CN110443798B - Autism detection method, device and system based on magnetic resonance image - Google Patents

Abstract

The invention discloses a method, a device and a system for detecting autism based on graph theory, and belongs to the technical field of medical image processing. Firstly, constructing a brain function network topological graph corresponding to each magnetic resonance image based on the lowest network consumption threshold value set by the invention as a binarization threshold value, determining abnormal brain regions based on the difference statistical information of node degree information of autism patients and normal people, and extracting characteristic information of the abnormal brain regions, namely node degree information of nodes corresponding to the brain regions; model training and detection of the autism binary classifier are carried out based on the extracted characteristic information, and auxiliary diagnosis assistance is provided for autism diagnosis. The multi-feature fusion mode autism detection system provides another auxiliary diagnosis aid for autism diagnosis, and has a certain improvement on the detection recognition rate of autism; the detection system can also be used for automatic detection of smoking addiction, network addiction, online game addiction and other health fields.

Description

Autism detection method, device and system based on magnetic resonance image

Technical Field

The invention belongs to the technical field of medical image processing, and particularly relates to a classification processing technology for autism based on processing of magnetic resonance images.

Background

Autism is a relatively common mental disorder, which is a subtype of pervasive developmental disorder, and the symptoms of autism are mainly represented by different degrees of speech developmental disorder, interpersonal disorder, interest stenosis, behavioral pattern inscription, intelligent disorder and the like. The incidence rate of the disease in China is not rare, but so far, no good method for diagnosing autism exists. The current diagnosis of autism is mostly performed by a doctor checking the usual growth and development history, medical history and spirit of children, and diagnosing the disease with reference to the currently internationally recognized autism diagnosis standard DSM-5. This is particularly demanding for the clinical experience of the doctor, and the diagnosis criteria are also controlled by the doctor, which also hinders the diagnosis of autism, so it is necessary to provide a technique for classifying autism based on the processing of magnetic resonance images in order to achieve auxiliary diagnosis.

Disclosure of Invention

The invention aims at: in order to solve the above-described problems, a method for detecting autism based on magnetic resonance images is provided.

The detection method of the invention adopts the following technical scheme:

collecting a training sample data set:

acquiring a certain number of resting state functional magnetic resonance data of normal people and a certain number of resting state functional magnetic resonance data of autism patients through a nuclear magnetic resonance scanner, so as to obtain two groups of magnetic resonance images, wherein one group is the magnetic resonance image of the autism patients, and the other group is the magnetic resonance image of the normal people;

the training sample data set is subjected to a data preprocessing step:

taking the position of the front link in each magnetic resonance image as the space coordinate origin of the front link, and performing head motion correction processing;

image segmentation is carried out on the image after the head movement correction treatment, and the image is segmented into three parts of brain gray matter, white matter and cerebrospinal fluid;

carrying out spatial normalization processing on the split brain gray image, and then carrying out spatial position correction processing;

performing spatial smoothing on the grey brain image subjected to spatial correction;

the method comprises the following steps of performing brain function network construction on each gray image in a training sample data set after data processing:

dividing each gray matter image into a plurality of brain regions based on a brain region template selected from a brain region standard template library;

calculating time sequence correlation coefficients of each brain region by utilizing the time sequence, so as to obtain a correlation coefficient matrix of each grey brain image on the brain region;

searching a threshold with the maximum network efficiency value in a binarization threshold range [0.05,0.4] based on a preset step length to serve as an optimal binarization threshold;

performing binarization processing on the correlation coefficient matrix based on the optimal binarization threshold value to obtain a binary matrix of the brain gray image about a brain region: if the correlation coefficient is greater than the optimal binarization threshold, the binarization result is 1; otherwise, 0;

taking each brain region as a node, and obtaining a brain function network topology graph of each brain gray image based on a binary matrix;

determining abnormal brain areas:

counting brain function network topology diagrams of normal people and autism patients based on node degrees of all nodes, and defining brain areas with difference exceeding a preset threshold as abnormal brain areas;

training a classifier:

performing feature extraction processing on each training sample in the training sample data set, performing model training processing on a preset binary classifier model for autism detection based on the extracted features, and stopping training when preset detection precision is met to obtain a trained classifier;

the characteristic information of each training sample is as follows: node degree information of abnormal brain areas;

a detection processing step of an object to be detected:

collecting resting state functional magnetic resonance data of an object to be detected through a nuclear magnetic resonance scanner to obtain a magnetic resonance image to be detected;

after data preprocessing and brain function network construction processing are carried out on the magnetic resonance image to be detected, a brain function network topology diagram of the magnetic resonance image to be detected is obtained, and node degree information of an abnormal brain region is extracted to serve as characteristic information of an object to be detected;

and inputting the characteristic information of the object to be detected into the trained classifier to obtain a detection result.

Meanwhile, the invention also provides a detection device based on the detection method, which comprises a data receiving and preprocessing module, a brain function network construction module, a feature extraction module, an abnormal brain region determination module, a classifier training module and a classification detection module;

the data receiving and preprocessing module is used for receiving the training sample and the magnetic resonance image of the object to be detected, preprocessing the received magnetic resonance image, and then sending the preprocessed magnetic resonance image to the brain function network construction module; wherein the training sample comprises magnetic resonance images of normal people and autism patients;

the brain function network construction module constructs a brain function network topological graph corresponding to each magnetic resonance image and sends the brain function network topological graph to the feature extraction module and the abnormal brain region determination module;

the abnormal brain region determining module is used for counting the brain function network topological graph of the normal person and the autism patient based on the node degree of each node of the brain function network topological graph, and defining the brain region with the difference exceeding a preset threshold value as an abnormal brain region; transmitting the abnormal brain region determination result to a feature extraction module;

the feature extraction module is used for extracting node degree information of abnormal brain regions in the brain function network topological graph, sending the feature information of the extracted training sample to the classifier training module and sending the feature information of the object to be detected to the classification detection module;

the classifier training module is used for receiving the binary classification model set by the user, training the classifier for detecting the autism of the binary classification model based on the characteristic information of the input training sample, obtaining a trained autism classifier when the classification detection precision meets the precision requirement set by the user, and sending the trained autism classifier to the classification detection module;

and the classifier detection module is used for carrying out the binary classification detection of the autism based on the trained autism classifier and the input characteristic information of the object to be detected and outputting the detection result of the object to be detected.

Meanwhile, the invention also provides an autism detection system in another mode;

the system comprises a data receiving and preprocessing module, a feature extracting and processing module, an SVM prediction model and a detection output module;

the data receiving and preprocessing module is used for receiving the magnetic resonance image, preprocessing the magnetic resonance image and then sending the magnetic resonance image to the characteristic extracting and processing module;

the feature extraction and processing module is used for extracting various feature information of the brain region, performing dimension reduction processing on the corresponding features based on a preset dimension reduction mode, performing multi-feature fusion on all the features to obtain detection feature vectors, and inputting the detection feature vectors into the SVM prediction model;

wherein the extracted feature information includes:

the node degree information of the abnormal brain region is not subjected to characteristic dimension reduction processing;

gray matter and/or white matter and/or cerebrospinal fluid information of each brain region, including volume, density and the like, wherein the dimension reduction mode is linear discriminant analysis LDA;

network attributes of the brain function network, including characteristic path length, clustering coefficient, small world attribute, global efficiency, layering degree, network synchronism and the like, wherein the set dimension reduction mode is principal component analysis PCA;

fALFF/ReHo signals of abnormal brain areas are subjected to PCA (principal component analysis) in a set dimension reduction mode;

VMHC values of abnormal brain regions, and the set dimension reduction mode is principal component analysis PCA;

the correlation properties of the functional network matrix comprise a triangular part on the functional network weight matrix, the rank of the matrix, determinant, characteristic values and the like, and the set dimension reduction mode is principal component analysis PCA;

the abnormal brain region is determined based on experience information judgment, for example, if the corresponding characteristic information of the current brain region reaches a preset threshold value, the current brain region is considered to be the abnormal brain region.

The SVM prediction model is offline training, corresponding training samples are acquired from a brain image database, various characteristic information of the training samples is extracted according to the characteristic information extracted by the characteristic extraction and processing module, the corresponding characteristics are subjected to dimension reduction processing based on a preset dimension reduction mode (the same as the setting of the characteristic extraction and processing module), and then all the characteristics are subjected to multi-characteristic fusion to obtain characteristic vectors of all the training samples, and model training is performed on the SVM prediction model based on the characteristic vectors of the training samples until the preset recognition precision or training requirement is met;

the SVM prediction model obtains an autism detection result of the object to be detected based on the input detection feature vector and sends the autism detection result to the detection output module;

and the detection output module outputs and displays the received autism detection result.

Further, the setting mode of the extracted characteristic information specifically includes:

selecting a quantity of initial feature sets;

setting a corresponding SVM classifier for each type of feature in the initial feature set;

based on the training of the corresponding SVM classifier by extracting various feature vectors from the brain image database, based on the trained SVM classification, the recognition accuracy corresponding to various features is counted, and the first K features with the highest recognition accuracy are selected as the extracted feature information.

In summary, due to the adoption of the technical scheme, the beneficial effects of the invention are as follows: according to the invention, firstly, a brain function network topological graph corresponding to each magnetic resonance image is constructed based on the binary threshold value set by the invention, then, an abnormal brain region is determined based on the difference statistical information of node degree information of an autism patient and a normal person, and only the characteristic information of the abnormal brain region is extracted for training and detecting, so that the dimension reduction processing of the characteristics is realized, and the training and the detection of a binary classification model are carried out by using the characteristic information of the node degree selection brain function network, thereby providing auxiliary diagnosis assistance for the autism diagnosis.

Meanwhile, the invention also discloses a multi-feature fusion mode autism detection system, which provides another auxiliary diagnosis aid for autism diagnosis, and the fusion multi-feature mode can improve the detection recognition rate of autism to a certain extent; the detection system can also be used for automatically detecting smoking addiction, network addiction, online game addiction and the like; health fields such as cognition, and other aspects. The method can be realized only by adjusting the output result when the SVM prediction model of the autism detection system is subjected to model training.

Drawings

FIG. 1 is a graph of ASD (autism) versus HC (normal) functional network normalized cluster coefficients;

FIG. 2 is a normalized characteristic path length comparison of ASD and HC functional networks;

FIG. 3 is a comparison of ASD and HC functional network small world attributes;

FIG. 4 is a global efficiency comparison of an ASD and HC functional network, where NCE represents the cost efficiency of network construction;

fig. 5 is a comparison of ASD versus HC functional network local efficiency.

Detailed Description

The present invention will be described in further detail with reference to the embodiments and the accompanying drawings, for the purpose of making the objects, technical solutions and advantages of the present invention more apparent.

According to the invention, through the magnetic resonance imaging technology and theoretical knowledge of graph theory, the specific characteristic information of the brain function network corresponding to the acquired magnetic resonance image is analyzed and judged, and based on the trained classification model, the binary classification processing of the magnetic resonance image on whether the autism exists is realized, so that the auxiliary diagnosis of the autism by doctors can be helped.

The self-disorder detection method based on graph theory comprises the following steps: data acquisition, data preprocessing, network construction and network detection.

The data acquisition comprises data acquisition of training samples and data acquisition of the current object to be detected, namely, acquiring resting state functional magnetic resonance data (magnetic resonance images) of a certain number of normal people and resting state functional magnetic resonance data of a corresponding number of autism patients through a nuclear magnetic resonance scanner, and taking the resting state functional magnetic resonance data as corresponding training samples; and acquiring resting state functional magnetic resonance data of the current object to be detected through a nuclear magnetic resonance scanner.

And then carrying out corresponding data preprocessing on the acquired magnetic resonance image to reduce the interference of external factors on training and detection processing, wherein the data preprocessing process comprises the following steps:

(1) The spatial origin is manually set, namely, the spatial origin is set at the front joint position, and relevant head movement correction is carried out, so that image data with larger head movement (for example, the head movement is more than 2mm in translation and 2 degrees in rotation) are eliminated;

(2) And (3) image segmentation is carried out on all acquired data, and the acquired data are segmented into three parts of brain gray matter, white matter and cerebrospinal fluid. The subsequent processing mainly aims at the segmented grey brain matter image;

(3) Carrying out space standardization treatment on the segmented brain gray image, wherein MNI (Montreal Neurological Institute) space is adopted in the specific embodiment, namely, the brain gray image is registered to MNI space again;

(4) Performing correction processing for errors occurring in spatial normalization;

(5) In order to reduce the existence of image noise, the grey brain image after error correction is subjected to spatial smoothing. Wherein the preferable mode is as follows: the 8mm gaussian kernel was spatially smoothed.

The network construction and the functional network construction are to construct by utilizing the brain gray matter images obtained by segmentation, divide the whole functional image into 264 brain areas by using Power templates in a brain area standard template library and calculate time sequence correlation coefficients of each brain area by utilizing a time sequence, thereby obtaining a correlation coefficient matrix of each brain gray matter image. In order to reduce the excessive number of edges in the network, the correlation coefficient matrix is transformed into a binary matrix through thresholding, namely a threshold value T is set, if the value in the correlation coefficient matrix is larger than the threshold value T, the value is set to be 1, otherwise, the value is set to be 0, so that the functional matrix is converted into the binary matrix, and the change of the attribute of the whole network is described through topology analysis on the binary matrix. In the specific embodiment, the selected binarization threshold range is 0.05-0.4, the step length is 0.01, the network topology results under 36 thresholds are calculated, and the binarization threshold with the best effect is selected from the network topology results, so that the detection accuracy is improved.

After the functional network is constructed, the corresponding characteristic information needs to be extracted so as to facilitate the realization and the learning of the two classification processing, mainly extracting the network topology attribute information of the constructed functional network, including small world attributes, cluster coefficients, characteristic path length, global efficiency, node degree, degree and the like.

In order to extract the characteristic information with better classification detection effect, the invention firstly compares the difference of two groups of data (autism and normal) on the global network topology structure, and carries out double-sample T test on the functional image data results of the autism and normal control groups under different thresholds, the results show that the functional image networks of the autism patient and the normal control group show the characteristics of a small world network, no significant difference is found (p > = 0.4157), and the two groups of data have higher clustering coefficient (p > =0.366) and lower characteristic path length (p > =0.312), the two groups of data show no difference (p > = 0.1401) on the global efficiency of the network, and the specific results are shown in fig. 1-5.

That is, the global network attributes such as small world attributes, cluster coefficients, characteristic path lengths and the like in the functional network of the autism patient are not significantly different from those in the normal control group. Therefore, in order to extract more remarkable characteristic information, the invention performs local network analysis by constructing a threshold value with the lowest network consumption as a research object. To describe local network changes in an autism-functional network, an appropriate threshold needs to be set to represent the network change over the entire threshold range of the network. To select this reasonable threshold, the processing mode of the invention is as follows: and searching the maximum cost efficiency value required by constructing the network in the whole threshold range by using the global efficiency as a reference basis, and taking the value as the threshold of the local network. Where network efficiency is the inverse of the harmonic mean of the shortest paths between all nodes in the network. In a network, a higher efficiency value indicates a lower cost for information exchange in the network. In this embodiment, the calculation method of the threshold with the lowest network consumption is as follows: in the whole threshold range, the difference processing between the network efficiency obtained under each threshold and the current threshold is calculated according to the preset step length, the maximum value of the calculated result is the required value, as shown in fig. 4, in the whole network construction process, the value is larger than 0, which indicates that the network construction between the autism and the normal control group meets the economic and economic attribute, and it can be seen that in the whole network, the cost efficiency value (NCE shown in fig. 4) of the whole network construction is the maximum at the 14 th threshold selection point, namely, when the threshold is 0.18, so that the threshold selected in the specific embodiment is 0.18. The correlation coefficient matrix is binarized based on a threshold value of 0.18, and then the double-sample T test is performed on the node degree under the threshold value based on the definition of the node degree (the number of edges directly connected with the node in the network), so as to compare the difference between the autism and the control group.

For node degree, firstly, for a group average network between two groups, a double-sample T test is carried out, the average network of the two groups of functional images is found to have no significance change (P > 0.05), but some significance differences exist between the two groups of data on node degree attributes, and a total of 35 brain regions are found to be abnormal, wherein the node degree of 18 brain regions in an autism patient is found to be obviously higher than that of a normal person, and the node degree of 17 brain regions is found to be obviously lower than that of a normal person. Specific results are shown in table 1 and table 2.

TABLE 1 functional image node degree comparison for autism (autism < control group)

TABLE 2 comparison of functional image node degree for autism (autism > control group)

In tables 1 and 2, the T value is the test statistic of the double sample T test and the P value is the corresponding significance value.

Meanwhile, by analyzing the node degree of the functional network, abnormal brain area distribution of the node degree of the autism patient can be clearly seen, wherein partial abnormal brain areas are consistent with differences in the brain local activities analyzed before, and the method provides help for the clearer understanding of the brain functional network of the autism. The orbitum region in the region with the remarkably increased appearance plays an important role in processing facial nerves and expressions, which indicates that the brain region of the autism patient in charge of emotion processing and processing is abnormal, and the other important brain region for emotion processing is a default network, and the specific brain region comprises regions of temporal gyrus, medial frontal gyrus and the like, and the abnormal appears in the regions, so that deviation can appear in the aspect of processing emotion information or emotion expression for the autism patient. In brain areas where node degree is reduced, top back plays an important role in handling visual information reception and transmission, which belongs to a dorsal attention network, node degree of the area is abnormal, indicating that abnormal connection of the area occurs in a functional network, which causes a patient suffering from autism to be hardly focused, and emotion expression is seriously affected.

The brain regions with significantly higher node degree than normal in autism patients are mainly on AAL (Anatomical Automatic Labeling) templates: the brain areas such as the infraorbital frontal center, rectus, infraorbital infrafrontal inferior, anterior wedge, lateral dorsally frontal superior, anterior cingulate and lateral cingulate cerebral palsy, infraorbital frontal superior, temporal center, occipital center, lingual center, occipital center, medial frontal superior, temporal inferior, and apical superior have significantly reduced node degree compared to normal, and the brain areas with autism patients have mainly on AAL templates: central posterior, temporal superior, central capping, supplementary motor regions, angular, frontal, thalamus, overhead, anterior wedge, temporal, etc. brain regions.

It can be seen that there are coincident brain region names in the two groups, that is, there are increases and decreases in the node degree of partial brain regions, because the brain region template used in this embodiment is 264 templates of Power, so in order to be convenient to correspond to the commonly used brain region namespace, the brain region template used in the naming of brain regions is an AAL template, and the brain regions are divided into 116 brain regions, that is, by using the MNI position to correspond to the brain region names, which is smaller than the divided brain regions, so that it is illustrated from a certain angle that some finer brain region differences can be found further by using finer brain region division, and also that a smaller scale study is required for the study of the brain to find more finer differences.

The method comprises the steps of performing model training through training samples based on a selected classification model by using characteristics selected by node degree (the number of edges directly connected with the node in a network), obtaining a classification classifier for autism, and inputting corresponding characteristic information of data to be detected into the trained classifier to obtain a classification result of a current object to be detected. In the specific embodiment, a plurality of classification models are used for model training, so that a better classifier is obtained, help is provided for autism diagnosis, and the result shows that the SVM (support vector machine) is used for disease diagnosis, so that the accuracy of the result can reach 81.87%.

TABLE 3 diagnosis of autism

Classification model	Accuracy (%)
		SVM	81.2672
Logistic regression	79.3468
		Decision tree	72.1245
Random forest	72.6573

The specific implementation process of the autism detection system is as follows:

setting an initial feature set, comprising:

1) Information derived from brain structure images, including gray matter/white matter/cerebrospinal fluid volume/density etc. of each brain region;

2) Network attributes of the brain function network, including characteristic path length, clustering coefficient, small world attribute, global efficiency, layering degree, network synchronism and the like;

3) fALFF/ReHo signals of abnormal brain areas in resting brain signal analysis and the like;

4) VMHC values of abnormal brain regions in brain symmetry analysis;

5) Clustering coefficients of brain cause effect network, etc.;

6) The correlation properties of the functional network matrix comprise triangle parts on the functional network weight matrix, the rank, determinant, eigenvalue and the like of the matrix.

Wherein the fALFF signal represents an amplitude analysis of the scoring low frequency, i.e. the energy of the calculated low frequency band signal is divided by the power of the whole frequency band on the basis of ALFF (a measure of the energy of the low frequency signal); the ReHo signal refers to local consistency and is used for measuring indexes of synchronicity and relativity of spontaneous activities of the neurons; VMHC refers to symmetric voxel homotopy.

The 6 characteristic information training models are adopted respectively, and the fitting effect of the model on the training set is tested, so that whether the 6 characteristic information is selected to have certain rationality and reliability is verified. The specific method comprises the following steps:

a training sample set usable for single-item detection of the initial feature is downloaded from the brain image database, in this embodiment from the second batch of data abidi in the ABIDE database (website http:// fcon 1000. Subjects. Nitrc. Org/indi/adibe) in the fcon 1000 project. The database is provided with a plurality of autism data sets, and the original data of the data sets comprise two types of brain structure images and resting brain function images, wherein the two types of the data sets comprise 57 autism patients (ASD) and 156 normal control groups (HC).

Extracting all the initial characteristics to be tested, training the SVM model, wherein the iteration number is 5, using default parameters for the rest, and then calculating the fitting condition of each model on a training set. The evaluation index uses Accuracy (Accuracy, i.e., the proportion of samples correctly classified), and the specific results are shown in table 4.

TABLE 4 fitting Effect of models trained using initial feature sets

Feature set	SVM accuracy
		Volume, density, etc. of ash	0.5138
Brain function network attributes	0.7248
		fALFF/ReHo signal of abnormal brain region	0.5688
VMHC value of abnormal brain region	0.6789
		Node degree value of abnormal brain region	0.7156
Clustering coefficients of brain cause effect network	0.6239
		Functional network matrix correlation properties	0.6697

When the data volume is small, the situation that the feature number is far larger than the sample number easily occurs, and at the moment, the model such as SVM, decision tree and the like easily has the over-fitting phenomenon. The dimension statistics for each feature set in the experiment and the results of training the SVM using five-fold cross-validation are shown in table 5.

TABLE 5 dimension and SVM accuracy contrast for different feature sets

It can also be seen from table 5 that the feature set with the larger dimension does not perform well on SVM, while the feature set accuracy is relatively high for several groups with feature numbers less than 100. In the case that the number of samples cannot be increased, the generalization performance of the model can be improved by means of data dimension reduction, feature selection and the like.

In this embodiment, two data dimension reduction modes are used for comparison, namely PCA (Principal components analysis) and LDA (Linear Discriminant Analysis), and are specifically shown in table 6.

Table 6 comparison of accuracy of SVM before and after PCA and LDA dimension reduction

In order to improve the generalization performance of the model and improve the model training prediction speed, feature selection is another method. The simplest of these is to use variance for screening. In particular, if the variance of the value of a feature is 0, where its values are all the same, then the contribution of this feature to the classification is obviously also 0. But this does not mean that the larger the variance, the greater the contribution to classification performance. For example: the method comprises the steps of having two positive samples and two negative samples, wherein the values of the two positive samples and the two negative samples on certain two characteristics are respectively F1= [0.1;0.2;0.8;0.9], f2= [1;10;2;9] it is possible to calculate that the variance of F1 is much smaller than that of F2, but in practice F1 can more simply separate positive and negative samples. Thus combining the actual scenarios. The characteristic selection method mainly comprises a filtering method, a packaging method and an embedding method.

In this embodiment, the filtering method and the packing method, specifically, the card method and the recursive elimination method (RFE) are used for feature selection, and the results are shown in table 7.

Table 7 SVM accuracy contrast after chi-square test and RFE feature selection

It can be seen from table 7 that the features such as gray matter volume and density, brain function network properties, etc. are not suitable for both methods because the features that contribute significantly to classification are not concentrated on a few features, though the dimensions are large; the fALFF/ReHo signals of abnormal brain regions have good effect after the VMHC values are subjected to feature selection; the correlation property of the functional network matrix is sparse, so that no obvious effect is achieved after feature selection is performed.

Therefore, the invention firstly carries out corresponding dimension reduction treatment (shown in table 8) on selected characteristics (brain gray matter information (including volume, density and the like), brain function network attributes, fALFF/ReHo signals of abnormal brain regions, VMHC values of abnormal brain regions and function network matrix related properties), the total dimension of the characteristics is reduced by 95 percent, and L1 regularization is added in the subsequent model training, so as to further control the dimension of the characteristics.

TABLE 8 dimension reduction processing of partial feature sets

Feature set	Feature dimension	Post-processing dimension	Treatment method
				Volume, density, etc. of ash	630	1	LDA
Brain function network attributes	216	40	PCA
				fALFF/ReHo signal of abnormal brain region	96	30	RFE
VMHC value of abnormal brain region	48	12	PCA
				Functional network matrix correlation properties	4005	45	PCA

That is, in the present invention, in the multi-feature mode, the extracted feature information includes:

the node degree information of the abnormal brain region is not subjected to characteristic dimension reduction processing;

gray matter and/or white matter and/or cerebrospinal fluid information of each brain region is set in a dimension reduction mode of linear discriminant analysis LDA;

the network attribute of the brain function network, the set dimension reduction mode is PCA (principal component analysis);

fALFF/ReHo signals of abnormal brain areas are subjected to PCA (principal component analysis) in a set dimension reduction mode;

VMHC values of abnormal brain regions, and the set dimension reduction mode is principal component analysis PCA;

and the related property of the functional network matrix is that the set dimension reduction mode is principal component analysis PCA.

In the specific embodiment, in model training, a five-fold cross validation method is adopted to train a support vector machine. Wherein the SVM uses a polynomial kernel function, up to 4 degrees. In addition, the sample category distribution is slightly unbalanced, so category weights are set in the model. The final results are shown in Table 9.

TABLE 9 results of 5-fold Cross validation and model fusion for each model

According to the invention, the SMV model is adopted to carry out classification training on the initial feature set, the fitting performance of the models trained by a plurality of groups of initial feature subsets on the training set is tested, and the average accuracy rate of the results can reach 70%. The accuracy rate of the model fusion trained by the multiple features reaches 80.73%, the specificity reaches 84.52%, and the accuracy rate exceeds that of a single feature subset or performance.

While the invention has been described in terms of specific embodiments, any feature disclosed in this specification may be replaced by alternative features serving the equivalent or similar purpose, unless expressly stated otherwise; all of the features disclosed, or all of the steps in a method or process, except for mutually exclusive features and/or steps, may be combined in any manner.

Claims (6)

1. The method for detecting the autism based on the magnetic resonance image is characterized by comprising the following steps of:

collecting a training sample data set:

acquiring a certain number of resting state functional magnetic resonance data of normal people and a certain number of resting state functional magnetic resonance data of autism patients through a nuclear magnetic resonance scanner, so as to obtain two groups of magnetic resonance images, wherein one group is the magnetic resonance image of the autism patients, and the other group is the magnetic resonance image of the normal people;

the training sample data set is subjected to a data preprocessing step:

taking the position of the front link in each magnetic resonance image as the space coordinate origin of the front link, and performing head motion correction processing;

image segmentation is carried out on the image after the head movement correction treatment, and the image is segmented into three parts of brain gray matter, white matter and cerebrospinal fluid;

carrying out spatial normalization processing on the split brain gray image, and then carrying out spatial position correction processing;

performing spatial smoothing on the grey brain image subjected to spatial correction;

the method comprises the following steps of performing brain function network construction on each gray image in a training sample data set after data processing:

dividing each gray matter image into a plurality of brain regions based on a brain region template selected from a brain region standard template library;

calculating time sequence correlation coefficients of each brain region by utilizing the time sequence, so as to obtain a correlation coefficient matrix of each grey brain image on the brain region;

searching a threshold with the maximum network efficiency value in a binarization threshold range [0.05,0.4] based on a preset step length to serve as an optimal binarization threshold;

performing binarization processing on the correlation coefficient matrix based on the optimal binarization threshold value to obtain a binary matrix of the brain gray image about a brain region: if the correlation coefficient is greater than the optimal binarization threshold, the binarization result is 1; otherwise, 0;

taking each brain region as a node, and obtaining a brain function network topology graph of each brain gray image based on a binary matrix;

determining abnormal brain areas:

counting brain function network topology diagrams of normal people and autism patients based on node degrees of all nodes, and defining brain areas with difference exceeding a preset threshold as abnormal brain areas;

training a classifier:

performing feature extraction processing on each training sample in the training sample data set, performing model training processing on a preset SVM-based binary classifier model for autism detection based on the extracted features, and stopping training when preset detection precision is met to obtain a trained classifier;

wherein, the characteristic information of the training sample includes:

the node degree information of the abnormal brain region is not subjected to characteristic dimension reduction processing;

gray matter and/or white matter and/or cerebrospinal fluid information of each brain region is set in a dimension reduction mode of linear discriminant analysis LDA;

the method comprises the steps of (1) analyzing PCA by using a set dimension reduction mode as a main component, wherein the network attribute of the brain function network comprises layering degree and network synchronism;

fALFF/ReHo signals of abnormal brain areas are subjected to PCA (principal component analysis) in a set dimension reduction mode;

VMHC values of abnormal brain regions, and the set dimension reduction mode is principal component analysis PCA;

the method comprises the steps of (1) analyzing PCA (principal component analysis) by using a set dimension reduction mode as a principal component, wherein the functional network matrix correlation property comprises a matrix rank, a matrix determinant and a characteristic value;

a detection processing step of an object to be detected:

collecting resting state functional magnetic resonance data of an object to be detected through a nuclear magnetic resonance scanner to obtain a magnetic resonance image to be detected;

after data preprocessing and brain function network construction processing are carried out on the magnetic resonance image to be detected, a brain function network topology diagram of the magnetic resonance image to be detected is obtained, and node degree information of an abnormal brain region is extracted to serve as characteristic information of an object to be detected;

and inputting the characteristic information of the object to be detected into the trained classifier to obtain a detection result.

2. The method according to claim 1, wherein the setting of the optimal binarization threshold is in particular: and within the threshold range [0.05,0.4], respectively calculating the difference value between the network efficiency obtained under each threshold and the current threshold according to the preset step length, and taking the threshold corresponding to the maximum difference value as the optimal binarization threshold.

3. The method of claim 1, wherein the optimal binarization threshold is set directly to 0.18.

4. The autism detection device based on the magnetic resonance image comprises a data receiving and preprocessing module, a brain function network construction module, a feature extraction module, an abnormal brain region determination module, a classifier training module and a classification detection module;

the data receiving and preprocessing module is used for receiving the training sample and the magnetic resonance image of the object to be detected, preprocessing the received magnetic resonance image, and then sending the preprocessed magnetic resonance image to the brain function network construction module; wherein the training sample comprises magnetic resonance images of normal people and autism patients; the data preprocessing mode is the data preprocessing step in the method of claim 1, 2 or 3;

the brain function network construction module is used for constructing brain function network topological graphs corresponding to the magnetic resonance images, sending the brain function network topological graphs of the training samples and the object to be detected to the feature extraction module, and sending the brain function network topological graphs of the training samples to the abnormal brain region determination module; the specific construction mode of the brain function network topological graph is the brain function network construction step in the method of claim 1, 2 or 3;

the abnormal brain region determining module is used for counting the brain function network topological graph of the normal person and the autism patient based on the node degree of each node of the brain function network topological graph, and defining the brain region with the difference exceeding a preset threshold value as an abnormal brain region; transmitting the abnormal brain region determination result to a feature extraction module;

the feature extraction module is used for extracting node degree information of abnormal brain regions in the brain function network topological graph, sending the feature information of the extracted training sample to the classifier training module and sending the feature information of the object to be detected to the classification detection module;

the classifier training module is used for receiving the binary classification model based on the SVM set by the user, training the classifier for detecting the autism of the binary classification model based on the characteristic information of the input training sample, obtaining a trained autism classifier when the classification detection precision meets the precision requirement set by the user, and sending the trained autism classifier to the classification detection module;

and the classifier detection module is used for carrying out the binary classification detection of the autism based on the trained autism classifier and the input characteristic information of the object to be detected and outputting the detection result of the object to be detected.

5. The autism detection system is characterized by comprising a data receiving and preprocessing module, a characteristic extracting and processing module, an SVM prediction model and a detection output module;

the data receiving and preprocessing module is used for receiving the magnetic resonance image, preprocessing the magnetic resonance image and then sending the magnetic resonance image to the characteristic extracting and processing module;

the feature extraction and processing module is used for extracting various feature information of the brain region, performing dimension reduction processing on the corresponding features based on a preset dimension reduction mode, performing multi-feature fusion on all the features to obtain detection feature vectors, and inputting the detection feature vectors into the SVM prediction model;

wherein the extracted feature information includes:

the node degree information of the abnormal brain region is not subjected to characteristic dimension reduction processing;

gray matter and/or white matter and/or cerebrospinal fluid information of each brain region is set in a dimension reduction mode of linear discriminant analysis LDA;

the method comprises the steps of (1) analyzing PCA by using a set dimension reduction mode as a main component, wherein the network attribute of the brain function network comprises layering degree and network synchronism;

fALFF/ReHo signals of abnormal brain areas are subjected to PCA (principal component analysis) in a set dimension reduction mode;

VMHC values of abnormal brain regions, and the set dimension reduction mode is principal component analysis PCA;

the method comprises the steps of (1) analyzing PCA (principal component analysis) by using a set dimension reduction mode as a principal component, wherein the functional network matrix correlation property comprises a matrix rank, a matrix determinant and a characteristic value;

the SVM prediction model is offline training, corresponding training samples are acquired from a brain image database, various characteristic information of the training samples is extracted according to characteristic information extracted by a characteristic extraction and processing module, after the corresponding characteristics are subjected to dimension reduction processing based on a preset dimension reduction mode, all the characteristics are subjected to multi-characteristic fusion to obtain characteristic vectors of all the training samples, and model training is performed on the SVM prediction model based on the characteristic vectors of the training samples until preset recognition precision or training requirements are met; the preset dimension reduction mode is the same as the setting of the feature extraction and processing module;

the SVM prediction model obtains an autism detection result of the object to be detected based on the input detection feature vector and sends the autism detection result to the detection output module;

and the detection output module outputs and displays the received autism detection result.

6. The system of claim 5, wherein the extracted feature information is specifically configured in the following manner:

selecting a quantity of initial feature sets;

setting a corresponding SVM classifier for each type of feature in the initial feature set;

based on the training of the corresponding SVM classifier by extracting various feature vectors from the brain image database, based on the trained SVM classification, the recognition accuracy corresponding to various features is counted, and the first K features with the highest recognition accuracy are selected as the extracted feature information.

Images