group to more accurately guide the identification of ASD patients; and compared with the existing large-sample research that only uses the information within the region as the classification feature, the present invention adds features representing the information between regions, and uses the multi-kernel learning method to effectively fuse the heterogeneous structure features containing complementary information, so as to achieve a better classification effect. Brief Description of the Drawings

[0034] Figure 1 It is a flowchart of an autism detection device for a specific age group according to an embodiment of the present invention.

[0035] Figure 2 It is a schematic diagram of the effect of using the specific features of each age group in an embodiment of the present invention. (a)-(c) are the between-group difference diagrams between ASD and normal people in three age groups; (a) represents the left-brain difference diagram and the right-brain difference diagram of the group under 12 years old, (b) represents the left-brain difference diagram and the right-brain difference diagram of the group from 12 to 18 years old, (c) represents the left-brain difference diagram and the right-brain difference diagram of the group over 18 years old; (d) is a comparison diagram of the accuracy rates of using the specific features of each age group and using the same features for all age groups; (e) is a schematic diagram of the optimal weight corresponding to each age group.

[0036] Figure 3 It is a schematic diagram of the algorithm framework of an autism detection device for a specific age group according to an embodiment of the present invention. Detailed Embodiment

[0037] In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] Figure 1 This is the flowchart of the autism detection device for a specific age group according to the embodiments of the present invention. As Figure 1 shown, the embodiments of the present invention provide an autism detection device for a specific age group, including a model training module and an autism detection module;

[0039] Among them, the model training module is trained based on the publicly available database ABIDE II with large samples and multiple centers; Step 101, extract the morphological features of the structural images within the regions of interest and the features of the relationships between the regions of interest; Step 102, use the filtering method and the recursive elimination method to perform feature selection on the two types of features within three age groups respectively, and obtain the specific features within each age group to guide the construction of the classifier; Step 103, use multi-kernel learning to fuse the two types of features, and determine the combined weights of the two types of features within three age groups respectively to construct the optimal classifier.

[0040] In the embodiments of the present invention, both feature selection and the construction of the classifier are performed separately within three age groups. Figure 2 (a)- Figure 2 (c) Taking cortical thickness as an example, it shows the differences in the inter-group difference features between two groups of people in three age groups, reflecting the importance of selecting specific features within each age group; Figure 2 (d) is the comparison of the model accuracies of constructing a classifier using the specific features of each age group and constructing a classifier using the features of all age groups. The results show that constructing a classifier using the specific features within each age group can achieve better results; when detecting ASD patients, compared with using the same classifier across all age groups, first selecting a specific classifier according to the patient's age can achieve more accurate identification, because the age-specific classifier is constructed from the specific inter-group difference features within each age group, which is more different and discriminative compared with the inter-group difference features of all age groups, so more accurate results can be obtained; Figure 2 (e) is the optimal weight allocation strategy for the two types of features within each age group, and the results show that the optimal weight allocation corresponding to each age group is different.

[0041] In step 101 of the model training module, features are extracted. Use a template with 68 regions of interest to extract 5 morphological features of structural images, and form the feature matrix X = [x 1 ,x 2 ,…,x n T ∈R n×d ​, n represents the number of brain regions, d represents the feature dimension, and the feature matrix is 68×5. Then, a multi-feature network between regions is obtained through a multiple regression process, that is, the feature vectors of the selected brain regions and the feature vectors of the remaining brain regions are used as the target variable and the predictor variable in the regression respectively. For example, in the i-th regression process, x i is used as the target vector, and the remaining n - 1 regions form the input matrix A of the regression. Note that the column of x i in A is set to 0. Define the linear regression model y = Aw, where y is the target vector x i , w is the regression coefficient vector we need, and finally a multi-feature network representing the relationship between regions is obtained.

[0042] Step 102 performs feature selection in two ways. The first step of feature selection uses the filtering method. For the morphological features within the region, the two-sample t-test is used; for the binary multi-feature network, the chi-square test is used. After completing the first step of feature selection, SVM-RFE is used for the second step of feature selection.

[0043] Step 103 calculates the kernel matrices using the features selected in Step 102 respectively, linearly combines the two kernel matrices into a new kernel matrix, and constructs a classifier using the fused kernel matrix.

[0044] In the autism detection module, for each subject, the features of the subject are first extracted, and through the determination of the subject's age, the age-specific features and weights determined within the model construction module are selected to construct a kernel matrix and input it into the classifier, and finally the classification information of the subject is obtained.

[0045] Figure 3 is an overall framework diagram of the algorithm for an autism detection device for a specific age group. After inputting the image, first, two types of structural image features are extracted, and then the age-specific features and optimal weights selected from the large sample data are used for the fusion of the kernel matrix. The kernel matrix is input into the trained support vector machine model to obtain the diagnostic group of the subject.

[0046] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. An autism detection device for a specific age group, characterized in that, the autism detection device for a specific age group includes a model training module and an autism detection module; The model training module is trained based on the large-sample multi-center database ABIDE II, and includes a feature extraction and construction module, three feature selection modules, and three classifier modules; The specific steps are as follows: Use the processing flow provided by Freesurfer to preprocess the brain regions, and use the Desikan-Killiany parcellation template to divide the brain regions in the individual space in detail. The Desikan-Killiany parcellation template divides the cortical surfaces of the bilateral cerebral hemispheres into 68 brain regions, which are used as regions of interest; (1) Extract the morphological features of the structural images within the regions of interest; Extract the morphological features within each region of interest, including cortical thickness, cortical surface area, gray matter volume, cortical folding index, and mean curvature; (2) Construct features representing the relationships between the regions of interest; Use the elastic net to quantify the relationships between the target brain region and the remaining multiple brain regions; use the 5 morphological features extracted within the regions of interest in step (1) to form the feature matrix of each subject in the database ABIDE II, and use the feature matrix to construct a multi-feature network MFN representing the relationship between the target brain region and the remaining brain regions; the multi-feature network MFN is obtained through a multiple regression process. The feature vectors of the selected target brain region and the remaining brain regions are used as the target variable and the predictive variable in the regression respectively. Define the linear regression model y = Aw, where y is the target vector, A is the feature matrix composed of the remaining brain regions, and w is the regression coefficient vector; obtain the sparse solution by solving the regularized optimization problem, and the formula is: Among them, λ 1 is a parameter for controlling the sparsity level. The larger the value of λ 1 , the sparser the regression coefficient vector; λ 2 is a control parameter for preventing overfitting and plays a constraining role; the non-zero values in the regression coefficient vector are set to 1 to obtain a binarized MFN matrix; λ 1 is traversed within a range, where λ 2 ={2 1 , 2 -1 , …, 2 -2 , …, 2 -10}, and λ 2 ={0.1, 0.2, …, 1}. The MFN features corresponding to the parameters λ 1 and λ 2 with the highest classification accuracy are selected to obtain the optimal solution of MFN; the multi-feature network MFN is an asymmetric matrix, where a value of 1 represents a strong relationship between the target brain region and other brain regions, and a value of 0 represents a weak relationship; Complete the construction of the feature extraction and construction module; (3) Perform feature selection on the morphological features and the features representing the relationships between the regions of interest within three age groups respectively; Use the filtering method and the recursive feature elimination method for two-step feature selection; divide the subjects into three groups according to age, under 12 years old, 12 - 18 years old, and over 18 years old, and select the age-specific features within each age group to form three feature selection modules respectively; The steps of feature selection are as follows: First, use the two-sample t-test and the chi-square test to screen out the features most relevant to the between-group differences, and exclude the features with a statistical test significance p-value greater than 0.05; For the first-step feature selection, for the morphological features, use the two-sample t-test to obtain the features with significant differences between the two groups of people for the second-step feature selection; for the multi-feature network MFN, since it is a binary feature matrix, use the chi-square test to obtain the features with significant differences between the two groups of people for the second-step feature selection; After the first feature selection, to further reduce the number of features used for classification and improve the classification effect, the second step of feature selection is carried out. The support vector machine recursive feature elimination algorithm SVM-RFE is used for the second step of feature selection. The SVM-RFE algorithm constructs a sorting coefficient based on the weight vector generated during SVM training. In each iteration, the feature with the smallest sorting coefficient is removed, and finally, a decreasing order sorting of all features is obtained, representing their contributions to classification. (4) Use multi-core learning to separately select weights and features within three age groups according to the results of step (3) and construct three classifiers. Let \(X\) be the input space and \(H\) be the Hilbert space. When there exists a mapping from \(X\) to \(H\), \(X\rightarrow H\) such that for all \(x\) i , \(x\) j \(\in X\), the function is called the kernel function \(k(x\) i , \(x\) j ). For the kernel matrix \(K\) m of the \(m\)-th feature, the calculation method is as follows: Among them, the linear kernel function is: k(x i , x j ) = x i T x j ; The kernel matrix K after fusing multiple features is K = ∑ m β m K m , where β m represents the weight of each feature; For each feature, a linear kernel function is used, and nested five-fold cross-validation is adopted; the optimal parameter C is searched on the training set in each cross-validation process, and the search range is {2 -8 , 2 -7 , …, 2 7 , 2 8}, and finally, the kernel matrix is constructed using the optimal parameter C and the selected features; when fusing two kernel matrices, the sum of the coefficients is 1, and finally, a classifier is constructed using the kernel matrix fused with the weights corresponding to the highest accuracy rate; The autism detection module includes the same feature extraction and feature construction modules as the model construction module, a feature selection module for the data to be tested, and a classifier module for the data to be tested. The specific steps are as follows: Input each subject, and extract the morphological features of the subject and the parameter λ selected based on step (2). 1 , λ 2 , construct the MFN features under it; through the determination of the subject's age, use one of the three feature selection modules obtained in the model training module to form a feature selection module for the data to be measured; use one of the three classifier modules obtained in the model training module to form a classifier module for the data to be measured, and finally obtain the classification information of the subject according to the classifier.

Images