Transcranial focused ultrasound stimulation model optimization system based on artificial intelligence

Abstract

The invention provides a transcranial focused ultrasound stimulation model optimization system based on artificial intelligence, and relates to the field of artificial intelligence and machine learning. The system comprises a brain three-dimensional simulation modeling module which constructs an individualized brain tissue structure and acoustic characteristic model based on MRI data and CT data; the ultrasonic stimulation parameter optimization module is used for generating optimal stimulation parameters through a space-time semantic graph convolutional neural network and a reinforcement learning algorithm, and the biological feedback real-time regulation and control module is used for constructing a real-time regulation and control mechanism to dynamically adjust key parameters of transcranial focused ultrasonic stimulation; and the multi-modal stimulation effect prediction and personalized adaptation module is used for extracting multi-modal features through a multi-strategy fusion framework, processing uncertainty among modals by utilizing a Dempster-Shafer evidence theory, constructing a stimulation effect prediction model in combination with an adaptive weighting mechanism, and outputting a personalized stimulation scheme with high credibility. In this way, the accuracy and individuation degree of transcranial ultrasonic nerve regulation are improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence and machine learning, and particularly relates to a transcranial focused ultrasound stimulation model optimization system based on artificial intelligence. BACKGROUND

[0002] With the development of neuroscience and artificial intelligence technology, non-invasive brain stimulation technology is showing broad application prospects in the fields of neural regulation, disease intervention and brain-computer interface. As a new generation of neural regulation method, transcranial focused ultrasound stimulation has gradually become a hot research and clinical direction due to its strong penetration, high focusing accuracy and small side effects.

[0003] However, transcranial focused ultrasound stimulation still faces the following technical difficulties in practical application: Large individual differences: skull thickness, brain tissue structure, etc. have significant differences among different individuals, resulting in complex sound wave propagation path, focusing point offset, and reduced stimulation accuracy.

[0004] Parameter setting relies on experience: parameters such as ultrasound power, pulse frequency and duty cycle need to be manually adjusted relying on expert experience, and there is a lack of systematic and data-driven optimization method.

[0005] Stimulation results are difficult to predict and the regulation closed loop is insufficient: real-time feedback means for stimulation effect is limited, and neural response evaluation relies on offline analysis, which cannot dynamically adjust the stimulation strategy.

[0006] It is difficult to fuse multi-modal data: electrophysiological signals, image data and behavioral response information have heterogeneous structures and inconsistent time steps, and traditional methods are difficult to fuse and model.

[0007] Therefore, it is urgent to build a new intelligent system that integrates deep learning, reinforcement learning and multi-modal perception to realize individualized modeling, parameter adaptive regulation, response real-time prediction and stimulation closed-loop optimization of transcranial focused ultrasound stimulation. SUMMARY

[0008] The present disclosure provides a transcranial focused ultrasound stimulation model optimization system based on artificial intelligence, aiming to realize individualized modeling, parameter adaptive regulation, response real-time prediction and stimulation closed-loop optimization of transcranial focused ultrasound stimulation.

[0009] According to a first aspect of the present disclosure, a transcranial focused ultrasound stimulation model optimization system based on artificial intelligence is provided, comprising: A brain three-dimensional simulation modeling module is configured to construct an individualized brain tissue structure and acoustic characteristic model based on magnetic resonance imaging and computed tomography data, and output an ultrasound wave propagation path simulation result and a "brain region-acoustic response" mapping database. An ultrasound stimulation parameter optimization module is configured to generate optimal stimulation parameters adapted to individual neural responses based on the anatomical and acoustic data provided by the individualized brain tissue structure and acoustic property model, using a spatiotemporal semantic graph convolutional neural network and a reinforcement learning algorithm, wherein the spatiotemporal semantic graph convolutional neural network includes temporal modeling, spatial modeling, semantic modeling, and multimodal information fusion; The biofeedback real-time control module is used to build a real-time control mechanism based on biofeedback, dynamically adjust the key parameters of transcranial focused ultrasound stimulation, and complete the real-time calibration of the stimulation process; The multimodal stimulation effect prediction and personalized adaptation module is used to extract multimodal features through a multi-strategy fusion framework, use the Dempster-Shafer evidence theory to deal with inter-modal uncertainty, and combine the adaptive weighting mechanism to build a stimulation effect prediction model to output a highly reliable personalized stimulation plan.

[0010] As a preferred embodiment, the brain three-dimensional simulation modeling module includes: a data acquisition and preprocessing unit, configured to acquire magnetic resonance imaging and electronic computer tomography data, and preprocess the magnetic resonance imaging and electronic computer tomography data; The three-dimensional brain structure modeling and simulation unit is used to construct a finite element simulation grid model containing multi-layer tissue acoustic parameters based on preprocessed data using the finite element method, simulate the ultrasonic propagation path according to the time-domain finite difference method, and output the ultrasonic propagation path simulation results and the "brain area-acoustic response" mapping database.

[0011] As a preferred embodiment, in the spatiotemporal semantic graph convolutional neural network, time modeling adopts a gated recurrent unit combined with a time-aware attention mechanism, and the calculation method is: ; in, For nodes State encoding in the time dimension, is the current node, For Adjacent nodes, For the edge timestamp when it was created, is the edge feature, For all connected nodes The edge set of is a gated recurrent unit network, is the time decay weight.

[0012] As a preferred embodiment, the spatial modeling aggregates information of related areas around the stimulus area through a graph attention mechanism, and the weight coefficient is: ; in, The node features after space modeling, For nodes The set of adjacent nodes of is the weight matrix of the spatial attention layer, For nodes The feature representation of For nodes and adjacent nodes The attention weight of .

[0013] As a preferred embodiment, the semantic modeling enhances the model's ability to identify node categories through a semantic embedding mechanism. The calculation formula is: ; in, is the feature vector after semantic modeling, For nodes The original eigenvector of is the embedding representation of the node type, is the semantic projection matrix, It is a vector concatenation operation.

[0014] As a preferred embodiment, the hidden layer update formula of the multimodal information fusion in the spatiotemporal semantic graph convolutional neural network is: ; in, For the The nodes of the hidden layer are represented by are the transformation matrices of time, space and semantic directions respectively, is the bias term, is the activation function.

[0015] As a preferred embodiment, in the reinforcement learning algorithm, the reward function comprehensively considers the neuromodulation effect and biosafety indicators, the strategy learning adopts a deep Q network with a normalized advantage function, and the loss function is defined by the Lyapunov function: ; in, is the Lyapunov function at time step The total amount, are stability indicators in different dimensions.

[0016] As a preferred embodiment, the biofeedback real-time control module includes: A real-time physiological signal acquisition and preprocessing unit is used to acquire original physiological signals in real time and preprocess the acquired original physiological signals; Constructing a neural state recognition and deviation assessment model unit, which uses a deep learning algorithm to build a neural state recognition and deviation assessment model, inputs preprocessed real-time physiological signals into the model for signal feature extraction and analysis, and outputs a classification judgment of the current neural response state; a deviation judgment and parameter adjustment unit, configured to compare the classification judgment with a preset target state, calculate a response error, and adjust key parameters of transcranial focused ultrasound stimulation in real time using a preset control strategy based on the error; The real-time calibration and feedback loop unit is used to continuously collect physiological signals and repeat the above-mentioned neural state identification, deviation assessment and parameter adjustment process to form a real-time closed-loop feedback control mechanism.

[0017] As a preferred embodiment, in the multimodal stimulation effect prediction and personalized adaptation module, the multi-strategy fusion framework includes early fusion, late fusion and deep late fusion.

[0018] As a preferred embodiment, the multimodal stimulation effect prediction and personalized adaptation module further includes: The multi-task joint training unit is used to unify the multimodal decision fusion loss and the adaptive weighted prediction loss into a comprehensive loss function with adjustable weights, optimize the prediction accuracy and uncertainty modeling capabilities, and output a highly reliable personalized stimulation plan. The comprehensive loss function is: ; in, 、 is the weight coefficient, is the total loss of the decision-making layer, is the mean absolute error loss.

[0019] Compared with the prior art, the present disclosure achieves the following beneficial effects: (1) This paper constructs a spatiotemporal semantic graph convolutional neural network structure that integrates temporal modeling, spatial modeling, and semantic modeling. It uses temporal attention mechanism, graph attention mechanism, and node semantic embedding to achieve deep representation of brain region structure and functional state. It proposes a unified hidden layer update expression to fuse multi-source information into high-quality graph embedding, providing state support for subsequent reinforcement learning.

[0020] (2) This paper introduces Lyapunov stability theory into the reinforcement learning module. To address the problem that traditional gradient descent is unstable and prone to falling into local optimality, a deep Q-network policy weight update mechanism based on Lyapunov function is proposed. This mechanism can ensure asymptotic stability during the policy optimization process and effectively improve the convergence speed of the training process and the policy security.

[0021] (3) In response to the problems of inter-modal conflict and inconsistent trust in multimodal prediction, this paper proposes to introduce the Dempster-Shafer evidence fusion rule into the stimulus effect prediction process, combining the confidence and uncertainty of each modality to achieve a more robust joint decision. This method can effectively improve the credibility of stimulus response prediction and reduce the risk of misjudgment.

[0022] (4) This paper designs an integrated multi-task learning framework, which unifies the multimodal decision fusion loss and the adaptive weighted prediction loss into a loss function with adjustable weights. The overall performance of the model is improved through a joint optimization method, so that the system can improve the prediction accuracy while taking into account uncertainty control and individual adaptability.

[0023] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which: Figure 1 A schematic diagram of an artificial intelligence-based transcranial focused ultrasound stimulation model optimization system according to an embodiment of the present disclosure is shown; Figure 2 A schematic diagram of a three-dimensional brain simulation modeling module according to an embodiment of the present disclosure is shown; Figure 3 A schematic diagram of an ultrasound stimulation parameter optimization module according to an embodiment of the present disclosure is shown; Figure 4 A schematic diagram of a biofeedback real-time control module according to an embodiment of the present disclosure is shown; Figure 5 A schematic diagram of a multimodal stimulation effect prediction and personalized adaptation module according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0025] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0026] In addition, the term "and / or" in this document merely describes an association relationship of associated objects, which means that there can be three relationships, for example, A and / or B can represent three cases of A alone, A and B together, and B alone. In addition, the character " / " in this document generally represents an "or" relationship between the front and rear associated objects.

[0027] Embodiment one As Figure 1 The system adopts a deep learning and data-driven method, combines brain tissue acoustic characteristic modeling, realizes intelligent optimization of stimulation parameters, and specifically includes: The brain three-dimensional simulation modeling module 100 is used for constructing an individualized brain tissue structure and acoustic characteristic model based on magnetic resonance imaging and computed tomography data, outputting an ultrasonic wave propagation path simulation result and a "brain region-acoustic response" mapping database; The ultrasonic stimulation parameter optimization module 200 is used for generating optimal stimulation parameters adapted to individual neural responses through a space-time semantic graph convolutional neural network and a reinforcement learning algorithm according to anatomical and acoustic data provided by the individualized brain tissue structure and acoustic characteristic model, wherein the space-time semantic graph convolutional neural network includes time modeling, space modeling, semantic modeling and multi-modal information fusion; The biological feedback real-time regulation module 300 is used for constructing a real-time regulation mechanism based on biological feedback, dynamically adjusting key parameters of transcranial focused ultrasound stimulation, and completing real-time calibration of the stimulation process; The multi-modal stimulation effect prediction and individualized adaptation module 400 is used for extracting multi-modal features through a multi-strategy fusion framework, processing inter-modal uncertainty by using Dempster-Shafer evidence theory, constructing a stimulation effect prediction model in combination with an adaptive weighting mechanism, and outputting a high-credibility individualized stimulation scheme.

[0028] As a preferred embodiment, the disclosure takes the brain three-dimensional simulation modeling module as a core link, constructs an individualized brain tissue structure and acoustic characteristic model based on fusion of high-resolution magnetic resonance imaging (MRI) and computed tomography (CT) data, provides accurate anatomical and propagation basis for downstream modules, and on this basis, the system runs the ultrasonic stimulation parameter optimization module 200, the biological feedback real-time regulation module 300 and the multi-modal stimulation effect prediction and individualized adaptation module 400 in parallel, realizes full-process intelligent modeling and individualized optimization of transcranial focused ultrasound stimulation path, parameters and reactions, and significantly improves the clinical adaptability and regulation precision of the system.

[0029] As a preferred embodiment, the brain three-dimensional simulation modeling module 100 is designed to construct a highly individualized and anatomically precise three-dimensional brain simulation model to provide basic data support for transcranial focused ultrasound stimulation path optimization and parameter control. The model uses high-resolution magnetic resonance imaging (MRI) and electronic computed tomography (CT) data as input, integrates medical image processing and acoustic simulation technology, and accurately simulates the propagation behavior of ultrasound in different brain tissues and their spatial sound pressure response characteristics, such as Figure 2 As shown, the brain three-dimensional simulation modeling module 100 specifically includes: The data acquisition and preprocessing unit 101 is used to acquire magnetic resonance imaging and electronic computer tomography data and preprocess the magnetic resonance imaging and electronic computer tomography data. The preprocessing of the magnetic resonance imaging and electronic computer tomography data includes: (1) Computed tomography (CT) data acquisition and registration: CT images provide information about the thickness, density, and structural differences of the skull, which significantly affects the attenuation and scattering of ultrasound. A mutual information-based rigid registration algorithm (such as the Elastix toolkit) can be used to align CT and MRI images to the MNI standard space to ensure that the spatial coordinates of the skull and brain regions are consistent. The registration accuracy is verified by normalized mutual information (NMI>0.95) and root mean square error (RMSE<0.3mm), ultimately achieving spatial matching of bone tissue and brain tissue.

[0030] (2) Data fusion and grid processing: After fusing MRI and CT information, a tissue label map is constructed in a unified anatomical coordinate system, and a mesh model suitable for finite element simulation is generated to ensure that the acoustic parameters (sound velocity, density, attenuation coefficient) of each type of tissue can be accurately mapped.

[0031] The three-dimensional brain structure modeling and simulation unit 102 is used to construct a finite element simulation grid model containing multi-layer tissue acoustic parameters based on the preprocessed data using the finite element method, simulate the ultrasonic propagation path according to the time-domain finite difference method, and output the ultrasonic propagation path simulation results and the "brain area-acoustic response" mapping database.

[0032] Specifically, the establishment of the high-precision three-dimensional brain model based on the fusion data includes multiple layers of tissue structures: skin, skull, meninges, gray matter, white matter, and brain ventricles, etc. Each layer of structure has an independent acoustic parameter setting, reflecting its true biophysical properties. The finite element method or k-Wave toolbox and other professional simulation platforms are used to simulate the propagation path of ultrasonic waves in the brain's multi-tissue structure. The offset vector of the actual focal point relative to the design target point is extracted from the simulation results to evaluate the spatial error existing in the penetration path, providing a reference basis for subsequent compensation algorithms. Finally, the simulation results of multiple individuals under different stimulation parameters (such as power 0.1-1.0 W / cm 2 , pulse frequency 50-500Hz, incident angle 0-30°) are stored in a structured manner to build a "brain region-acoustic response" mapping database, with each record associated with brain region ID, sound pressure value, focusing error, tissue attenuation coefficient, etc. to provide data support for subsequent algorithms.

[0033] As a preferred embodiment, the ultrasonic stimulation parameter optimization module 200 is used to build a graph neural network model for spatio-temporal-semantic modeling of the structure and functional state of the brain stimulation region, as shown in Figure 3 , the module specifically includes: Time modeling unit 201: Wherein, the time modeling adopts a gated recurrent unit combined with a time-aware attention mechanism to dynamically capture the influence of historical states on the current stimulation response, and its calculation method is: ; Wherein, is the state encoding of node in the time dimension, is the current node, is the node adjacent to , is the timestamp formed by the edge , is the feature of the edge, is the set of edges connected to node , is the gated recurrent unit network, is the time decay weight.

[0034] ; Wherein, is the time decay coefficient, is the current time point, is the timestamp of the historical edge, is the historical information, the farther the time, the smaller the weight, represents the th historical edge in the set of historical edges of node The timestamp of the edge, For the Feature information of historical edges.

[0035] Spatial Modeling Unit 202: Spatial modeling aggregates information about related areas around the stimulus area through the graph attention mechanism, with the weight coefficient being: ; in, The node features after space modeling, For nodes The set of adjacent nodes of is the weight matrix of the spatial attention layer, For nodes The feature representation of For nodes and adjacent nodes The attention weight of (see the following formula):

[0036] ; in, : trainable attention weight vector, : vector concatenation operation, is the activation function with leakage correction, For nodes The feature representation of For nodes The feature representation of For nodes The feature representation of For nodes The set of adjacent nodes of Representation node The first node in the adjacent node set of nodes.

[0037] Semantic Modeling Unit 203: Semantic modeling enhances the model's ability to identify node categories through semantic embedding mechanisms: ; in, is the feature vector after semantic modeling, For nodes The original eigenvector of is the embedding representation of the node type, is the semantic projection matrix, It is a vector concatenation operation.

[0038] Multimodal fusion unit 204: The hidden layer update formula of the three multimodal information in the spatiotemporal semantic graph convolutional neural network is expressed as follows: ; in, For the The nodes of the hidden layer are represented by are the transformation matrices of time, space and semantic directions respectively, is the bias term, is the activation function.

[0039] Reinforcement Learning Unit 205: The system transforms the task of optimizing stimulation parameters into a Markov decision process problem in reinforcement learning. The state space is represented by a multidimensional graph embedding output by a spatiotemporal semantic graph convolutional neural network. The action space includes adjustable combinations of parameters such as stimulation power, pulse repetition frequency, duty cycle, and focal position. The reward function comprehensively considers the effects of neural regulation (such as neural activation levels in electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS)) and biosafety indicators. Policy learning uses a deep Q-network with a normalized advantage function, and introduces the Lyapunov stability criterion to ensure the robustness of the controller. Its loss function is defined by the Lyapunov function: ; in, is the Lyapunov function at time step The total amount, are stability indicators in different dimensions.

[0040] Among them, the parameter update formula of the Q network is replaced by the Lyapunov gradient descent update formula: ; in, is the current strategy parameter, is the learning rate, is the Lyapunov equilibrium coefficient, is the deviation error between the current state and the expected state, is the gradient of the error with respect to the current parameter, are the updated policy parameters.

[0041] Proximal policy optimization is used to make policy updates smoother during intensive training. Under the framework of partially observable Markov games, the intelligent agent updates its strategy through environmental interaction sampling, effectively balancing the matching degree between stimulus intensity and neural responses in the target area.

[0042] As a preferred embodiment, the biofeedback real-time control module 300 is used to continuously collect and analyze brain physiological signals, dynamically optimize the parameter settings of transcranial focused ultrasound stimulation, use artificial intelligence models to judge the neural response state, and continuously self-correct during the stimulation process to achieve an individualized, precise, and dynamic neural control strategy. Figure 4 As shown, this module specifically includes: The real-time physiological signal acquisition and preprocessing unit 301 is used to acquire original physiological signals in real time and preprocess the acquired original physiological signals. The physiological signals used in this part include: Electroencephalogram (EEG): EEG can reflect the synchronous discharge behavior of neuronal groups with high temporal resolution and is a direct indicator of the brain's electrical activity state; Functional near-infrared spectroscopy (fNIRS): By measuring changes in the concentrations of oxyhemoglobin and deoxyhemoglobin, fNIRS can reflect the metabolic activity and hemodynamic response of local brain regions and is suitable for evaluating the response of cortical regions to stimulation.

[0043] Heart rate variability, galvanic skin response, eye movements, etc. can serve as supplementary basis for judging neural status, helping to comprehensively evaluate the subject's arousal, emotional state and stimulus acceptance.

[0044] A neural state recognition and deviation assessment model building unit 302 is configured to use a deep learning algorithm to build a neural state recognition and deviation assessment model, input the preprocessed real-time physiological signal into the model for signal feature extraction and analysis, and output a classification judgment of the current neural response state (e.g., "underactivation," "overactivation," or "normal response"); The deviation judgment and parameter adjustment unit 303 is used to compare the classification judgment with the preset target state, calculate the response error, and make real-time adjustments based on the error using a preset multi-level adjustment strategy. The preset multi-level adjustment strategy specifically includes: Level 1 adjustment (mild deviation): fine-tune the pulse repetition frequency (±10Hz) or duty cycle (±5%); Secondary adjustment (moderate deviation): switch stimulation mode, including: sound power: adjust stimulation energy; pulse repetition frequency: control stimulation rhythm; stimulation duration: determine the duration of a single stimulation; focal point position: dynamically fine-tune the stimulation target position; incident angle: optimize the sound beam path to avoid bone impedance, etc.

[0045] Level 3 adjustment (severe deviation): Pause stimulation and prompt a safety alarm, which requires manual confirmation before restarting.

[0046] If EEG alpha waves drop significantly, potentially indicating brain overexcitation, the system will reduce power and shorten stimulation duration to prevent neural overactivation. If fNIRS detects local blood oxygen levels falling short of expectations, it will automatically increase the pulse repetition frequency or adjust the target area. The control model continuously operates in a closed loop of "perception-judgment-control-verification," continuously narrowing the gap between expected stimulation effects and actual responses, thereby enhancing the system's robustness and individual adaptability.

[0047] The real-time calibration and feedback loop unit 304 is used to continuously collect physiological signals and repeat the above-mentioned neural state identification, deviation assessment and parameter adjustment process to form a real-time closed-loop feedback control mechanism.

[0048] As a preferred embodiment, the multimodal stimulation effect prediction and personalized adaptation module 400 is used to build a stimulation effect prediction model through multimodal data fusion to achieve high reliability and personalized adaptation of the stimulation scheme, such as Figure 5 As shown, this module specifically includes: Multi-strategy fusion framework construction unit 401: This preprocesses input data using a multimodal encoder, mapping continuous time series signals and structured text information to a unified latent dimension. After adding a learnable positional encoding, it then extracts sequence fragments using a sliding window mechanism. These are then fed separately or jointly into a standard Transformer encoding module to form a multimodal feature representation.

[0049] For multimodal information fusion, this system introduces a multi-strategy fusion framework, including early fusion, late fusion and deep late fusion: Early fusion: After the transformed multimodal data is spliced, it is uniformly input into the Transformer module for processing; Late fusion: Each modality is encoded by an independent Transformer, and then concatenated after feature extraction; Deep Late Fusion: Based on late fusion, an additional Transformer layer is introduced to further deeply encode the splicing features.

[0050] The outputs of each fusion strategy will be spliced ​​along the sequence dimension and compressed by the bottleneck layer and filtered by the gating mechanism before being input into the alarm decoder to predict whether the stimulus effect is abnormal and locate its cause label.

[0051] Multimodal decision uncertainty estimation unit 402: This unit combines the prediction results and uncertainties of different modalities based on the Dempster-Shafer evidence theory. Suppose there are three belief functions for speech, image, and text 、 and , then the joint distribution after fusion is: ; ; ; in, After fusion, it belongs to The trustworthiness of the class, They are the first, second and third modes respectively. The trustworthiness of the class, are the uncertainties of the first, second and third modes respectively, c is the conflict between the three belief functions, 1-c is the normalization factor of the conflict, and u is the overall uncertainty after fusion.

[0052] The loss function used during training : ; in, is the total loss of the decision layer, λ is the hyperparameter for adjusting the weight between KL divergence and cross entropy, is the KL divergence.

[0053] Cross Entropy Loss The expression is: ; in, is the true label value of the jth class of the i-th sample (0 or 1), is the probability that the i-th sample is predicted to be of the j-th class.

[0054] KL divergence term The expression is: ; Where KL is the Kullback-Leibler divergence function; is the distribution calculated according to the adjusted Dirichlet distribution; is an ideal (uniform or deterministic) Dirichlet distribution; is the target distribution parameter corresponding to the i-th target, integrating the true label and the predicted credibility.

[0055] Adaptive weighted integration modal feature unit 403: This unit assigns weights to the uncertainties of the three modalities of speech, image, and text based on subjective logic theory, and integrates modal features in a weighted manner: ; Among them, U is the fused feature vector, is the original feature extracted from the nth mode, is the weighted modal feature, To represent three modalities: speech, image, and text.

[0056] The final prediction output is: ; in, is the final prediction result, Linear is the fully connected linear layer, ReLU is the rectified linear unit activation function, is the fused feature vector.

[0057] This module is trained using mean absolute error loss. : ; in, is the predicted value of the i-th sample, is the true label of the i-th sample, and N is the total number of samples.

[0058] Multi-task joint training unit 404: It is used to unify the multimodal decision fusion loss and the adaptive weighted prediction loss into a comprehensive loss function with adjustable weights, optimize the prediction accuracy and uncertainty modeling capabilities, and output a highly reliable personalized stimulation plan. as follows: ; In this embodiment, The value is 0.98, The value is 0.02, is the total loss of the decision-making layer, is the mean absolute error loss.

[0059] The overall optimization strategy simultaneously optimizes prediction accuracy and uncertainty modeling capabilities during backpropagation. Through this design, the system described in the present invention can model and predict brain responses in different individuals and under different conditions, generating personalized, responsive transcranial focused ultrasound stimulation protocols. This significantly improves the stability and adaptability of the stimulation effect, particularly in scenarios such as non-invasive treatment of neurological diseases such as Parkinson's disease and depression, brain-computer interface adaptation and debugging, and rehabilitation training, thereby enhancing the clinical practical value of transcranial focused ultrasound stimulation.

[0060] According to the above embodiments of the present disclosure, the following technical effects are achieved: (1) This paper constructs a spatiotemporal semantic graph convolutional neural network structure that integrates temporal modeling, spatial modeling, and semantic modeling. It uses temporal attention mechanism, graph attention mechanism, and node semantic embedding to achieve deep representation of brain region structure and functional state. It proposes a unified hidden layer update expression to fuse multi-source information into high-quality graph embedding, providing state support for subsequent reinforcement learning.

[0061] (2) This paper introduces Lyapunov stability theory into the reinforcement learning module. To address the problem that traditional gradient descent is unstable and prone to falling into local optimality, a deep Q-network policy weight update mechanism based on Lyapunov function is proposed. This mechanism can ensure asymptotic stability during the policy optimization process and effectively improve the convergence speed of the training process and the policy security.

[0062] (3) In response to the problems of inter-modal conflict and inconsistent trust in multimodal prediction, this paper proposes to introduce the Dempster-Shafer evidence fusion rule into the stimulus effect prediction process, combining the confidence and uncertainty of each modality to achieve a more robust joint decision. This method can effectively improve the credibility of stimulus response prediction and reduce the risk of misjudgment.

[0063] (4) This paper designs an integrated multi-task learning framework, which unifies the multimodal decision fusion loss and the adaptive weighted prediction loss into an objective function with adjustable weights. The overall performance of the model is improved through a joint optimization method, so that the system can improve the prediction accuracy while taking into account uncertainty control and individual adaptability.

[0064] In the technical solutions disclosed herein, the acquisition, storage, and application of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0065] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions of this disclosure can be achieved, and this document is not limited here.

[0066] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. An artificial intelligence-based transcranial focused ultrasound stimulation model optimization system, characterized in that: include: A 3D brain simulation modeling module, which builds a personalized model of brain tissue structure and acoustic properties based on magnetic resonance imaging and computed tomography data, and outputs ultrasound propagation path simulation results and a "brain region-acoustic response" mapping database; An ultrasound stimulation parameter optimization module is configured to generate optimal stimulation parameters adapted to individual neural responses based on the anatomical and acoustic data provided by the individualized brain tissue structure and acoustic property model, using a spatiotemporal semantic graph convolutional neural network and a reinforcement learning algorithm, wherein the spatiotemporal semantic graph convolutional neural network includes temporal modeling, spatial modeling, semantic modeling, and multimodal information fusion; The biofeedback real-time control module is used to build a real-time control mechanism based on biofeedback, dynamically adjust the key parameters of transcranial focused ultrasound stimulation, and complete the real-time calibration of the stimulation process; The multimodal stimulation effect prediction and personalized adaptation module is used to extract multimodal features through a multi-strategy fusion framework, use the Dempster-Shafer evidence theory to deal with inter-modal uncertainty, and combine the adaptive weighting mechanism to build a stimulation effect prediction model to output a highly reliable personalized stimulation plan.

2. The system according to claim 1, wherein: The brain three-dimensional simulation modeling module includes: A data acquisition and preprocessing unit, configured to acquire magnetic resonance imaging and electronic computer tomography data, and preprocess the magnetic resonance imaging and electronic computer tomography data; The three-dimensional brain structure modeling and simulation unit is used to construct a finite element simulation mesh model containing multi-layer tissue acoustic parameters based on preprocessed data using the finite element method, simulate the ultrasonic propagation path according to the time-domain finite difference method, and output the ultrasonic propagation path simulation results and the "brain area-acoustic response" mapping database.

3. The system according to claim 1, wherein: In the spatiotemporal semantic graph convolutional neural network, time modeling uses a gated recurrent unit combined with a time-aware attention mechanism, and the calculation method is: ； in, For nodes State encoding in the time dimension, is the current node, For Adjacent nodes, For the edge timestamp when it was created, is the edge feature, For all connected nodes The edge set of is a gated recurrent unit network, is the time decay weight.

4. The system according to claim 3, characterized in that The spatial modeling aggregates information of related areas around the stimulus area through the graph attention mechanism, and the weight coefficient is: ； in, The node features after space modeling, For nodes The set of adjacent nodes of is the weight matrix of the spatial attention layer, For nodes The feature representation of For nodes and adjacent nodes The attention weight of .

5. The system according to claim 4, characterized in that The semantic modeling enhances the model's ability to identify node categories through a semantic embedding mechanism. The calculation formula is: ； in, is the feature vector after semantic modeling, For nodes The original eigenvector of is the embedding representation of the node type, is the semantic projection matrix, It is a vector concatenation operation.

6. The system according to claim 4, characterized in that The hidden layer update formula of the multimodal information fusion in the spatiotemporal semantic graph convolutional neural network is: ； in, For the The nodes of the hidden layer are represented by are the transformation matrices of time, space and semantic directions respectively, is the bias term, is the activation function.

7. The system according to claim 1, wherein: In the reinforcement learning algorithm, the reward function comprehensively considers the neuromodulation effect and biosafety indicators. The strategy learning adopts a deep Q network with a normalized advantage function. The loss function is defined by the Lyapunov function: ； in, is the Lyapunov function at time step The total amount, are stability indicators in different dimensions.

8. The system according to claim 1, wherein: The biofeedback real-time control module includes: A real-time physiological signal acquisition and preprocessing unit is used to acquire original physiological signals in real time and preprocess the acquired original physiological signals; Constructing a neural state recognition and deviation assessment model unit, which uses a deep learning algorithm to build a neural state recognition and deviation assessment model, inputs preprocessed real-time physiological signals into the model for signal feature extraction and analysis, and outputs a classification judgment of the current neural response state; a deviation judgment and parameter adjustment unit, configured to compare the classification judgment with a preset target state, calculate a response error, and adjust key parameters of transcranial focused ultrasound stimulation in real time using a preset control strategy based on the error; The real-time calibration and feedback loop unit is used to continuously collect physiological signals and repeat the above-mentioned neural state identification, deviation assessment and parameter adjustment process to form a real-time closed-loop feedback control mechanism.

9. The system according to claim 1, wherein: In the multimodal stimulation effect prediction and personalized adaptation module, the multi-strategy fusion framework includes early fusion, late fusion and deep late fusion.

10. The system according to claim 1, wherein: The multimodal stimulation effect prediction and personalized adaptation module further includes: The multi-task joint training unit is used to unify the multimodal decision fusion loss and the adaptive weighted prediction loss into a comprehensive loss function with adjustable weights, optimize the prediction accuracy and uncertainty modeling capabilities, and output a highly reliable personalized stimulation plan. The comprehensive loss function is: ； in, 、 is the weight coefficient, is the total loss of the decision-making layer, is the mean absolute error loss.

Images