A neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task-based perceptual reconstruction

Abstract

The present invention discloses a neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task-aware reconstruction. The system comprises an functional magnetic resonance image processing module, an attention-based signal masking module, a spatiotemporal feature extraction module, a risk prediction module, and a reconstruction module. The functional magnetic resonance image processing module preprocesses the collected images and accurately extracts time series signals from brain regions. In the risk prediction task, the spatiotemporal feature extraction module extracts features from the complete fMRI time series and generates an attention matrix. Based on the extracted features, the risk prediction module outputs the corresponding diagnosis results. In the task-aware reconstruction task, the signal masking module selectively masks fMRI signals from important brain regions based on the attention matrix. The spatiotemporal feature extraction module extracts features from the masked fMRI time series, and the reconstruction module reconstructs the complete signal based on the masked signal data. The present invention makes the reconstruction task-aware, enabling the model to capture true fMRI features and improve the accuracy of the risk prediction task.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction. Background Art

[0002] Neurodevelopmental disorders are primarily diseases associated with abnormal nervous system and brain function, with autism spectrum disorder (ASD) and Alzheimer's disease (AD) being common examples. While symptom-based risk prediction methods have been used in the clinical diagnosis of neurodevelopmental disorders, their subjective decision-making process often leads to inaccurate risk predictions, resulting in misdiagnosis and overdiagnosis of subjects. Functional magnetic resonance imaging (fMRI) is a non-invasive technique that reveals brain activity by measuring blood oxygenation level-dependent (BOLD) signals. Due to its extremely high spatial resolution, the common practice in the neuroimaging community is to use predefined three-dimensional atlases to map 4D voxel-level signals to 2D brain region-level signals. Researchers have begun employing supervised deep learning models to extract information from brain activity and predict neurodevelopmental disorder risk. These supervised models can be broadly categorized into two types: static and dynamic. Static models typically assume that functional interactions between brain regions remain constant and directly model the signals of brain regions or the functional connectivity (FC) derived from these signals. In contrast, dynamic models aim to explore changes and state transitions in brain regions over time.

[0003] However, these models, which rely solely on supervised learning, are unable to learn the underlying true representation of fMRI, resulting in poor performance in risk prediction tasks. Even with reduced spatial resolution, existing neuroimaging datasets still pose challenges due to their high dimensionality and insufficient sample size. In this case, prediction models often overfit, resulting in poor performance and severely limiting the potential for obtaining interpretable biomarkers. Masked autoencoders (MAEs), as a self-supervised learning framework, have achieved remarkable results in natural language processing (NLP) and computer vision (CV), and the trained large models have good generalization capabilities. Inspired by MAEs, a recent study used fMRI data to formulate a reconstruction task, which involves randomly selecting the signal of a ROI at different time periods, masking it, and then reconstructing it. Finally, the pre-trained model was fine-tuned on the risk prediction task, improving prediction accuracy by reusing the same data samples and their labels.

[0004] However, directly applying this learning framework to fMRI may lead to suboptimal results. First, random masking of fMRI data is ineffective. fMRI signals are more redundant than sentences or images, and random masking is unlikely to hide key information in the data unless the masking strategy is very aggressive. Therefore, using random masking for reconstruction results in a model that lacks robust representation learning capabilities. Second, reconstruction is not task-aware, ignoring an intriguing phenomenon in neuroimaging: brain regions exhibit differential contributions depending on the type of neurodevelopmental disorder. For example, the sensorimotor network (SMN) has the highest attention scores in autism spectrum disorder (ASD) diagnosis, suggesting that brain regions within the SMN play an important role in ASD diagnosis. Furthermore, many studies have found that individuals with cognitive impairment exhibit abnormalities in the default mode network. Based on these observations, reconstructing signals from important brain regions across different time periods should yield better performance in risk prediction tasks than reconstructing from randomly selected ROIs. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task-aware reconstruction, so that the reconstruction becomes task-aware, enabling the model to capture the true fMRI features and improve the accuracy of the neurodevelopmental disorder risk prediction task.

[0006] To solve the above technical problems, the present invention provides a neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task-perceived reconstruction, comprising: a functional magnetic resonance image processing module, an attention-based signal masking module, a spatiotemporal feature extraction module, a risk prediction module, and a reconstruction module; the functional magnetic resonance image processing module preprocesses the collected images and accurately extracts the time series signals of the brain regions; in the disease diagnosis task, the spatiotemporal feature extraction module extracts the features of the complete fMRI time series and generates an attention matrix; based on the extracted features, the risk prediction module is responsible for analyzing the extracted features and predicting the risk results; in the task-perceived reconstruction task, the signal masking module selectively masks the fMRI signals of important brain regions based on the attention matrix generated in the risk prediction; the spatiotemporal feature extraction module extracts the features of the masked fMRI time series, and the reconstruction module reconstructs the complete signal based on the masked signal data; during the training process of the entire system, the risk prediction task and the reconstruction task share the spatiotemporal feature extraction module, first executing the risk prediction task to generate the attention matrix, and then using the attention matrix to guide the reconstruction task to reconstruct the important brain region signals, and the reconstruction task in turn provides feedback to the risk prediction task to improve the accuracy of risk prediction.

[0007] Preferably, the functional magnetic resonance image processing module can preprocess four-dimensional functional magnetic resonance data, including an image preprocessing submodule, a time series extraction submodule and a sliding window submodule; the image preprocessing submodule is used to reduce noise in the original data; the time series extraction submodule maps the 4D data into a 2D regional time series through a predefined brain map; the sliding window submodule divides a complete time series into non-overlapping time periods.

[0008] Preferably, the image preprocessing submodule maps the 4D fMRI data into a 2D brain region time series, uses the corresponding region of interest in the brain atlas to accurately define each brain region node, and extracts the time series of each brain region from it; in order to eliminate the differences in the time dimension of different brain regions, the time series of each brain region will be standardized to obtain the brain region time series matrix N represents the number of brain regions, and M represents the length of the time series. In addition, a target category label y∈[0,...C-1] is assigned, where C represents the number of categories in the classification task. This label serves as a supervisory signal for the risk prediction task.

[0009] Preferably, the sliding window submodule uses a non-overlapping sliding window of length τ to The time dimension is divided into T time periods, where M = Tτ, thus obtaining a set

[0010] Preferably, the spatiotemporal feature extraction module extracts useful spatiotemporal features from fMRI. Its parameters are shared by the risk prediction task and the reconstruction task for subsequent reconstruction and risk prediction. It includes a spatial feature extraction submodule and a temporal feature extraction submodule. The spatial feature extraction submodule is used to capture the functional connectivity between brain regions. The temporal feature extraction submodule is used to capture the temporal dynamics of brain activity. Both submodules are composed of multiple Transformer encoders. The Transformer encoder has an L-layer multi-head attention module MHSA for modeling the spatial / temporal dependency between ROIs / time periods, which is expressed as follows:

[0011]

[0012] where Z l Represents the output of the lth layer, l is the index of the layer, Z 0 It represents the original input, || is the concatenation operator, and M is the number of heads. are all learnable parameters of the model, yes the size of the first dimension of

[0013] Both the risk prediction task and the reconstruction task use the spatiotemporal feature extraction module for feature extraction; for the risk prediction task, each fMRI segment X is first transformed through a linear layer. t Mapping to d spat dimensional space, and then use the spatial feature extraction submodule as input to perform spatial feature extraction, and the output will form a matrix Each row represents the features learned by each brain region in time period t; in addition to outputting the spatial features of each brain region, the spatial feature extraction submodule of risk prediction also generates an attention matrix for each time period Used to guide the signal masking module in the reconstruction task; the average pooling layer aggregates the feature vectors of all brain regions into one vector Then, the time feature extraction submodule maps the feature sets of all time periods to For a series of potential in

[0014] Preferably, for the reconstruction task, the input is the masked fMRI time segment Then the spatial feature extraction submodule takes it as input and outputs Then use the time feature extraction submodule to extract the spatial features of all time periods The set is mapped to a set of potential features

[0015] Preferably, in the reconstruction task, the input of the spatiotemporal feature extraction module is In the reconstruction task, the attention-based signal masking module will be activated, and the brain area important for the risk prediction task will be selectively selected according to the attention matrix output in the risk prediction task, and its signal will be masked, thereby outputting In the signal masking module, the core is its attention-guided masking strategy Ω, which can determine the set of important brain regions. In order to define the concept of important brain regions, the attention matrix generated by the forward propagation of the spatial feature extraction submodule in risk prediction is used. The attention matrix shows the weights between brain regions and reflects the importance of each brain region to a certain extent. For a time period X, t , the attention matrices generated by each layer and each head of the spatial feature extraction submodule are averaged, and finally we get At the same time, remove the self-loop, that is, set the value of the diagonal position to 0; use A t , calculate a weight vector as follows:

[0016]

[0017] Using α tTo quantify the importance of all brain regions in time period t; then, select α t The brain regions corresponding to the first k maximum values ​​in , where k is set to Mask ratio δ∈(0,1); the signals of these selected brain regions will be masked.

[0018] Preferably, since the same training data is used in each round of training, t The calculated set of important brain regions may remain unchanged between training rounds, which may lead to model overfitting, that is, the model may only remember to select the same set of brain regions from a specific sample; to solve this problem, this system adopts a random sampling strategy to ensure that in each training round, for all fMRI segments of each sample, the model will randomly explore which brain regions are identified as important and extract a random subset; randomly sample k important brain regions brain regions, where μ is a hyperparameter with a value range of 0<μ<1; in this way, more randomness can be introduced during the training process, thereby helping the model to have stronger learning ability; the masked signal Through Ω from and It is derived that The signal of the masked brain area is replaced by a mask mark, which is a dimension d spat A learnable parameter vector is used to indicate the presence of missing brain region signals.

[0019] Preferably, the risk prediction module first Perform average pooling to calculate the global fMRI feature h global , and then passes through two fully connected layers and RELU activation function to output the final classification results, as follows:

[0020]

[0021] is a learnable parameter, f represents the RELU activation function, and the predicted value Through the softmax function, we get the probability distribution p and use the cross entropy loss. for supervised training.

[0022] Preferably, the reconstruction module utilizes the spatiotemporal feature extraction module based on Generated latent representation sequence To reconstruct fMRI signals, the signal reconstruction module includes a decoder, which is also composed of multiple Transformer encoders. The decoder first gives Add position encoding, then use Transformer to process and finally output the reconstructed fMRI segment The labels for the reconstruction task are the original fMRI data In order to ensure the accuracy of reconstruction, the mean square error (MSE) of the signal of the masked brain area was calculated as follows:

[0023]

[0024] where Φ t is the set of brain regions masked in time period t, X t (i) is the tth fMRI segment of the ith brain region, and is the corresponding reconstructed fMRI segment; given that fMRI is a time series data with strong temporal correlation, the loss of the reconstructed unmasked part is also considered, as shown below:

[0025]

[0026] in Represents the set of brain regions, the final reconstruction loss yes and The weighted sum of Given, the final loss function is

[0027] The beneficial effects of the present invention are: the present invention adopts a new fMRI feature learning framework, which is different from the traditional masked autoencoder. The masked autoencoder randomly masks the signals of each brain region, then reconstructs the signals, and finally fine-tunes them in the downstream risk prediction task. However, the random masking and reconstruction tasks are separated from the risk prediction task. These two factors make it impossible to effectively capture the underlying true features. The present invention designs an attention-related masking module, which integrates the reconstruction task and the prediction task together, and can reconstruct the brain region signals related to downstream neurodevelopmental disorders, so that our model can further extract feature representations related to neurodevelopmental disorders and further improve the accuracy of neurodevelopmental disorder risk prediction. In addition, due to its excellent interpretability, it has become a powerful tool for revealing the internal mechanisms and dynamics of the brain, which is of great significance for in-depth understanding and research of brain functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the system structure of the present invention.

[0029] Figure 2 Schematic diagram of the system structure of the present invention.

[0030] Figure 3This is a schematic diagram of the visualization of important brain regions of the present invention.

[0031] Figure 4 Schematic diagram showing the visualization of the reconstruction results of the ADNI dataset of the present invention.

[0032] Figure 5 Schematic diagram showing the visualization of the reconstruction results of the ABIDE dataset of the present invention. DETAILED DESCRIPTION

[0033] like Figure 1 As shown, a brain disease detection system based on functional magnetic resonance imaging and task-aware reconstruction includes an fMRI processing module, an attention-based signal masking module, a spatiotemporal feature extraction module, and a reconstruction module. The fMRI processing module preprocesses the collected images and accurately extracts time series signals from brain regions. In the risk prediction task, the spatiotemporal feature extraction module extracts features from the complete fMRI time series and generates an attention matrix. Based on the extracted features, the risk prediction module outputs the corresponding diagnosis results. In the task-aware reconstruction task, the signal masking module selectively masks fMRI signals from important brain regions based on the attention matrix. The spatiotemporal feature extraction module extracts features from the masked fMRI time series, and the reconstruction module reconstructs the complete signal based on the masked signal data. The entire model uses a multi-task training method, allowing the reconstruction task to feed back into the neurodevelopmental disorder risk prediction task, improving the accuracy of neurodevelopmental disorder risk prediction.

[0034] The functional magnetic resonance imaging processing module consists of three submodules: image preprocessing submodule, time series extraction submodule, and sliding window submodule. The image preprocessing submodule first preprocesses the functional magnetic resonance data, including head motion correction and bandpass filtering operations to ensure the stability and accuracy of the data; the time series extraction submodule uses the brain regions in the corresponding brain region segmentation template to accurately define each brain region node and extract the time series data of each brain region from it. In addition, in order to eliminate the differences in the time dimension between different brain regions, the time series values ​​of each brain region will be standardized, and finally the brain region time series matrix will be obtained. N represents the number of brain regions, and M represents the length of the time series. In addition, a target category label y∈[0,...C-1] is assigned, where C represents the number of categories in the classification task. This label serves as a supervisory signal for the classification task. The sliding window submodule divides the matrix X into T time periods in the time dimension, which is achieved through a non-overlapping sliding window of size τ, thereby obtaining a set of subsequences.

[0035] like Figure 2As shown in Figure 3, the spatiotemporal feature extraction module extracts useful spatiotemporal features from fMRI. It consists of two submodules: the spatial feature extraction submodule and the temporal feature extraction submodule, and its parameters are shared by both the risk prediction and reconstruction tasks. The spatial feature extraction submodule and the temporal feature extraction submodule are used to capture the functional connectivity between brain regions and the temporal dynamics of brain activity, respectively. Both submodules are composed of multiple Transformer encoders.

[0036] For the risk prediction task, the input of the spatiotemporal feature extraction module is After mapping the spatial feature extraction submodule and the temporal feature extraction submodule, a series of potential features are output. (for risk prediction) and attention matrix (used for signal masking). For the reconstruction task, the attention-based signal masking module selectively masks the original signal data to obtain At this time, the input of the spatiotemporal feature extraction module is It also undergoes the mapping of the spatial feature extraction submodule and the temporal feature extraction submodule, outputting a series of potential features for signal reconstruction.

[0037] The attention-based signal masking module selectively masks the original signal. Its core is the attention-guided strategy Ω, namely This strategy can determine the set of important brain regions to be masked. In order to define the concept of important brain regions, the attention matrix generated by the forward propagation of the spatial feature extraction submodule in the risk prediction process is used. The attention matrix shows the weights between brain regions and reflects the importance of each brain region to a certain extent. t , by averaging the attention matrices generated by each head of each layer of the spatial feature extraction submodule, we finally get At the same time, remove the self-loop and set the value of the diagonal position to 0; calculate as follows:

[0038]

[0039] Using α t to quantify the importance of all brain regions at time t; then, select α t The brain regions corresponding to the first k maximum values ​​in , where k is set to Mask ratio δ∈(0,1); The signals of these selected brain regions will be masked. It can focus on the brain regions that are important for the classification task. Since the same training data is used in each round of training, the tThe calculated set of important brain regions may remain unchanged between training rounds. This may lead to model overfitting, that is, the model may only remember to select the same set of brain regions from a specific sample. To solve this problem, this system adopts a random sampling strategy to ensure that in each training round, for all fMRI segments of each sample, the model will randomly explore which brain regions are identified as important and extract a random subset. Random sampling among k important brain regions brain regions, where μ is a hyperparameter with a value range of 0<μ<1. In this way, more randomness can be introduced during the training process, thus helping the model to have stronger learning ability. Through Ω from and It is derived that The signal of the masked brain area is replaced by a mask mark, which is a dimension d spat A learnable parameter vector is used to indicate the presence of missing brain region signals.

[0040] The risk prediction module Perform average pooling to calculate the global fMRI feature h global After that, through two fully connected layers and RELU activation function, the final classification results are output as follows:

[0041]

[0042] is a learnable parameter, f represents the RELU activation function, and the predicted value Through the softmax function, we get the probability distribution p and use the cross entropy loss. for supervised training.

[0043] The reconstruction module uses the spatiotemporal feature extraction module based on Generated latent representation sequence To reconstruct fMRI signals, the signal reconstruction module includes a decoder composed of Transformer. The decoder first gives Add position encoding, then use Transformer to process and finally output the reconstructed fMRI segment The labels for the reconstruction task are the original fMRI data In order to ensure the accuracy of reconstruction, the mean square error (MSE) of the signal of the masked brain area was calculated as follows:

[0044]

[0045] where Φ t is the set of brain regions masked in time period t, Xt (i) is the tth fMRI segment of the ith brain region, and is the corresponding reconstructed fMRI segment; given that fMRI is a time series data with strong temporal correlation, the loss of the reconstructed unmasked part is also considered, as shown below:

[0046]

[0047] Where V represents the set of brain regions. The final reconstruction loss yes and The weighted sum of The final loss function is given as

[0048] Our model, TARDRL, conducted extensive experiments on the ABIDE autism classification dataset and the ADNI cognitive impairment classification dataset. The ABIDE dataset contains samples from 1,009 subjects, including 516 with autism spectrum disorder (ASD). We used this dataset to perform a two-class classification problem between normal controls and patients (NC vs ASD). The ADNI dataset includes 517 samples from 143 subjects, including 140 samples from 35 normal controls, 138 samples from 36 participants with early mild cognitive impairment (EMCI), 130 samples from 40 participants with late mild cognitive impairment (LMCI), and 109 samples from 32 participants with Alzheimer's disease (AD). For ABIDE and ADNI, we used the Craddock 200 and AAL 90 brain atlases, respectively, to define brain regions. We first compared performance with existing state-of-the-art methods, as shown in Table 1. The experimental results show that compared to methods trained using fully self-supervised learning, our method outperforms other methods on both static and dynamic datasets and in two metrics (risk prediction accuracy and AUROC). This demonstrates that methods based on self-supervised learning are indeed superior to supervised learning methods and shows that self-supervised learning can better enable the model to capture the underlying information of fMRI.

[0049] Table 1 Comparison of the method of the present invention and existing advanced methods (superscript * represents a method based on self-supervised learning)

[0050]

[0051] We also conducted ablation experiments to demonstrate the effectiveness of our method, as shown in Table 2. We further demonstrate the effectiveness of the attention-based signal masking module and once again emphasize the enhanced effect of introducing self-supervised learning (reconstruction). We designed three ablation models: DRL, TARDRL-Random, and TARDRL-Opp. DRL uses the same architecture for the classification task as the original model and is trained from scratch. TARDRL-Random disables the policy Ω, randomly selecting brain regions for masking and reconstruction. In addition, TARDRL-Opp uses the exact opposite policy of Ω, masking the signals of brain regions that are not important for predicting neurodevelopmental disorder risk. We observe that DRL performs worse than all other self-supervised learning methods except TARDRL-Opp. Furthermore, removing policy Ω or using its inverse policy results in a decrease in prediction accuracy. This emphasizes the importance of reconstructing the signals of important brain regions for the final task, enabling the model to learn more relevant information.

[0052] Table 2 Ablation experiment verification table

[0053]

Claims

1. A neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction, characterized in that: include: Functional MRI image processing module, attention-based signal masking module, spatiotemporal feature extraction module, risk prediction module and reconstruction module; the functional MRI image processing module pre-processes the collected images, accurately extracts the time series signals of the brain region, and divides them into multiple subsequences using a sliding window; in the neurodevelopmental disorder risk prediction task, the spatiotemporal feature extraction module extracts the features of the complete fMRI time series and generates an attention matrix; based on the extracted features, the risk prediction module is responsible for analyzing the extracted features and predicting the risk results; in the task perception reconstruction task, the signal masking module selectively masks the fMRI signals of important brain regions based on the attention matrix generated in the risk prediction; the spatiotemporal feature extraction module extracts the features of the masked fMRI time series again, and the reconstruction module reconstructs the complete signal based on the masked signal data; The spatiotemporal feature extraction module extracts useful spatiotemporal features from fMRI. Its parameters are shared by the risk prediction task and the reconstruction task for subsequent reconstruction and risk prediction. It includes a spatial feature extraction submodule and a temporal feature extraction submodule. The spatial feature extraction submodule is used to capture the functional connectivity between brain regions; The temporal feature extraction submodule is used to capture the temporal dynamics of brain activity. Both submodules are composed of multiple Transformer encoders. The Transformer encoder has an L-layer multi-head attention module MHSA, which is used to model the spatial / temporal dependencies between ROIs / time periods, as shown below: where Z l Represents the output of the lth layer, l is the index of the layer, Z 0 It represents the original input, || is the concatenation operator, and M is the number of heads. are all learnable parameters of the model, yes the size of the first dimension of Both the risk prediction task and the reconstruction task use the spatiotemporal feature extraction module for feature extraction; for the risk prediction task, each fMRI segment X is first transformed through a linear layer. t Mapping to d spat dimensional space, and then use the spatial feature extraction submodule as input to perform spatial feature extraction, and the output will form a matrix Among them, each row represents the features learned by each brain region in time period t; In addition to outputting the spatial features of each brain region, the spatial feature extraction submodule of risk prediction also generates an attention matrix for each time period. Used to guide the signal masking module in the reconstruction task; the average pooling layer aggregates the feature vectors of all brain regions into one vector Then, the time feature extraction submodule maps the feature sets of all time periods to For a series of potential in During the training process of the entire system, the risk prediction task and the reconstruction task share the spatiotemporal feature extraction module. The risk prediction task is first performed to generate an attention matrix, and the attention matrix is ​​used to guide the reconstruction task to reconstruct important brain area signals. The reconstruction task, in turn, provides feedback to the risk prediction task to improve the accuracy of risk prediction.

2. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 1, characterized in that: The functional magnetic resonance image processing module can preprocess four-dimensional functional magnetic resonance data, including an image preprocessing submodule, a time series extraction submodule, and a sliding window submodule; the image preprocessing submodule is used to reduce noise in the original data; The time series extraction submodule maps 4D data into 2D regional time series through a pre-defined brain map; the sliding window submodule divides a complete time series into non-overlapping time periods.

3. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 2, characterized in that: The image preprocessing submodule maps the 4D fMRI data into a 2D brain region time series, uses the corresponding region of interest in the brain atlas to accurately define each brain region node, and extracts the time series of each brain region from it; the time series of each brain region is standardized to obtain the brain region time series matrix N represents the number of brain regions, and M represents the length of the time series. In addition, a target category label y∈[0,...C-1] is assigned, where C represents the number of categories in the classification task. This label serves as a supervisory signal for the risk prediction task.

4. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 2, characterized in that: The sliding window submodule uses a non-overlapping sliding window of length τ to The time dimension is divided into T time periods, where M = Tτ, thus obtaining a set 5. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 1, characterized in that: For the reconstruction task, the input is the masked fMRI time segment. Then the spatial feature extraction submodule takes it as input and outputs Then use the time feature extraction submodule to extract the spatial features of all time periods The set is mapped to a set of potential features 6. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 1, characterized in that: In the reconstruction task, the input of the spatiotemporal feature extraction module is In the reconstruction task, the attention-based signal masking module will be activated, and the brain area important for the risk prediction task will be selectively selected according to the attention matrix output in the risk prediction task, and its signal will be masked, thereby outputting In the signal masking module, the attention-guided masking strategy Ω can determine the set of important brain regions. To define the concept of important brain regions, the attention matrix generated by the forward propagation of the spatial feature extraction submodule in risk prediction is used. The attention matrix shows the weights between brain regions and, to a certain extent, reflects the importance of each brain region. For a time period X t , the attention matrices generated by each layer and each head of the spatial feature extraction submodule are averaged, and finally we get At the same time, remove the self-exchange, that is, set the value of the diagonal position to 0; use A t , calculate a weight vector as follows: Using α t To quantify the importance of all brain regions in time period t; then, select α t The brain regions corresponding to the first k maximum values ​​in , where k is set to Mask ratio δ∈(0,1); the signals of these selected brain regions will be masked.

7. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 6, characterized in that: Since the same training data is used in each round of training, t The calculated set of important brain regions may remain unchanged between training rounds, leading to model overfitting, that is, the model may only remember to select the same set of brain regions from a specific sample; a random sampling strategy is adopted to ensure that in each training round, for all fMRI segments of each sample, the model will randomly explore which brain regions are identified as important and extract a random subset; random sampling is performed among k important brain regions brain regions, where μ is a hyperparameter with a value range of 0<μ<1; in this way, more randomness can be introduced during the training process, thereby helping the model to have stronger learning ability; the masked signal Through Ω from and It is derived that The signal of the masked brain area is replaced by a mask mark, which is a dimension d spat A learnable parameter vector is used to indicate the presence of missing brain region signals.

8. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 1, characterized in that: The risk prediction module first Perform average pooling to calculate the global fMRI feature h global , and then passes through two fully connected layers and RELU activation function to output the final classification results, as follows: is a learnable parameter, f represents the RELU activation function, and the predicted value Through the softmax function, we get the probability distribution p and use the cross entropy loss. to conduct supervised training.

9. The neurodevelopmental disorder risk prediction system based on functional magnetic resonance imaging and task perception reconstruction according to claim 1, characterized in that: The reconstruction module uses the spatiotemporal feature extraction module based on Generated latent representation sequence To reconstruct fMRI signals, the signal reconstruction module includes a decoder, which is also composed of multiple Transformer encoders. The decoder first gives Add position encoding, then use Transformer to process and finally output the reconstructed fMRI segment The labels for the reconstruction task are the original fMRI data In order to ensure the accuracy of reconstruction, the mean square error (MSE) of the signal of the masked brain area was calculated as follows: where Φ t is the set of brain regions masked in time period t, X t (i) is the tth fMRI segment of the ith brain region, and is the corresponding reconstructed fMRI segment; given that fMRI is a time series data with strong temporal correlation, the loss of the reconstructed unmasked part is also considered, as shown below: in Represents the set of brain regions, the final reconstruction loss yes and The weighted sum of Given, the final loss function is

Images