Autism auxiliary diagnosis system based on lightweight geometric manifold deep network

Abstract

The invention discloses an autism auxiliary diagnosis system based on a lightweight geometric manifold deep network, and relates to the technical field of medical image analysis and computer-aided diagnosis, in particular to the autism auxiliary diagnosis system. The method aims at solving the problems that an existing model is generally large in model parameter quantity, low in calculation efficiency, limited in three-dimensional topological structure modeling capacity and the like. The system comprises a training set acquisition module, a trained lightweight geometric manifold deep network model LGMD-Net acquisition module and a to-be-tested functional magnetic resonance imaging fMRI prediction module. The to-be-tested functional magnetic resonance imaging fMRI prediction module is used for obtaining to-be-tested functional magnetic resonance imaging fMRI; the functional magnetic resonance imaging fMRI is processed, and an ALFF image to be detected after mean value normalization is obtained; and inputting the ALFF to-be-detected image after mean normalization into the trained lightweight geometric manifold deep network LGMD-Net, and outputting a classification probability according to the type of autism.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and computer-aided diagnosis technology, specifically to an autism auxiliary diagnosis system. Background Technology

[0002] With the continued rise in the prevalence of autism spectrum disorder (ASD), its early and accurate diagnosis has become a crucial topic in clinical neurodevelopmental disorder research. Currently, ASD diagnosis primarily relies on behavioral observation and scale assessments, lacking objective biomarkers, and is highly subjective, prone to missed diagnoses and misdiagnoses. Functional magnetic resonance imaging (fMRI) provides important imaging biomarkers for ASD, especially the amplitude of low-frequency fluctuation (ALFF) index, which is widely used in brain function analysis due to its data-driven approach and lack of prior assumptions.

[0003] In fMRI-based computer-aided diagnostic methods, convolutional neural networks (CNNs) have become one of the mainstream technologies. For example, Thomas et al. (2019) proposed using 3D-CNN combined with ensemble learning strategies to fully utilize the three-dimensional structure of fMRI to improve classification performance, but problems such as model complexity, large training data requirements, and high computational costs still exist. Ali et al. (2020) constructed a deep, dense convolutional network to enhance feature representation capabilities, but the model parameter burden was significant, limiting its practical application in clinical settings. To reduce complexity, Min et al. (2021) used 2D-CNN to process fMRI slices, which improved model robustness, but due to ignoring the three-dimensional context and geometric structure, spatial feature extraction was incomplete, and conventional pooling operations further caused spatial information loss.

[0004] Therefore, although existing CNN methods have shown good feature mining capabilities in disease classification, they generally suffer from problems such as large number of model parameters, low computational efficiency, and limited ability to model three-dimensional topological structures, making it difficult to achieve an ideal balance between accuracy and efficiency. Summary of the Invention

[0005] The present invention addresses the problems of existing models, such as large number of model parameters, low computational efficiency, and limited ability to model three-dimensional topological structures, making it difficult to achieve an ideal balance between accuracy and efficiency. Therefore, it proposes an autism auxiliary diagnostic system based on lightweight geometric manifold deep networks.

[0006] Autism-assisted diagnostic systems based on lightweight geometric manifold deep networks include:

[0007] The training set acquisition module, the pre-trained lightweight geometric manifold deep network model LGMD-Net acquisition module, and the fMRI prediction module for the functional magnetic resonance imaging (fMRI) test are all included.

[0008] The training set acquisition module is used to obtain functional magnetic resonance imaging (fMRI); the fMRI is preprocessed to obtain preprocessed fMRI; the preprocessed fMRI is further processed to obtain mean-normalized ALFF images, which are used as the training set.

[0009] The module for obtaining the trained lightweight geometric manifold deep network model LGMD-Net is used to construct the lightweight geometric manifold deep network LGMD-Net. The lightweight geometric manifold deep network LGMD-Net is trained based on the mean-normalized ALFF images in the training set to obtain the trained lightweight geometric manifold deep network LGMD-Net.

[0010] The fMRI prediction module is used to obtain the fMRI image of the target functional magnetic resonance imaging (fMRI). The fMRI image is preprocessed to obtain a preprocessed fMRI image of the target functional magnetic resonance imaging (fMRI). The preprocessed fMRI image of the target functional magnetic resonance imaging (fMRI) is further processed to obtain a mean-normalized ALFF image of the target functional magnetic resonance imaging (ALFF). The mean-normalized ALFF image of the target functional magnetic resonance imaging (ALFF) is then input into a trained lightweight geometric manifold deep network (LGMD-Net). The trained lightweight geometric manifold deep network (LGMD-Net) outputs a classification probability, with the category being whether or not the individual has autism.

[0011] The beneficial effects of this invention are as follows:

[0012] The experiment divided the processed ALFF data into training, validation, and test sets at 80%, 10%, and 10% ratios, respectively. The training set contained 1207 samples, and the test set contained 151 samples. Evaluation metrics used included accuracy (ACC), precision (PRE), sensitivity (SEN), F1 score (F1), area under the receiver operating characteristic curve (AUC), and specificity (SPE).

[0013] To ensure the validity and reliability of the experiments, LGMD-Net was systematically compared and analyzed with classic convolutional networks (2D / 3D ResNet, DenseNet, Auto-Encoder), advanced convolutional models (ACS CNN), and hybrid Transformer models (ViT, SwinTransformer). The performance results of the comparative experiments are shown in Table 1.

[0014]

[0015] An asterisk indicates the optimal metric for the classification result;

[0016] In addition, the time complexity performance of all the above 2D models and traditional models was evaluated under the same hyperparameter settings, and the experimental results are shown in Table 2.

[0017]

[0018] An asterisk (*) indicates the optimal measure of time complexity.

[0019] LGMD-Net demonstrates significant advantages in both classification accuracy and engineering practicality. Experimental results show that on the ABIDE dataset, LGMD-Net outperforms traditional convolutional networks and hybrid Transformer models in classification accuracy, F1 score, and robustness. It also exhibits significant advantages in inference speed and memory usage, demonstrating the method's comprehensive competitiveness in both accuracy and engineering practicality. The experimental results show that LGMD-Net significantly reduces its dependence on computational resources while maintaining high classification performance, exhibiting better practicality and generalization potential. Attached Figure Description

[0020] Figure 1 The overall network framework diagram for a lightweight deep feature extraction network;

[0021] Figure 2 This is an architecture diagram of a multi-channel 2D CNN model;

[0022] Figure 3 This is a schematic diagram of the Basic Block module structure;

[0023] Figure 4 This is a schematic diagram of a spatial pyramid structure. Detailed Implementation

[0024] Specific Implementation Method 1: This implementation method, based on a lightweight geometric manifold deep network, includes an autism auxiliary diagnostic system comprising:

[0025] The training set acquisition module, the pre-trained lightweight geometric manifold deep network model LGMD-Net acquisition module, and the fMRI prediction module for the functional magnetic resonance imaging (fMRI) test are all included.

[0026] The training set acquisition module is used to obtain functional magnetic resonance imaging (fMRI); the fMRI is preprocessed to obtain preprocessed fMRI.

[0027] The preprocessed functional magnetic resonance imaging (fMRI) images were processed to obtain mean-normalized low-frequency amplitude (ALFF) images, which were used as the training set.

[0028] The pre-trained Lightweight Geometric Manifold Deep Network (LGMD-Net) model acquisition module is used to construct the Lightweight Geometric Manifold Deep Network (LGMD-Net). The Lightweight Geometric Manifold Deep Network (LGMD-Net) is trained based on the mean-normalized low-frequency amplitude (ALFF) images in the training set to obtain the pre-trained Lightweight Geometric Manifold Deep Network (LGMD-Net).

[0029] The fMRI prediction module is used to obtain the fMRI image of the target functional magnetic resonance imaging (fMRI). The fMRI image is preprocessed to obtain a preprocessed fMRI image of the target functional magnetic resonance imaging (fMRI). The preprocessed fMRI image is further processed to obtain a mean-normalized ALFF low-frequency amplitude image of the target functional magnetic resonance imaging (fMRI). The mean-normalized ALFF low-frequency amplitude image of the target functional magnetic resonance imaging (fMRI) is then input into a trained lightweight geometric manifold deep network (LGMD-Net). The trained lightweight geometric manifold deep network (LGMD-Net) outputs a classification probability, where the category is autism. The closer the probability is to 1, the higher the probability of autism; the closer the probability is to 0, the lower the probability of autism.

[0030] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: the training set acquisition module is used to obtain functional magnetic resonance imaging (fMRI); the fMRI is preprocessed to obtain preprocessed fMRI.

[0031] The preprocessed functional magnetic resonance imaging (fMRI) images were processed to obtain mean-normalized low-frequency amplitude (ALFF) images, which were used as the training set.

[0032] The specific process is as follows:

[0033] 1) Use the RESTplus tool to systematically process functional magnetic resonance imaging (fMRI) images to obtain preprocessed fMRI images; the specific process is as follows:

[0034] The systematic processing is as follows: removing fMRI from the first 10 time points, time-layer correction, head motion correction, registration, spatial smoothing, removal of linear trends, covariate regression analysis, and filtering.

[0035] 2) The preprocessed functional magnetic resonance imaging (fMRI) images are processed to obtain mean-normalized low-frequency amplitude (ALFF) images; the specific process is as follows:

[0036] 21) The brain map of the preprocessed functional magnetic resonance imaging (fMRI) at one time point is cut into small blocks of 3mm×3mm×3mm, each block is 1 voxel, and each voxel corresponds to 1 BOLD signal at each time point in the preprocessed fMRI. The mean value of all BOLD signals corresponding to all time points of each voxel in the preprocessed fMRI is taken to obtain the BOLD temporal mean value of each voxel at each time point.

[0037] 22) Subtract the mean BOLD time series value of the same voxel at each time point from the BOLD signal value of each voxel at each time point to obtain the mean BOLD time series value of each voxel at each time point.

[0038] 23) Perform Fast Fourier Transform (FFT) on the BOLD time series value of each voxel after removing the mean at each time point to obtain the complex spectrum of each voxel at each time point after Fast Fourier Transform (FFT obtains a complex spectrum, which changes from multiple time points to multiple frequency points).

[0039] 24) Obtain the amplitude spectrum by taking the modulus of the complex spectral coefficients corresponding to each voxel at each time point. (Make all frequency values ​​positive);

[0040] The amplitude spectrum that meets the requirements for each voxel at all time points The ALFF value of the low-frequency amplitude image for each voxel is obtained by averaging.

[0041] Amplitude spectrum that meets the requirements for: ;

[0042] The low-frequency amplitude image (ALFF) is a three-dimensional image in the axial, coronal, and sagittal directions;

[0043] 25) Take the average ALFF value of the low-frequency amplitude image of all voxels to obtain the average ALFF value of the low-frequency amplitude image;

[0044] Divide the ALFF value of the low-frequency amplitude image of each voxel by the mean ALFF value of the low-frequency amplitude image to obtain the mALFF of each voxel in the whole brain; mALFF represents the ALFF image after mean normalization; it is used to suppress differences at the whole brain scale and highlight the relative strength distribution of local low-frequency activities; in order to further suppress noise, ALFF needs to be standardized.

[0045] The dataset used was the ABIDE (Autism Brain Imaging Data Exchange) dataset, which includes ABIDE-I and ABIDE-II. It collected functional magnetic resonance imaging (fMRI) and phenotypic data from 2226 participants across 36 international sites. The experiment used a quality-checked subset of 1509 participants, including 673 individuals with autism spectrum disorder (ASD) and 836 healthy controls (HC).

[0046] In MATLAB 2022, the RESTplus package is used to preprocess the ABIDE dataset;

[0047] The purpose of preprocessing is to minimize errors caused by data acquisition and physiological characteristics.

[0048] fMRI scanning process: There are generally more than 100 time points (different machines have different time points). One time point (usually 2s or 3s) scans one brain (three-dimensional, axial, coronal and sagittal three-dimensional planes). This brain is composed of several time layers. It is scanned one by one (due to machine limitations). One scan is counted as one time layer. There are generally more than 30 or 40 time layers.

[0049] After removing the first 10 time points, the time required for the machine to be turned on and for the patient to stabilize after being placed in the MRI machine is generally considered to be 20 seconds.

[0050] Temporal correction (the terminology in the text is incorrect) uses the intermediate layer as the reference layer and performs temporal correction according to the scanning order of the temporal layers. It uses linear interpolation to obtain data from the same time point in multiple brain slices. Interpolation is the process of predicting the value of a location point by using the signal values ​​of known points adjacent to the unknown point.

[0051] For head movement correction, the target image is the first image; the images at other time points are translated and rotated (rigid body transformation) to make them coincide with the target image; this is head movement correction; subjects with head movements >3mm3° will be excluded here;

[0052] Spatial registration involves registering the average map of the subject at multiple time points with a standard template (MNI standard space) using rigid body transformation to obtain a set of parameters. Applying these transformation parameters to the mind maps at all time points completes the registration from individual space to standard space.

[0053] Spatial smoothing, Gaussian smoothing with a 6mm Gaussian kernel;

[0054] To de-linearize the trend, due to machine heat and prolonged operation, the signal may drift upwards. Linear regression is used to bring the signal back to a horizontal fluctuation state.

[0055] Regression covariates: Linear regression of the different head-motion rigid body transformation parameters for each individual.

[0056] The other steps and parameters are the same as in Specific Implementation Method 1.

[0057] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that: the trained Lightweight Geometric Manifold Deep Network (LGMD-Net) acquisition module is used to construct the Lightweight Geometric Manifold Deep Network (LGMD-Net). The Lightweight Geometric Manifold Deep Network (LGMD-Net) is trained based on the mean-normalized ALFF images in the training set to obtain the trained Lightweight Geometric Manifold Deep Network (LGMD-Net);

[0058] The specific process is as follows:

[0059] 21. Construct a lightweight geometric manifold deep network LGMD-Net; the specific process is as follows:

[0060] The lightweight geometric manifold deep network LGMD-Net consists of a multi-channel shallow extraction layer, a lightweight residual layer, a manifold pyramid pooling layer, and a classifier.

[0061] The multi-channel shallow extraction layer includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a first batch normalization layer (BN), and a first ReLU activation function layer;

[0062] The kernel size of the first convolutional layer is Step size is 2, fill size is 3;

[0063] The kernel size of the second convolutional layer is Step size is 2, fill size is 3;

[0064] The kernel size of the third convolutional layer is Step size is 2, fill size is 3;

[0065] The lightweight residual layer includes: a fourth convolutional layer, a second batch normalization layer (BN), a second ReLU activation function layer, a fifth convolutional layer, a third batch normalization layer (BN), a third ReLU activation function layer, a first downsampling layer, a sixth convolutional layer, a fourth batch normalization layer (BN), a fourth ReLU activation function layer, a seventh convolutional layer, a fifth batch normalization layer (BN), a fifth ReLU activation function layer, and a second downsampling layer;

[0066] The manifold pyramid pooling layer includes: a first pooling layer, a second pooling layer, a third pooling layer, and a fully connected layer;

[0067] The feature map size corresponding to the first pooling layer is [size missing]. ;

[0068] The feature map size corresponding to the second pooling layer is [size missing]. ;

[0069] The feature map size corresponding to the third pooling layer is [size missing]. ;

[0070] The classifier includes: a flattening layer, a first fully connected layer (output dimension 512), a second fully connected layer (output dimension 2), and a Softmax activation function layer;

[0071] 22. The lightweight geometric manifold deep network LGMD-Net is trained based on the mean-normalized ALFF images in the training set to obtain the trained lightweight geometric manifold deep network LGMD-Net.

[0072] Other steps and parameters are the same as in specific implementation method one or two.

[0073] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: step 22 involves training the Lightweight Geometric Manifold Deep Network LGMD-Net based on the mean-normalized ALFF image from the training set, thereby obtaining the trained Lightweight Geometric Manifold Deep Network LGMD-Net; the specific process is as follows:

[0074] 221. Input the mean-normalized ALFF images from the training set into the multi-channel shallow extraction layer, and the multi-channel shallow extraction layer outputs feature map E;

[0075] 222. Input the feature map E into the lightweight residual layer, and the lightweight residual layer outputs the feature map M;

[0076] 223. The feature map M is input to the manifold pyramid pooling layer, and the feature map R is output by the manifold pyramid pooling layer.

[0077] 224. The feature map R is input into the classifier, and the classifier outputs the classification probability;

[0078] 225. Repeat steps 221 to 224 until the maximum number of iterations is reached to obtain the trained lightweight geometric manifold deep network LGMD-Net.

[0079] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0080] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step 221, the ALFF image after mean normalization in the training set is input into the multi-channel shallow extraction layer, and the multi-channel shallow extraction layer outputs feature map E; the specific process is as follows:

[0082] 2211. Standardize the 3D ALFF images in the training set after mean normalization to obtain the standardized 2D image ALFF. The dimensions of the standardized 2D image ALFF are as follows: , , The specific process is as follows:

[0083] The mean-normalized 3D ALFF images in the training set are sliced ​​in the axial, coronal, and sagittal directions respectively, generating 61 slices in the axial direction, 61 slices in the coronal direction, and 61 slices in the sagittal direction.

[0084] Sixty-one slices were generated based on the axial direction, sixty-one slices were generated in the coronal direction, and sixty-one slices were generated in the sagittal direction, which were then reconstructed into axial and coronal sections, respectively. Two-dimensional images of the channels, axial and sagittal Two-dimensional images of the channels, coronal and sagittal Two-dimensional image of the channel;

[0086] 2212. The axial coronal The two-dimensional image ALFF of the channel is input into the first convolutional layer, and the first convolutional layer outputs feature map A;

[0087] axial sagittal The two-dimensional image ALFF of the channel is input into the second convolutional layer, and the second convolutional layer outputs feature map B;

[0088] coronal sagittal The two-dimensional image ALFF of the channel is input into the third convolutional layer, and the third convolutional layer outputs the feature map C;

[0089] The kernel size of the first convolutional layer is Step size is 2, fill size is 3;

[0090] The kernel size of the second convolutional layer is Step size is 2, fill size is 3;

[0091] The kernel size of the third convolutional layer is Step size is 2, fill size is 3;

[0092] Feature map A, feature map B, and feature map C are input into the first batch normalization layer BN, and the first batch normalization layer BN outputs feature map D.

[0093] The feature map D is input into the first ReLU activation function layer, and the first ReLU activation function layer outputs the feature map E.

[0094] The other steps and parameters are the same as in any of the specific implementation methods one to four.

[0095] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: in step 222, the feature map E is input into the lightweight residual layer, and the lightweight residual layer outputs the feature map M; the specific process is as follows:

[0096] The feature map E is sequentially input into the fourth convolutional layer, the second batch normalization layer BN, the second ReLU activation function layer, the fifth convolutional layer, and the third batch normalization layer BN. The third batch normalization layer BN outputs the feature map F.

[0097] The feature map E is input into the first downsampling layer, and the first downsampling layer outputs the feature map G.

[0098] Add feature map F and feature map G element by element to obtain feature map H;

[0099] The feature map H is input into the third ReLU activation function layer, and the third ReLU activation function layer outputs the feature map I.

[0100] Feature map I is sequentially input into the sixth convolutional layer, the fourth batch normalization layer (BN), the fourth ReLU activation function layer, the seventh convolutional layer, and the fifth batch normalization layer (BN). The fifth batch normalization layer (BN) outputs feature map J.

[0101] Feature map I is input into the second downsampling layer, and the second downsampling layer outputs feature map K;

[0102] Feature map K and feature map J are added element by element to obtain feature map L;

[0103] The feature map L is input to the fifth ReLU activation function layer, and the fifth ReLU activation function layer outputs the feature map M.

[0104] The other steps and parameters are the same as those in any of the specific implementation methods one to five.

[0105] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that: in 223, the feature map M is input to the manifold pyramid pooling layer, and the manifold pyramid pooling layer outputs the feature map R;

[0106] The specific process is as follows:

[0107] 2231. Based on feature map M and another 3D ALFF image (which can be of one person or not), the output feature map N of the lightweight residual layer is obtained. ; indicates as:

[0108]

[0109] in, ;

[0111] 2232. Feature map According to size Divided into One grid cell;

[0112] feature map According to size Divided into One grid cell;

[0113] feature map According to size Divided into One grid cell;

[0114] Will Each grid cell is input to the first pooling layer, and the output size of the first pooling layer is [value missing]. Feature map;

[0115] Will Each grid cell is input to the second pooling layer, and the output size of the second pooling layer is... Feature map;

[0116] Will Each grid cell is input to the third pooling layer, and the output size of the third pooling layer is... Feature map;

[0117] The output size of the first pooling layer is... The feature map and the output size of the second pooling layer are The feature map and the output size of the third pooling layer are The feature map is input into the fully connected layer, and the fully connected layer outputs a feature map R.

[0118] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0119] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that: in the 224, the feature map R is input to the classifier, and the classifier outputs the classification probability;

[0120] The specific process is as follows:

[0121] Flatten the feature map R to obtain the feature map S.

[0122] The feature map S is sequentially input into the first fully connected layer (output dimension 512), the second fully connected layer (output dimension 2), and the Softmax activation function layer. The Softmax activation function layer outputs the classification probability.

[0123] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0124] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that: the fMRI prediction module is used to obtain the fMRI of the target functional magnetic resonance imaging (fMRI); preprocessing the fMRI to obtain the preprocessed fMRI of the target functional magnetic resonance imaging (fMRI); processing the preprocessed fMRI of the target functional magnetic resonance imaging (fMRI) to obtain the mean-normalized ALFF low-frequency amplitude image of the target; inputting the mean-normalized ALFF low-frequency amplitude image of the target ALFF into the trained lightweight geometric manifold deep network LGMD-Net, the trained lightweight geometric manifold deep network LGMD-Net outputs the classification probability, the category being whether or not it is autism, the closer to 1 the higher the probability of autism, the closer to 0 the lower the probability of autism; the specific process is as follows:

[0125] 31. Obtain the functional magnetic resonance imaging (fMRI) image to be tested; preprocess the fMRI image to obtain the preprocessed low-frequency amplitude image (ALFF) to be tested; process the preprocessed ALFF to obtain the mean-normalized ALFF low-frequency amplitude image to be tested.

[0126] 32. Input the mean-normalized ALFF image of the test subject into the trained lightweight geometric manifold deep network LGMD-Net. The trained lightweight geometric manifold deep network LGMD-Net outputs the classification probability; the specific process is as follows:

[0127] 321. Input the mean-normalized ALFF image of the test image into the multi-channel shallow extraction layer. The multi-channel shallow extraction layer outputs a feature map. The specific process is as follows:

[0128] The mean-normalized ALFF image to be tested is input into the first convolutional layer, and the first convolutional layer outputs a feature map. ;

[0129] The mean-normalized ALFF image to be tested is input into the second convolutional layer, and the second convolutional layer outputs a feature map. ;

[0130] The mean-normalized ALFF image to be tested is input into the third convolutional layer, and the third convolutional layer outputs a feature map. ;

[0131] The kernel size of the first convolutional layer is 7×7, the stride is 2, and the padding is 3.

[0132] The second convolutional layer has a kernel size of 7×7, a stride of 2, and padding of 3.

[0133] The kernel size of the third convolutional layer is 7×7, the stride is 2, and the padding is 3.

[0134] feature map Feature map Feature map Input the first batch normalization layer (BN), and output the feature map. ;

[0135] feature map The first ReLU activation function layer is input, and the first ReLU activation function layer outputs a feature map. ;

[0136] 322. Feature map Input a lightweight residual layer; output a feature map from the lightweight residual layer. The specific process is as follows:

[0137] feature map The input layers are sequentially the fourth convolutional layer, the second batch normalization layer (BN), the second ReLU activation function layer, the fifth convolutional layer, and the third batch normalization layer (BN). The third batch normalization layer (BN) outputs the feature map. ;

[0138] feature map The first downsampling layer is input, and the first downsampling layer outputs a feature map. ;

[0139] feature map and feature map By adding elements one by one, we obtain the feature map. ;

[0140] feature map The input is a third ReLU activation function layer, and the output is a feature map. ;

[0141] Feature map The input layers are sequentially the sixth convolutional layer, the fourth batch normalization layer (BN), the fourth ReLU activation function layer, the seventh convolutional layer, and the fifth batch normalization layer (BN). The fifth batch normalization layer (BN) outputs the feature map. ;

[0142] Feature map The input is the second downsampling layer, and the output of the second downsampling layer is the feature map. ;

[0143] Feature map and feature map By adding elements one by one, we obtain the feature map. ;

[0144] Feature map The input is the fifth ReLU activation function layer, and the output is a feature map. ;

[0145] 323. Feature map According to size Divided into One grid cell;

[0146] feature map According to size Divided into One grid cell;

[0147] feature map According to size Divided into One grid cell;

[0148] Will Each grid cell is input to the first pooling layer, and the output size of the first pooling layer is [value missing]. Feature map;

[0149] Will Each grid cell is input to the second pooling layer, and the output size of the second pooling layer is... Feature map;

[0150] Will Each grid cell is input to the third pooling layer, and the output size of the third pooling layer is... Feature map;

[0151] The output size of the first pooling layer is... The feature map and the output size of the second pooling layer are The feature map and the output size of the third pooling layer are The feature map is input into the fully connected layer, and the fully connected layer outputs the feature map. ;

[0152] 324. Feature Map Input the classifier, and the classifier outputs the classification probability.

[0153] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0154] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that: the feature diagram in 324... The input is a classifier, and the classifier outputs the classification probability; the specific process is as follows:

[0155] For feature maps Flatten the image to obtain a feature map. ;

[0156] Feature map The inputs are sequentially processed by the first fully connected layer (output dimension 512), the second fully connected layer (output dimension 2), and the Softmax activation function layer. The Softmax activation function layer outputs the classification probability.

[0157] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0158] (1) Overall framework

[0159] The overall framework diagram of the Lightweight Geometric Manifold Deep Network (LGMD-Net) is as follows: Figure 1As shown, this study uses a residual network framework. The axial, coronal, and sagittal views of the ALFF index image from fMRI are input as three channels into three parallel convolutional networks for feature extraction, simultaneously extracting the three-dimensional information of the image. Then, a designed lightweight residual module is used to perform deep feature extraction on the features extracted by the convolutional blocks. Subsequently, the features extracted by the lightweight network are subjected to spatial pyramid pooling (SPP) dimensionality reduction and manifold mixing to preserve the manifold features of the geometric image. Finally, the features are passed through fully connected layers to downstream tasks.

[0160] (2) Lightweight multi-channel feature extraction module based on geometric prior

[0161] To address the issues of low computational efficiency, redundant feature extraction, and insufficient sample size associated with traditional 3D convolution in fMRI data, LGMD-Net constructs a two-stage lightweight multi-channel 2D residual network structure. This structure first performs preliminary feature extraction on the 3D brain map using a multi-channel convolution module, and then incorporates geometric priors from the medical image coordinate system to extract spatial 3D features, such as... Figure 2 As shown in the diagram. Specifically, the multi-channel geometrical convolutional network segments the 3D brain map along three directions: axial (xy-plane), coronal (yz-plane), and sagittal (xz-plane), resulting in several 2D images. Then, each slice's 2D image is treated as a channel, converting the 3D brain map data into a multi-channel 2D image format. The number of channels equals the total number of slices on the segmentation axes of the 3D brain map. In this way, the resulting multi-channel 2D brain map not only fully preserves the features of all voxels but also maintains, to some extent, the 3D spatial characteristics of the original data.

[0162] The parallel multi-channel 2D convolutional neural network consists of three parallel multi-channel 2D convolutional sub-networks, including batch normalization layers and ReLU layers. Each sub-network receives a multi-channel 2D image in one direction as input and performs convolution operations independently. The output features of all branches are concatenated into a unified one-dimensional feature vector, which is then normalized and activated before being fed into the subsequent network structure. This concatenation operation integrates information from three orthogonal planes, improving the model's overall ability to model 3D spatial features.

[0163] During feature extraction, two-dimensional convolution slides the window in two directions.

[0164] Then, a lightweight multi-channel residual structure is constructed in the deep feature extraction module, which can enhance the sensitivity of small sample data to spatial detail features and improve the efficiency and expressive power of high-dimensional fMRI feature extraction. The lightweight deep feature extraction network consists of multi-channel convolutional layers, residual structure layers, batch normalization layers, and ReLU activation layers, aiming to maintain low computational complexity while improving the model's expressive power. Figure 3 As shown, firstly, the input image is processed by a basic convolutional module consisting of multi-channel two-dimensional convolution, batch normalization, and ReLU activation, which effectively extracts local information such as low-level edges and textures. Subsequently, the feature map is sequentially input into two stacked Basic Block residual structure units. Each unit contains two sets of convolution-normalization-activation operations and a residual connection path, realizing efficient extraction of deep semantic features and cross-layer information fusion.

[0165] (3) Feature dimensionality reduction method for manifold constraints

[0166] Manifold Mixup performs data augmentation in the feature space, extending the sample mixing operation to the intermediate feature layers (i.e., the manifold space) of the neural network. By performing sample interpolation on feature representations at higher semantic levels, it can effectively enhance the generalization and discriminative performance of the model.

[0167] To address the limitation of traditional global pooling in losing local details, LGMD-Net employs both SPP and Manifold Mixup spatial regularization methods for optimization. This network introduces multi-scale pooling operations during the feature dimensionality reduction stage, concatenating outputs from different scales into a unified vector to comprehensively capture global and local multi-level geometric features. Furthermore, the method performs random linear interpolation on features in the manifold space, thereby generating enhanced samples online and smoothing decision boundaries, significantly improving the structural integrity of 3D features during dimensionality reduction and the model's generalization performance.

[0168] SPP structure such as Figure 4 As shown, this structure uses a pooling operation on the input in a manner similar to the Bag-of-Words model. A two-dimensional feature map of arbitrary size is input to the SPP layer, which sets up various spatial buckets. The spatial map is then sequentially divided into these buckets, and pooling is performed within each bucket. Finally, the pooling results from all buckets are arranged in order and combined into a feature vector of uniform size. Compared to the Bag-of-Words model, this method preserves the spatial information of the two-dimensional feature map by setting up local spatial buckets, achieving the fusion of local and global features and enhancing the ability to extract features at different scales.

[0169] SPP borrows the idea of ​​Spatial Pyramid Matching, dividing the entire feature map into different scales {1×1, 2×2, 4×4} sequentially. Each grid cell is used for partitioning. Pooling is performed independently within each grid cell to obtain local statistical features corresponding to that scale. By combining multiple scales, both global semantics and fine-grained regional information can be preserved.

[0170] At each scale Below, for shapes that are Adaptive pooling is performed on the feature maps, and the output is directly obtained. The result of the size calculation does not require manual calculation of the pooling kernel size and step size.

[0171] Then, the pooling results of each scale are flattened and concatenated along the channel dimension to obtain the length of the vector;

[0172] For each vector Its vector length is the same as that of the original image or feature map. This independence allows the entire network to accept inputs of any size, while eliminating the need to redesign weights for different sizes in subsequent fully connected layers.

[0173] Finally, they are spliced ​​together along the channel dimension.

[0174] This enables the fusion of multi-scale spatial information and ensures that the output dimension is fixed, facilitating subsequent fully connected layer processing.

[0175] The beneficial effects of the present invention are verified using the following embodiments:

[0176] Example 1:

[0177] The lightweight geometric manifold deep network (LGMD-Net) based on functional magnetic resonance imaging (fMRI) described in this invention can be implemented in various ways. A preferred embodiment is provided below to illustrate the technical solution of this invention in detail and demonstrate its feasibility and effectiveness. This embodiment was trained and tested on the public dataset ABIDE for the auxiliary diagnosis of autism spectrum disorder (ASD).

[0178] 1. Data preparation and preprocessing:

[0179] The dataset used was the ABIDE (Autism Brain Imaging Data Exchange) dataset, which includes ABIDE-I and ABIDE-II, collecting fMRI and phenotypic data from 2226 participants across 36 international sites. The experiment used a quality-checked subset of 1509 participants, including 673 ASD patients and 836 healthy controls (HC). Preprocessing was performed using the RESTplus package in MATLAB 2022, including removing the first ten time points, correcting time slices, head motion correction, spatial registration, segmentation, spatial standardization, spatial smoothing, de-linearization, and regression of covariates. Subsequently, time-series data for specific voxels were subjected to Fast Fourier Transform (FFT) to calculate their amplitude spectra. The mean ALFF value was obtained for all frequency-related amplitudes in the 0.01–0.08 Hz range, and the mean ALFF value for the whole brain was standardized to obtain the mALFF value.

[0180] 2. Network construction and training details:

[0181] The LGMD-Net constructed in this embodiment is implemented using the PyTorch framework, and its core structure and hyperparameter settings are as follows:

[0182] Multi-channel input construction: The 3D mALFF image is normalized and sliced ​​in the axial, coronal and sagittal directions respectively. 61, 61 and 61 two-dimensional slices are generated in each direction, and then recombined into 61, 61 and 61-channel two-dimensional images respectively (the normalized image sizes are 61×61, 61×61 and 61×61 respectively).

[0183] Parallel convolutional sub-networks: The sub-network structure is identical in each direction, each containing a convolutional layer with a kernel size of 7×7, a stride of 2, and padding of 3, followed by batch normalization (BN) and ReLU activation functions. The output feature map has 64 channels.

[0184] Lightweight Residual Module: The concatenated feature input consists of a lightweight residual module composed of two stacked Basic Blocks. In each Basic Block, the kernel size of the two convolution operations is 3×3, and the number of output channels is [64, 64] respectively.

[0185] Spatial Pyramid Pooling (SPP): The SPP module uses a three-level pyramid: {1×1, 2×2, 4×4} to perform adaptive max pooling on input feature maps of arbitrary size, and outputs a fixed-length feature vector (1+4+16=21 grids, and the output feature dimension is 21×number of channels).

[0186] Manifold Mixup: During training, features from either shallow (after the first convolutional layer) or deep (after the second residual block) layers of the network are randomly selected for interpolation. Interpolation coefficients. From the Beta distribution ( , Random sampling is performed within the range of )

[0187] Classifier: The SPP output is flattened and fed into two fully connected layers (intermediate dimension 512, output dimension 2), and finally the classification probability is obtained through the Softmax function.

[0188] Training settings: The model uses the Adam optimizer (initial learning rate 0.001, weight decay 1e-5), the loss function is cross-entropy loss, the training batch size is 32, the training is conducted for a total of 100 epochs, and the number of early stops is set to 15.

[0189] 3. Replacement options and optimized variants:

[0190] Lightweight Residual Module Replacement Structure: This invention can use either a standard residual block or a bottleneck structure for feature extraction. In scenarios with extremely limited computational resources, depthwise separable convolution can be used to replace the standard convolution in the basic block, further reducing the number of parameters and computational cost, but this may slightly sacrifice feature extraction capabilities.

[0191] Alternative solutions for feature dimensionality reduction and regularization: SPP can be replaced by Global Average Pooling (GAP) or Global Max Pooling (GMP) to greatly reduce model complexity, but this will result in the loss of multi-scale spatial information. If multi-scale capability needs to be preserved, weighted aggregation methods such as Attention Pooling can also be used. Manifold Mixup can be replaced by traditional input space Mixup or by applying other feature space enhancement strategies such as CutMix in the hidden layer, or by alternating between multiple enhancement methods during training to achieve different regularization effects.

[0192] Variations of multi-scale pooling and manifold mixing: The SPP scale can be adjusted according to the resolution of the input image and task requirements. For example, a denser scale (e.g., 8×8) can be added to capture more local features, or the scale can be reduced to reduce computational overhead. Manifold Mixup can be applied at variable layers. It can be applied to a fixed layer of the network (e.g., after the first residual block) or randomly selected from multiple candidate layers during training to enhance the diversity of regularization effects.

[0193] 4. Implementation Results:

[0194] This embodiment was tested on the ABIDE dataset. The results show that the LGMD-Net model of this specific implementation outperforms the baseline 3D-CNN, Transformer, and GNN models in terms of classification accuracy, F1 score, and AUC. The model has significantly fewer parameters than traditional 3D-CNN models, faster training and inference speeds, and lower memory usage, demonstrating its lightweight characteristics and engineering application potential.

[0195] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. An autism-assisted diagnostic system based on lightweight geometric manifold deep networks, characterized in that: The system includes: The training set acquisition module, the pre-trained lightweight geometric manifold deep network model LGMD-Net acquisition module, and the fMRI prediction module for the functional magnetic resonance imaging (fMRI) test are all included. The training set acquisition module is used to obtain functional magnetic resonance imaging (fMRI); the fMRI is preprocessed to obtain preprocessed fMRI; the preprocessed fMRI is further processed to obtain mean-normalized ALFF images, which are used as the training set. The module for obtaining the trained lightweight geometric manifold deep network model LGMD-Net is used to construct the lightweight geometric manifold deep network LGMD-Net. The lightweight geometric manifold deep network LGMD-Net is trained based on the mean-normalized ALFF images in the training set to obtain the trained lightweight geometric manifold deep network LGMD-Net. The fMRI prediction module is used to obtain the fMRI image of the target functional magnetic resonance imaging (fMRI). The fMRI image is preprocessed to obtain a preprocessed fMRI image of the target functional magnetic resonance imaging (fMRI). The preprocessed fMRI image of the target functional magnetic resonance imaging (fMRI) is further processed to obtain a mean-normalized ALFF image of the target functional magnetic resonance imaging (ALFF). The mean-normalized ALFF image of the target functional magnetic resonance imaging (ALFF) is then input into a trained lightweight geometric manifold deep network (LGMD-Net). The trained lightweight geometric manifold deep network (LGMD-Net) outputs a classification probability, with the category being whether or not the individual has autism.

2. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 1, characterized in that: The training set acquisition module is used to obtain functional magnetic resonance imaging (fMRI); preprocess the fMRI to obtain preprocessed fMRI; process the preprocessed fMRI to obtain mean-normalized ALFF images, and use the mean-normalized ALFF images as the training set. The specific process is as follows: 1) Use the RESTplus tool to systematically process functional magnetic resonance imaging (fMRI) images to obtain preprocessed fMRI images; the specific process is as follows: The systematic processing is as follows: removing fMRI from the first 10 time points, time-layer correction, head motion correction, registration, spatial smoothing, removal of linear trends, covariate regression analysis, and filtering. 2) The preprocessed functional magnetic resonance imaging (fMRI) images are processed to obtain mean-normalized ALFF images; the specific process is as follows: 21) The brain map of the preprocessed functional magnetic resonance imaging (fMRI) at one time point is cut into small blocks of 3mm×3mm×3mm, each block is 1 voxel, and each voxel corresponds to 1 BOLD signal at each time point in the preprocessed fMRI. The mean value of all BOLD signals corresponding to all time points of each voxel in the preprocessed fMRI is taken to obtain the BOLD temporal mean value of each voxel at each time point. 22) Subtract the mean BOLD time series value of the same voxel at each time point from the BOLD signal value of each voxel at each time point to obtain the mean BOLD time series value of each voxel at each time point. 23) Perform a Fast Fourier Transform on the BOLD time series value of each voxel after removing the mean at each time point to obtain the complex spectrum of each voxel at each time point after the Fast Fourier Transform. 24) Obtain the amplitude spectrum by taking the modulus of the complex spectral coefficients corresponding to each voxel at each time point. ; The amplitude spectrum that meets the requirements for each voxel at all time points The average value is used to obtain the ALFF value of each voxel. Amplitude spectrum that meets the requirements for: ; The ALFF image is a three-dimensional image in the axial, coronal, and sagittal directions; 25) Take the average ALFF value of the image for all voxels to obtain the average ALFF value of the image; Divide the ALFF value of each voxel by the mean ALFF value of the images to obtain the mALFF of each voxel in the whole brain; mALFF represents the ALFF image after mean normalization.

3. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 2, characterized in that: The pre-trained Lightweight Geometric Manifold Deep Network (LGMD-Net) acquisition module is used to construct the LGMD-Net (Lightweight Geometric Manifold Deep Network). The LGMD-Net is trained based on mean-normalized ALFF images from the training set to obtain the pre-trained LGMD-Net. The specific process is as follows:

21. Construct a lightweight geometric manifold deep network LGMD-Net; the specific process is as follows: The lightweight geometric manifold deep network LGMD-Net consists of a multi-channel shallow extraction layer, a lightweight residual layer, a manifold pyramid pooling layer, and a classifier. The multi-channel shallow extraction layer includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a first batch normalization layer (BN), and a first ReLU activation function layer; The kernel size of the first convolutional layer is Step size is 2, fill size is 3; The kernel size of the second convolutional layer is Step size is 2, fill size is 3; The kernel size of the third convolutional layer is Step size is 2, fill size is 3; The lightweight residual layer includes: a fourth convolutional layer, a second batch normalization layer (BN), a second ReLU activation function layer, a fifth convolutional layer, a third batch normalization layer (BN), a third ReLU activation function layer, a first downsampling layer, a sixth convolutional layer, a fourth batch normalization layer (BN), a fourth ReLU activation function layer, a seventh convolutional layer, a fifth batch normalization layer (BN), a fifth ReLU activation function layer, and a second downsampling layer; The manifold pyramid pooling layer includes: a first pooling layer, a second pooling layer, a third pooling layer, and a fully connected layer; The feature map size corresponding to the first pooling layer is [size missing]. ; The feature map size corresponding to the second pooling layer is [size missing]. ; The feature map size corresponding to the third pooling layer is [size missing]. ; The classifier includes: a flattening layer, a first fully connected layer, a second fully connected layer, and a Softmax activation function layer; 22. The lightweight geometric manifold deep network LGMD-Net is trained based on the mean-normalized ALFF images in the training set to obtain the trained lightweight geometric manifold deep network LGMD-Net.

4. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 3, characterized in that: In step 22, the lightweight geometric manifold deep network LGMD-Net is trained based on the mean-normalized ALFF images in the training set to obtain the trained lightweight geometric manifold deep network LGMD-Net; the specific process is as follows:

221. Input the mean-normalized ALFF images from the training set into the multi-channel shallow extraction layer, and the multi-channel shallow extraction layer outputs feature map E; 222. Input the feature map E into the lightweight residual layer, and the lightweight residual layer outputs the feature map M; 223. The feature map M is input to the manifold pyramid pooling layer, and the feature map R is output by the manifold pyramid pooling layer.

224. The feature map R is input into the classifier, and the classifier outputs the classification probability; 225. Repeat steps 221 to 224 until the maximum number of iterations is reached to obtain the trained lightweight geometric manifold deep network LGMD-Net.

5. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 4, characterized in that: In step 221, the ALFF image after mean normalization in the training set is input into the multi-channel shallow extraction layer, and the multi-channel shallow extraction layer outputs a feature map E; the specific process is as follows: 2211. Standardize the 3D ALFF images in the training set after mean normalization to obtain the standardized 2D image ALFF. The dimensions of the standardized 2D image ALFF are as follows: , , The specific process is as follows: The mean-normalized 3D ALFF images in the training set are sliced ​​in the axial, coronal, and sagittal directions respectively, generating 61 slices in the axial direction, 61 slices in the coronal direction, and 61 slices in the sagittal direction. Sixty-one slices were generated based on the axial direction, sixty-one slices were generated in the coronal direction, and sixty-one slices were generated in the sagittal direction, which were then reconstructed into axial and coronal sections, respectively. Two-dimensional images of the channels, axial and sagittal Two-dimensional images of the channels, coronal and sagittal Two-dimensional image of the channel; 2212. The axial coronal The two-dimensional image ALFF of the channel is input into the first convolutional layer, and the first convolutional layer outputs feature map A; axial sagittal The two-dimensional image ALFF of the channel is input into the second convolutional layer, and the second convolutional layer outputs feature map B; coronal sagittal The two-dimensional image ALFF of the channel is input into the third convolutional layer, and the third convolutional layer outputs the feature map C; The kernel size of the first convolutional layer is Step size is 2, fill size is 3; The kernel size of the second convolutional layer is Step size is 2, fill size is 3; The kernel size of the third convolutional layer is Step size is 2, fill size is 3; Feature map A, feature map B, and feature map C are input into the first batch normalization layer BN, and the first batch normalization layer BN outputs feature map D. The feature map D is input into the first ReLU activation function layer, and the first ReLU activation function layer outputs the feature map E.

6. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 5, characterized in that: In step 222, the feature map E is input into the lightweight residual layer, and the lightweight residual layer outputs the feature map M. The specific process is as follows: The feature map E is sequentially input into the fourth convolutional layer, the second batch normalization layer BN, the second ReLU activation function layer, the fifth convolutional layer, and the third batch normalization layer BN. The third batch normalization layer BN outputs the feature map F. The feature map E is input into the first downsampling layer, and the first downsampling layer outputs the feature map G. Add feature map F and feature map G element by element to obtain feature map H; The feature map H is input into the third ReLU activation function layer, and the third ReLU activation function layer outputs the feature map I. Feature map I is sequentially input into the sixth convolutional layer, the fourth batch normalization layer (BN), the fourth ReLU activation function layer, the seventh convolutional layer, and the fifth batch normalization layer (BN). The fifth batch normalization layer (BN) outputs feature map J. Feature map I is input into the second downsampling layer, and the second downsampling layer outputs feature map K; Feature map K and feature map J are added element by element to obtain feature map L; The feature map L is input to the fifth ReLU activation function layer, and the fifth ReLU activation function layer outputs the feature map M.

7. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 6, characterized in that: In step 223, the feature map M is input to the manifold pyramid pooling layer, and the manifold pyramid pooling layer outputs the feature map R; the specific process is as follows: 2231. Based on feature map M and the lightweight residual layer output feature map N corresponding to another 3D ALFF image, the feature map is obtained. ; indicates as: in, ; 2232、 feature map According to size Divided into One grid cell; feature map According to size Divided into One grid cell; feature map According to size Divided into One grid cell; Will Each grid cell is input to the first pooling layer, and the output size of the first pooling layer is [value missing]. Feature map; Will Each grid cell is input to the second pooling layer, and the output size of the second pooling layer is... Feature map; Will Each grid cell is input to the third pooling layer, and the output size of the third pooling layer is... Feature map; The output size of the first pooling layer is... The feature map and the output size of the second pooling layer are The feature map and the output size of the third pooling layer are The feature map is input into the fully connected layer, and the fully connected layer outputs a feature map R.

8. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 7, characterized in that: In the 224, the feature map R is input into the classifier, and the classifier outputs the classification probability. The specific process is as follows: Flatten the feature map R to obtain the feature map S. The feature map S is sequentially input into the first fully connected layer, the second fully connected layer, and the Softmax activation function layer. The Softmax activation function layer outputs the classification probability.

9. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 8, characterized in that: The fMRI prediction module is used to obtain the fMRI image to be tested; preprocess the fMRI image to obtain the preprocessed fMRI image to be tested; process the preprocessed fMRI image to obtain the mean-normalized ALFF image to be tested; input the mean-normalized ALFF image to be tested into the trained lightweight geometric manifold deep network LGMD-Net, and the trained lightweight geometric manifold deep network LGMD-Net outputs the classification probability, the category being whether or not it is autism; the specific process is as follows:

31. Obtain the functional magnetic resonance imaging (fMRI) image to be tested; preprocess the fMRI image to obtain the preprocessed ALFF image to be tested; The preprocessed ALFF image to be tested is processed to obtain the mean-normalized ALFF image to be tested.

32. Input the mean-normalized ALFF image to be tested into the trained lightweight geometric manifold deep network LGMD-Net. The trained lightweight geometric manifold deep network LGMD-Net outputs the classification probability. The specific process is as follows:

321. Input the mean-normalized ALFF image to be tested into the multi-channel shallow extraction layer. The multi-channel shallow extraction layer outputs a feature map. ; The specific process is as follows: The mean-normalized ALFF image to be tested is input into the first convolutional layer, and the first convolutional layer outputs a feature map. ; The mean-normalized ALFF image to be tested is input into the second convolutional layer, and the second convolutional layer outputs a feature map. ; The mean-normalized ALFF image to be tested is input into the third convolutional layer, and the third convolutional layer outputs a feature map. ; The kernel size of the first convolutional layer is 7×7, the stride is 2, and the padding is 3. The second convolutional layer has a kernel size of 7×7, a stride of 2, and padding of 3. The kernel size of the third convolutional layer is 7×7, the stride is 2, and the padding is 3. feature map Feature map Feature map Input the first batch normalization layer (BN), and output the feature map. ; feature map The first ReLU activation function layer is input, and the first ReLU activation function layer outputs a feature map. ; 322. Feature map Input a lightweight residual layer; output a feature map from the lightweight residual layer. ; The specific process is as follows: feature map The input layers are sequentially the fourth convolutional layer, the second batch normalization layer (BN), the second ReLU activation function layer, the fifth convolutional layer, and the third batch normalization layer (BN). The third batch normalization layer (BN) outputs the feature map. ; feature map The first downsampling layer is input, and the first downsampling layer outputs a feature map. ; feature map and feature map By adding elements one by one, we obtain the feature map. ; feature map The input is a third ReLU activation function layer, and the output is a feature map. ; Feature map The input layers are sequentially the sixth convolutional layer, the fourth batch normalization layer (BN), the fourth ReLU activation function layer, the seventh convolutional layer, and the fifth batch normalization layer (BN). The fifth batch normalization layer (BN) outputs the feature map. ; Feature map The input is the second downsampling layer, and the output of the second downsampling layer is the feature map. ; Feature map and feature map By adding elements one by one, we obtain the feature map. ; Feature map The input is the fifth ReLU activation function layer, and the output is a feature map. ; 323、 feature map According to size Divided into One grid cell; feature map According to size Divided into One grid cell; feature map According to size Divided into One grid cell; Will Each grid cell is input to the first pooling layer, and the output size of the first pooling layer is [value missing]. Feature map; Will Each grid cell is input to the second pooling layer, and the output size of the second pooling layer is... Feature map; Will Each grid cell is input to the third pooling layer, and the output size of the third pooling layer is... Feature map; The output size of the first pooling layer is... The feature map and the output size of the second pooling layer are The feature map and the output size of the third pooling layer are The feature map is input into the fully connected layer, and the fully connected layer outputs the feature map. ; 324. Feature Map Input the classifier, and the classifier outputs the classification probability.

10. The autism auxiliary diagnostic system based on lightweight geometric manifold deep networks according to claim 9, characterized in that: The 324 feature maps Input the classifier, and the classifier outputs the classification probability; The specific process is as follows: For feature maps Flatten the image to obtain a feature map. ; Feature map The first fully connected layer, the second fully connected layer, and the Softmax activation function layer are input sequentially. The Softmax activation function layer outputs the classification probability.

Images