Method and System for Constructing an Autism Classifier Based on Functional Magnetic Resonance Imaging of the Human Brain

Abstract

The present invention discloses a method and system for constructing an autism classifier based on human brain functional magnetic resonance imaging. The brain functional magnetic resonance images of normal subjects and autism subjects are preprocessed to obtain feature vectors. For each feature of the feature vectors, a feature selection method based on the difference in gradient distribution curves is used to obtain significant features. The obtained significant features are combined into a new feature vector, and the new feature vectors of all subjects are used as training samples, and the training samples include a training set and a validation set. The variational autoencoder is pre-trained using the training set. After the pre-training is completed, the encoder parameters of the variational autoencoder are transferred to the multi-layer perceptron, and the multi-layer perceptron is supervised-trained using the training set, the parameters of the multi-layer perceptron are fine-tuned, and the validation set is used for evaluation after each round of training until the set number of rounds of training or the evaluation accuracy of the validation set reaches the standard. The multi-layer perceptron with fine-tuned parameters is used as the autism classifier.

Description

Technical Field

[0001] The present invention relates to the field of medical image classification, and in particular, to a method and system for constructing an autism classifier based on human brain functional magnetic resonance imaging. Background Art

[0002] Autism spectrum disorder (ASD) is a common complex neurodevelopmental disorder that occurs in early childhood, and its core features are social and restricted repetitive sensorimotor behaviors. Traditional symptom-based classification methods cannot reveal the pathogenesis behind ASD, so they are often unreliable. With the development of neuroimaging, non-invasive brain imaging technology has become a powerful tool for studying and revealing neurological diseases such as ASD. Among them, functional magnetic resonance imaging (rs-fMRI) measures the change signals related to blood oxygen levels, which can help clinicians and neuroscientists visually evaluate the functional characteristics or properties of the brain, and has become a powerful tool for early classification of ASD. In recent years, the combination of rs-fMRI with machine learning and deep learning technologies for ASD classification has achieved good results and has become one of the most promising imaging methods for ASD classification.

[0003] In recent years, machine learning (including deep learning) methods have been widely used in the classification of autism. In recent years, machine learning (including deep learning) methods have been widely used in the prediction and research of cell images. Plitt et al. [1] Only used support vector machines to classify autism examples and normal person examples in the ABIDE dataset, achieving a classification accuracy of 69%. Heinsfeld et al. [2] Utilized stacked autoencoders (SAE) and fully connected neural networks to achieve the highest prediction accuracy of 70% at that time on the public dataset ABIDE I. Parisot et al. [3] Proposed a general framework that uses imaging and non-imaging information for brain analysis of large populations. This framework is based on graph convolutional neural networks (GCN) and achieved a classification accuracy of 70.4% on the ABIDE dataset. Zhi-An Huang et al. [4] Utilized deep belief networks (DBN) to achieve a prediction accuracy of 76.4% on the ABIDE I dataset. In addition, feature selection methods are also often combined with machine learning to obtain better classification performance.

[0004] However, the main challenge of current brain functional magnetic resonance imaging is that the preprocessed data contains a large amount of redundant information, which will lead to the deterioration of the performance of the classification model. At present, the accuracy of machine learning classification models based on functional magnetic resonance imaging still needs to be improved. The current classification models cannot flexibly adjust sensitivity and specificity, so that they cannot adapt to certain specific actual needs.

[0005] 1. Plitt, M., Barnes, K. A., and Martin, A. (2015). Functional connectivity classification of autism identifies highly predictive brain features but falls short of biomarker standards. YNICL 7, 359–366. doi:10.1016 / j.nicl.2014.12.013

[0006] 2. Heinsfeld, A. S.; Franco, A. R.; Craddock, R. C.; Buchweitz, A.; Meneguzzi, F., Identification of autism spectrum disorder using deep learning and the ABIDE dataset. NeuroImage: Clinical 2018, 17, 16 - 23.

[0007] 3. Parisot, S.; Ktena, S. I.; Ferrante, E.; Lee, M.; Guerrero, R.; Glocker, B.; Rueckert, D., Disease prediction using graph convolutional networks: application to autism spectrum disorder and Alzheimer’s disease. Medical image analysis 2018, 48, 117 - 130.

[0008] 4. Huang, Z.-A.; Zhu, Z.; Yau, C. H.; Tan, K. C., Identifying autism spectrum disorder from resting-state fMRI using deep belief network. IEEE Transactions on Neural Networks and Learning Systems 2020. Summary of the Invention

[0009] The purpose of the present invention is to provide a method and system for constructing an autism classifier based on human brain functional magnetic resonance imaging, so as to overcome the defects of the prior art. The present invention can select more refined useful features than the commonly used methods in the past, and has high accuracy, sensitivity, specificity and training speed.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A method for constructing an autism classifier based on human brain functional magnetic resonance imaging includes:

[0012] Preprocessing the brain functional magnetic resonance imaging of normal subjects and autistic subjects to obtain feature vectors;

[0013] Performing feature selection on each feature of the feature vector based on the difference in the gradient distribution curve to obtain significant features, forming a new feature vector with the obtained significant features, and using the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set;

[0014] Using the training set to pre-train the variational autoencoder. After the pre-training is completed, migrating the encoder parameters of the variational autoencoder to the multi-layer perceptron, and using the training set to perform supervised training on the multi-layer perceptron, fine-tuning the parameters of the multi-layer perceptron, and evaluating with the validation set after each round of training until the set number of rounds of training is reached. Taking the multi-layer perceptron with fine-tuned parameters as the autism classifier.

[0015] Further, the process of preprocessing the brain functional magnetic resonance imaging of normal subjects and autistic subjects to obtain feature vectors is specifically as follows: calculating the functional connectivity between regions of interest based on the time series extracted from the human brain functional magnetic resonance imaging, and all the functional connectivities form a feature vector, with each functional connectivity being a feature of the feature vector;

[0016] The specific calculation method of the functional connectivity is: calculating the Pearson correlation coefficient of the time series corresponding to all pairwise different regions of interest.

[0017] Further, the process of performing feature selection on each feature of the feature vector based on the difference in the gradient distribution curve to obtain significant features is specifically as follows:

[0018] Calculating the DSDC score for each feature in the feature vector;

[0019] Selecting the features with DSDC scores greater than the preset threshold as significant features, and discarding the features with DSDC scores less than or equal to the preset threshold.

[0020] Further, the process of calculating the DSDC score for each feature in the feature vector is specifically as follows:

[0021] Divide the range where the features in all feature vectors are located into several equal-length sub-intervals;

[0022] Calculate the DSDC score of the feature through the following formula:

[0023]

[0024] In the formula, b 0 and b 1 are the lower and upper bounds of the feature value; δ is the length of the sub-interval; i represents the i-th sub-interval, n i + and n i - are the numbers of autistic subjects and normal subjects falling in the interval [i - δ, i), N + and N – are the numbers of autistic subjects and normal subjects in the training set.

[0025] Furthermore, the supervised training of the multi-layer perceptron using the training set is specifically unconstrained training, training using sensitivity constraint conditions, or training using specificity constraint conditions;

[0026] When unconstrained training is adopted, it is evaluated using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the parameters after this round of training are saved; otherwise, they are not saved;

[0027] When training using sensitivity constraint conditions, it is evaluated using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation sensitivity and the evaluation specificity of the validation set improves and the improvement value is less than the preset value, the parameters after this round of training are saved; otherwise, they are not saved;

[0028] When training using specificity constraint conditions, it is evaluated using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation specificity and the evaluation sensitivity of the validation set improves and the improvement value is less than the preset value, the parameters after this round of training are saved; otherwise, they are not saved.

[0029] An autism classifier construction system based on human brain functional magnetic resonance imaging, including a preprocessing module, a training set acquisition module, and a training module, where:

[0030] Preprocessing module: used to preprocess the brain functional magnetic resonance imaging of normal subjects and autistic subjects to obtain feature vectors;

[0031] Training set acquisition module: used to perform feature selection for each feature of the feature vector based on the difference in the gradient distribution curve to obtain significant features, form a new feature vector with the obtained significant features, and use the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set;

[0032] Training module: used to pre-train the variational autoencoder with the training set. After the pre-training is completed, transfer the encoder parameters of the variational autoencoder to the multi-layer perceptron, and perform supervised training on the multi-layer perceptron with the training set, fine-tune the parameters of the multi-layer perceptron, and evaluate with the validation set after each round of training until the set number of rounds of training is reached, and use the multi-layer perceptron with fine-tuned parameters as the autism classifier.

[0033] Further, the process of preprocessing the brain functional magnetic resonance images of normal subjects and autistic subjects to obtain feature vectors is specifically as follows: calculate the functional connectivity between regions of interest based on the time series extracted from the human brain functional magnetic resonance images, and all functional connectivities form a feature vector, and each functional connectivity is used as a feature of the feature vector;

[0034] The specific calculation method of the functional connectivity is: calculate the Pearson correlation coefficient of the time series corresponding to all pairwise different regions of interest.

[0035] Further, the process of performing feature selection for each feature of the feature vector based on the difference in the gradient distribution curve to obtain significant features is specifically as follows:

[0036] For each feature in the feature vector, calculate the DSDC score;

[0037] Select the features with DSDC scores greater than the preset threshold as significant features, and discard the features with DSDC scores less than or equal to the preset threshold.

[0038] Further, the process of calculating the DSDC score for each feature in the feature vector is specifically as follows:

[0039] Divide the range where the features in all feature vectors are located into several equal-length sub-intervals;

[0040] Calculate the DSDC score of the feature through the following formula:

[0041]

[0042] In the formula, b 0 and b 1 are the lower and upper bounds of the feature value; δ is the length of the sub-interval; i represents the i-th sub-interval, n i + and n i- is the number of autistic and normal subjects falling within the interval [i - δ, i), N + and N – are the numbers of autistic and normal subjects in the training set.

[0043] Further, the supervised training of the multi - layer perceptron using the training set is specifically unconstrained training, training using a sensitivity constraint condition, or training using a specificity constraint condition;

[0044] When using unconstrained training, after each round of training, the validation set is used for evaluation. If the evaluation accuracy of the validation set improves, the parameters after this round of training are saved; otherwise, they are not saved.

[0045] When using training with a sensitivity constraint condition, after each round of training, the validation set is used for evaluation. If the evaluation accuracy of the validation set improves, the difference between the evaluation sensitivity and the evaluation specificity of the validation set improves, and the improvement value is less than a preset value, the parameters after this round of training are saved; otherwise, they are not saved.

[0046] When using training with a specificity constraint condition, after each round of training, the validation set is used for evaluation. If the evaluation accuracy of the validation set improves, the difference between the evaluation specificity and the evaluation sensitivity of the validation set improves, and the improvement value is less than a preset value, the parameters after this round of training are saved; otherwise, they are not saved.

[0047] Compared with the prior art, the present invention has the following beneficial technical effects:

[0048] First, the feature selection method based on the difference of step - distribution curves (DSDC) used in the method of the present invention can select more refined useful features than the commonly used methods in the past; second, the accuracy, sensitivity, specificity, and training speed of the classifier constructed by the present invention exceed the most advanced experimental results in previous studies based on the same data set; finally, the present invention can flexibly adjust the sensitivity and specificity of the classifier by imposing constraint conditions during the training process, enabling the model to meet different actual application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings in the specification are used to provide a further understanding of the present invention, and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0050] Figure 1 is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0054] The present invention provides a method for constructing an autism classifier based on human brain functional magnetic resonance imaging, which specifically includes:

[0055] The first step is to preprocess the functional magnetic resonance imaging of the brains of 455 normal subjects and 477 autistic subjects to obtain feature vectors. Specifically, the functional connections between the regions of interest are calculated based on the time series extracted from the functional magnetic resonance imaging of the human brain. Each functional connection is used as a feature, and all functional connections form a feature vector. The specific calculation method of functional connectivity is to calculate the Pearson correlation coefficient of the time series corresponding to all two different regions of interest.

[0056] The second step is to perform feature selection based on the gradient distribution curve difference based on the feature vector obtained in the first step. For each feature, calculate the DSDC score of the feature: first, divide the range of the feature in all feature vectors into 20 equal-length sub-intervals, then count the number of samples in each sub-interval, and divide the number by the total number of samples in the corresponding category for normalization, and generate a step distribution curve according to the normalized value of the number of samples in each interval. This process is equivalent to a coarse-grained fitting of the original distribution curve of the feature, that is, the DSDC score of the feature is calculated by the following formula.

[0057]

[0058] In the formula, b 0 and b1 Lower and upper bounds of the value of the feature; δ is the length of the sub-interval; i represents the i-th sub-interval, and the value of i is an integer within the range of [1, 20]; n i + and n i - are the numbers of autistic subjects and normal subjects falling within the interval [i - δ, i), N + and N – are the numbers of autistic subjects and normal subjects in the training set.

[0059] The present invention pre-sets a filtering threshold, selects features with DSDC scores greater than the threshold as significant features, discards features with DSDC scores less than or equal to the threshold, and uses the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set.

[0060] Step 3: The present invention simplifies the encoder structure of the variational autoencoder, uses the same neural network to generate two parameters of the latent space, and pre-trains the variational autoencoder using the training set.

[0061] Step 4: After the pre-training is completed, transfer the encoder parameters of the variational autoencoder to the multi-layer perceptron, and perform supervised training on the multi-layer perceptron using the training set, fine-tune the parameters of the multi-layer perceptron, and evaluate using the validation set after each round of training until the set number of rounds of training is reached. Use the multi-layer perceptron with fine-tuned parameters as an autism classifier for autism classification. During the training process of the multi-layer perceptron, the present invention designs two constraint conditions (sensitivity constraint and specificity constraint) as optional items. During the training process of the multi-layer perceptron, adding the sensitivity constraint and the specificity constraint can significantly improve the sensitivity and specificity respectively at the cost of a small reduction in accuracy.

[0062] The main functions of the sensitivity constraint and the specificity constraint are to determine whether the parameters obtained by the classifier after each round of training are saved. In the present invention, one-tenth of the data in the training set is taken out as the validation set, which is mainly used to evaluate each round of training models and does not participate in the training process. Without constraint conditions, after a certain round of training, if the evaluation accuracy of the validation set improves, the parameters obtained through this round of training will be saved; otherwise, they will not be saved. When using the sensitivity constraint, after a certain round of training, if the evaluation accuracy of the validation set improves, the difference between the evaluation sensitivity and the evaluation specificity of the validation set improves and is less than 0.3, the parameters obtained through this round of training will be saved; otherwise, they will not be saved. When using the specificity constraint, after a certain round of training, if the evaluation accuracy of the validation set improves, the difference between the evaluation specificity and the evaluation sensitivity of the validation set improves and is less than 0.3, the parameters obtained through this round of training will be saved; otherwise, they will not be saved. Using the sensitivity constraint helps to improve the model sensitivity; using the specificity constraint helps to improve the model specificity.

[0063] The present invention also provides an autism classifier construction system based on human brain functional magnetic resonance imaging, including a preprocessing module, a training set acquisition module, and a training module, where:

[0064] The preprocessing module: is used to preprocess the brain functional magnetic resonance images of normal subjects and autistic subjects to obtain feature vectors;

[0065] The training set acquisition module: is used to perform a feature selection method based on the difference in gradient distribution curves for each feature of the feature vectors to obtain significant features, form new feature vectors from the obtained significant features, and use the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set;

[0066] The training module: is used to pre-train a variational autoencoder using the training set. After the pre-training is completed, the encoder parameters of the variational autoencoder are migrated to a multi-layer perceptron, and the multi-layer perceptron is supervised-trained using the training set, the parameters of the multi-layer perceptron are fine-tuned, and the validation set is used for evaluation after each round of training until the set number of rounds of training is reached. The multi-layer perceptron with fine-tuned parameters is used as the autism classifier.

[0067] The present invention has been experimented and tested on the ABIDE I dataset. The experimental results show that the accuracy (78.12%) (without using constraint conditions), sensitivity (87.20%) (using the sensitivity constraint), specificity (88.55%) (using the specificity constraint), and training speed (10-fold cross-validation takes 85 seconds) of the classifier constructed by the present invention all exceed the most advanced experimental results in previous studies based on the same dataset.

[0068] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0069] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0070] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that realizes the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the scope of its protection. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: after reading the present invention, those skilled in the art can still make various changes, modifications, or equivalent replacements to the specific implementation manners of the invention, but these changes, modifications, or equivalent replacements are all within the scope of the claims of the invention pending approval.

Claims

1. Method for constructing an autism classifier based on human brain functional magnetic resonance imaging, characterized in that, it includes: Preprocessing the brain functional magnetic resonance imaging of normal subjects and autistic subjects to obtain feature vectors; Performing feature selection on each feature of the feature vector based on the difference in gradient distribution curves to obtain significant features, forming a new feature vector with the obtained significant features, and using the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set; Pre-training a variational autoencoder using the training set. After the pre-training is completed, transfer the encoder parameters of the variational autoencoder to a multi-layer perceptron, and perform supervised training on the multi-layer perceptron using the training set, fine-tune the parameters of the multi-layer perceptron, and evaluate using the validation set after each round of training until the set number of rounds of training is reached. Use the multi-layer perceptron with fine-tuned parameters as the autism classifier; Among them, the process of preprocessing the brain functional magnetic resonance imaging of normal subjects and autistic subjects to obtain feature vectors is specifically: calculating the functional connectivity between regions of interest based on the time series extracted from the human brain functional magnetic resonance imaging. All functional connectivities form a feature vector, and each functional connectivity is a feature of the feature vector; The specific calculation method of the functional connectivity is: calculating the Pearson correlation coefficient of the time series corresponding to all pairwise different regions of interest; The process of performing feature selection on each feature of the feature vector based on the difference in gradient distribution curves to obtain significant features is specifically: For each feature in the feature vector, calculate the DSDC score, specifically: Dividing the range where the features in all feature vectors are located into several equal-length sub-intervals; Calculating the DSDC score of the feature through the following formula: where b 0 and b 1 are the lower and upper bounds of the feature value; δ is the length of the sub - interval; i represents the i - th sub - interval, n i + and n i - are the numbers of autistic and normal subjects falling in the interval [i - δ, i), N + and N – are the numbers of autistic and normal subjects in the training set; Select the features with DSDC scores greater than the preset threshold as significant features, and discard the features with DSDC scores less than or equal to the preset threshold.

2. The method for constructing an autism classifier based on human brain functional magnetic resonance imaging according to claim 1, characterized in that, The process of performing supervised training on the multi-layer perceptron using the training set is specifically unconstrained training, training using sensitivity constraint conditions, or training using specificity constraint conditions; When using unconstrained training, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, save the parameters after this round of training, otherwise do not save; When using sensitivity constraint conditions for training, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation sensitivity and the evaluation specificity of the validation set improves and the improvement value is less than the preset value, save the parameters after this round of training, otherwise do not save; When using specificity constraint conditions for training, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation specificity and the evaluation sensitivity of the validation set improves and the improvement value is less than the preset value, save the parameters after this round of training, otherwise do not save.

3. System for constructing an autism classifier based on human brain functional magnetic resonance imaging, characterized in that, It includes a preprocessing module, a training set acquisition module, and a training module, where: The preprocessing module: is used to preprocess the brain functional magnetic resonance images of normal subjects and autistic subjects to obtain feature vectors; The training set acquisition module: is used to perform feature selection on each feature of the feature vector based on the difference in gradient distribution curves to obtain significant features, form a new feature vector with the obtained significant features, and use the new feature vectors of all subjects as training samples, where the training samples include a training set and a validation set; The training module: is used to pre-train a variational autoencoder using the training set. After the pre-training is completed, transfer the encoder parameters of the variational autoencoder to a multi-layer perceptron, and perform supervised training on the multi-layer perceptron using the training set, fine-tune the parameters of the multi-layer perceptron, and evaluate using the validation set after each round of training until the set number of rounds of training is reached, and use the multi-layer perceptron with fine-tuned parameters as an autism classifier; Among them, the process of preprocessing the brain functional magnetic resonance images of normal subjects and autistic subjects to obtain feature vectors is specifically as follows: calculate the functional connectivity between regions of interest based on the time series extracted from the human brain functional magnetic resonance images, and all functional connectivities form a feature vector, and each functional connectivity is a feature of the feature vector; The specific calculation method of the functional connectivity is: calculate the Pearson correlation coefficient of the time series corresponding to all pairwise different regions of interest; The process of performing feature selection on each feature of the feature vector based on the difference in gradient distribution curves to obtain significant features is specifically as follows: For each feature in the feature vector, calculate the DSDC score specifically as: Divide the range where the features in all feature vectors are located into several equal-length sub-intervals; Calculate the DSDC score of the feature through the following formula: where b 0 and b 1 are the lower and upper bounds of the feature value; δ is the length of the sub - interval; i represents the i - th sub - interval, n i + and n i - are the numbers of autistic subjects and normal subjects falling in the interval [i - δ, i); N + and N – are the numbers of autistic subjects and normal subjects in the training set; Select the features with DSDC scores greater than the preset threshold as significant features, and discard the features with DSDC scores less than or equal to the preset threshold.

4. The autism classifier construction system based on human brain functional magnetic resonance images according to claim 3, characterized in that The process of performing supervised training on the multi-layer perceptron using the training set is specifically unconstrained training, training using a sensitivity constraint condition, or training using a specificity constraint condition; When performing unconstrained training, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, save the parameters after this round of training, otherwise do not save; When training using a sensitivity constraint condition, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation sensitivity and the evaluation specificity of the validation set improves and the improvement value is less than the preset value, save the parameters after this round of training, otherwise do not save; When training using a specificity constraint condition, evaluate using the validation set after each round of training. If the evaluation accuracy of the validation set improves, the difference between the evaluation specificity and the evaluation sensitivity of the validation set improves and the improvement value is less than the preset value, save the parameters after this round of training, otherwise do not save.

Images