An intelligent pulp regeneration regulation method based on targeted outer vesicles

Abstract

The application discloses an intelligent pulp regeneration regulation method based on targeted exovesicles, comprising the following steps: generating a root canal center line prior graph according to a CBCT image, combining a dual-channel fusion image to form network input data, inputting the network input data into a 3D U-Net++ network for segmentation, and outputting a pulp cavity binary mask and a pulp damage probability graph; constructing a functional partition graph according to the output result and inputting the functional partition graph into a type perception graph neural network for information aggregation and feature learning, obtaining the final feature representation of a node, and predicting a pulp regeneration factor demand vector by using a partition multi-task regression head; generating a targeted exovesicle formula parameter and a targeted ligand amino acid sequence by a double-head hybrid action near-end strategy optimization agent based on the regeneration factor demand vector; treating a patient based on a targeted exovesicle delivery scheme, calculating a dynamic regeneration score based on postoperative monitoring data, and regenerating the targeted exovesicle delivery scheme when the score is lower than a preset threshold. The application can improve the accuracy of pulp regeneration regulation.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and regenerative medicine, and relates to, but is not limited to, an intelligent method for regulating dental pulp regeneration based on targeted external vesicles. Background Technology

[0002] Pulp disease is a common oral disease. Traditional root canal treatment typically involves removing infected tissue and filling the root canal with an inert material. While this treatment can control infection, the replacement of the pulp with the inert material causes the tooth to lose its biological activity, often resulting in long-term problems such as increased tooth fragility and a high risk of reinfection. Pulp regeneration therapy aims to restore the structure and function of the pulp-dentin complex and is an ideal alternative to traditional root canal treatment.

[0003] In current technologies, dental pulp regeneration research mainly revolves around stem cell transplantation, growth factor delivery, and scaffold materials. Extracellular vesicles such as exosomes, as natural nanocarriers, are considered highly promising cell-free regeneration carriers due to their advantages of low immunogenicity, high biocompatibility, and ease of engineering modification. Studies have shown that exosomes derived from dental pulp stem cells can regulate inflammation, promote angiogenesis, and induce dentin formation. However, existing dental pulp regeneration treatments generally suffer from the following drawbacks: First, in complex and narrow root canal systems, treatment carriers are difficult to deliver effectively to the target site, especially in the apical region. Second, existing treatment protocols are mostly based on group experience, and the protocols are largely fixed, failing to dynamically adjust according to the patient's specific anatomical morphology, degree of damage, and changes in the microenvironment during the regeneration process. Furthermore, different regions of the pulp cavity have different biological functions and regeneration needs, but existing methods often treat them as a homogeneous whole, leading to incomplete functional reconstruction. Finally, in the diagnosis and assessment of pulp status, multimodal imaging technologies such as cone-beam computed tomography, magnetic resonance imaging, and optical coherence tomography are usually used for comprehensive evaluation, as well as gingival crevicular fluid and serum biomarker detection. Artificial intelligence technologies such as deep learning are used for medical image segmentation and prediction. However, existing technologies cannot deeply integrate multidimensional information with intelligent treatment decisions, execution, and feedback to form a closed-loop precision control system.

[0004] Therefore, there is an urgent need for a more comprehensive method for regulating pulp regeneration to address the problems of insufficient targeting, insufficient individualization, and neglect of spatial heterogeneity in existing methods, thereby significantly improving the accuracy and intelligence of pulp regeneration regulation. Summary of the Invention

[0005] This application provides an intelligent method for regulating dental pulp regeneration based on targeted external vesicles.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide an intelligent pulp regeneration regulation method based on targeted external vesicles. The method includes: collecting preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data from the patient; preprocessing the multimodal medical images to obtain a dual-channel fused image, wherein the dual-channel fused image is constructed by stitching together CBCT and MRI images along the channel dimension; generating a priori map of the root canal centerline based on the CBCT images, and combining it with the dual-channel fused image to form network input data; and inputting the network input data into a 3D... The pulp chamber is segmented using a U-Net++ network, outputting a binary mask and a pulp injury probability map. Based on the pulp chamber binary mask, the pulp injury probability map, and the root canal centerline, a functional partition map is constructed. This functional partition map is input into a type-aware graph neural network, where an attention mechanism based on node type consistency is introduced for information aggregation and feature learning to obtain the final feature representation of each node. Based on the final feature representation of each node, a partitioned multi-task regression head is used for prediction to obtain a pulp regeneration factor demand vector. Based on the regeneration factor demand vector, a dual-head hybrid action proximal strategy is used to optimize the proxy to generate corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, determining the targeted exovesicle delivery scheme. The patient is treated based on the targeted exovesicle delivery scheme, and postoperative monitoring data is collected on days 7 and 14 after treatment. A dynamic regeneration score is calculated based on the postoperative monitoring data. When the dynamic regeneration score is lower than a preset threshold, a new targeted exovesicle delivery scheme is generated.

[0007] The technical solution provided in this application collects preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data from patients. The multimodal medical images are preprocessed to obtain a dual-channel fused image. This dual-channel fused image is constructed by stitching together CBCT and MRI images along the channel dimension, providing a comprehensive and accurate data foundation for subsequent intelligent analysis and solving the problems of single information and insufficient support in traditional methods. A priori map of the root canal centerline is generated based on the CBCT images and combined with the dual-channel fused image to form network input data. The network input data is then input into a 3D... The segmentation is performed in the U-Net++ network, outputting a binary mask of the pulp cavity and a probability map of pulp injury, thereby achieving highly connected and unbroken precise segmentation of millimeter-narrow root canals, providing a reliable spatial structure model for subsequent precise drug delivery and zonal demand assessment. Based on the binary mask of the pulp cavity, the probability map of pulp injury, and the root canal centerline, a functional zoning map is constructed. The functional zoning map is input into a type-aware graph neural network, and an attention mechanism based on node type consistency is introduced for information aggregation and feature learning to obtain the final feature representation of the nodes. Based on the final feature representation of the nodes, a zonal multi-task regression head is used for prediction to obtain the pulp regeneration factor demand vector, thereby achieving precise quantification and spatial decoupling assessment of the regeneration demand of different functional areas in the apex, middle, and coronal regions of the pulp cavity, overcoming the limitation of treating the pulp as a homogeneous whole in traditional methods; based on the regeneration factor demand... The vector, through a dual-headed hybrid proximal strategy optimization agent, generates corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, determining the targeted exovesicle delivery plan. The abstract regeneration factor demand vector is automatically transformed into an executable targeted exovesicle delivery plan, enabling precision drug manufacturing. This solves the problem of fixed treatment plans and inability to personalize treatment in traditional methods. Patients are treated based on the targeted exovesicle delivery plan, and postoperative monitoring data is collected on days 7 and 14 after treatment. A dynamic regeneration score is calculated based on the postoperative monitoring data. When the dynamic regeneration score falls below a preset threshold, a new targeted exovesicle delivery plan is generated. This forms an intelligent adaptive closed-loop control of pulp regeneration, encompassing assessment, decision-making, treatment, reassessment, and feedback re-intervention, significantly improving treatment success rate and controllability.

[0008] Optionally, the step of generating a priori root canal centerline image based on the CBCT image and combining it with the dual-channel fused image to form network input data includes: performing coarse segmentation on the CBCT image to obtain an initial pulp mask, and extracting the root canal centerline from the initial pulp mask; calculating the Euclidean distance from each voxel to the root canal centerline in the three-dimensional voxel coordinate space defined by the CBCT image, generating a distance transformation map, normalizing the distance transformation map to obtain the priori root canal centerline image; and using the priori root canal centerline image as a third channel to stitch together with the dual-channel fused image to form three-channel network input data.

[0009] Optionally, the 3D U-Net++ network includes a shared encoder-decoder backbone and two independent output heads. The step of inputting the network input data into the 3D U-Net++ network for segmentation and outputting a pulp cavity binary mask and a pulp injury probability map includes: inputting the network input data into the 3D U-Net++ network for forward propagation; outputting a pulp cavity probability map through the first output head to characterize the confidence that each voxel belongs to the pulp tissue; thresholding the pulp cavity probability map to generate a pulp cavity binary mask; and outputting a pulp injury probability map through the second output head to characterize the probability that each voxel has damage.

[0010] Optionally, the step of constructing a functional zoning map based on the binary mask of the pulp cavity, the pulp injury probability map, and the root canal centerline includes: dividing the pulp cavity space into three functional zones along the root canal centerline: the apical zone, the middle zone, and the coronal zone; defining each voxel in the binary mask of the pulp cavity as a graph node, constructing a node set, and assigning a type label to the functional zone to which each node belongs based on its spatial location; establishing connection edges between each node and adjacent voxel nodes based on three-dimensional neighborhood rules, constructing an edge set; and constructing an initial feature vector for each node, wherein the initial feature vector includes the pulp injury probability value, local curvature, normalized distance to the apex, and type embedding feature corresponding to the node, and the type embedding feature is determined based on the type label.

[0011] Optionally, the step of inputting the functional partition map into a type-aware graph neural network and introducing an attention mechanism based on node type consistency for information aggregation and feature learning to obtain the final feature representation of the nodes includes: performing information aggregation and feature learning on the functional partition map through a multi-layer type-aware graph neural network, wherein, in the neighbor information aggregation process of each layer, the attention weight between a node and its neighbor nodes is determined by an attention mechanism based on node type consistency according to the consistency between the type label of the node and the type labels of its neighbor nodes. When calculating the attention weight between a node and its neighbor nodes, the attention mechanism introduces an indicator function as one of the input features. The indicator function outputs a first predetermined value when the type label of the node is the same as the type label of its neighbor node, and outputs a second predetermined value when they are different, so that the attention weight between nodes belonging to the same functional partition is higher than the attention weight between nodes not belonging to the same functional partition; the features of neighbor nodes are weighted and aggregated according to the attention weight, and combined with the node's own features, the feature representation of the node is updated through linear transformation and nonlinear activation function, and the final feature representation of each node is obtained after iterative update.

[0012] Optionally, the step of predicting the pulp regeneration factor demand vector using a partitioned multi-task regression head based on the final feature representation of the node includes: configuring independent regression prediction modules for the apical, mid-region, and coronal regions respectively; inputting the final feature representation into the corresponding regression prediction module according to the type label of each node, wherein the regression prediction module corresponding to the apical region outputs the demand prediction value of anti-inflammatory related factors, the regression prediction module corresponding to the mid-region outputs the demand prediction value of angiogenesis related factors, and the regression prediction module corresponding to the coronal region outputs the demand prediction value of dentin-related factors; and generating the pulp regeneration factor demand vector through an aggregation operation based on the demand prediction values ​​of all nodes in each functional region.

[0013] Optionally, the dual-head hybrid action proximal strategy optimization agent includes a shared encoder, a continuous action head, and a discrete action head. The shared encoder is connected to the continuous action head and the discrete action head, respectively. The step of generating corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences based on the regeneration factor demand vector through the dual-head hybrid action proximal strategy optimization agent includes: encoding the regeneration factor demand vector into an implicit semantic vector through the shared encoder; inputting the implicit semantic vector into the continuous action head and the discrete action head, and outputting a continuous action vector, which corresponds to the targeted exovesicle formulation parameters, including particle size, zeta potential, drug loading, ligand density, and concentration parameters; and outputting a discrete action sequence in an autoregressive manner through the discrete action head, which corresponds to the targeted ligand amino acid sequence.

[0014] Optionally, the dual-head hybrid action proximal strategy optimization agent is trained through the following process: constructing a composite reward function including a dose-response function and a simulation environment based on an intracanal fluid dynamics and pharmacodynamics model; interacting with the simulation environment through the dual-head hybrid action proximal strategy optimization agent according to the composite reward function and the simulation environment, and collecting a state-action-reward trajectory dataset; constructing a training sample set, which includes postoperative monitoring features and actual long-term regeneration scores from historical patient data; training a value prediction network through the training sample set, outputting a predicted long-term regeneration score; calculating the loss value of the predicted long-term regeneration score and the actual long-term regeneration score, and using the loss... The parameters of the value prediction network are optimized until the loss value converges to obtain a trained value prediction network. Based on the states in the state-action-reward trajectory dataset collected in the early stage of training, the corresponding predicted long-term regeneration score is determined as the agent reward through the trained value prediction network. Combined with the composite reward function, the policy network parameters of the dual-headed hybrid action proximal policy optimization agent are updated through the proximal policy optimization algorithm. After obtaining the true long-term regeneration score, the true reward is calculated based on the true long-term regeneration score through the composite reward function, and the policy network parameters are updated through the true reward until the policy converges to obtain the trained dual-headed hybrid action proximal policy optimization agent.

[0015] Secondly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the above-described intelligent pulp regeneration regulation method based on targeted external vesicles.

[0016] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the aforementioned intelligent pulp regeneration regulation method based on targeted external vesicles.

[0017] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides an intelligent pulp regeneration regulation method based on targeted external vesicles. It collects preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data from patients. The multimodal medical images are preprocessed to obtain a dual-channel fused image. This dual-channel fused image is constructed by stitching CBCT and MRI images along the channel dimension, providing a comprehensive and accurate data foundation for subsequent intelligent analysis and solving the problems of single information and insufficient support in traditional methods. A priori map of the root canal centerline is generated based on the CBCT images and combined with the dual-channel fused image to form network input data. The network input data is then input into a 3D... The segmentation is performed in the U-Net++ network, outputting a binary mask of the pulp cavity and a probability map of pulp injury, thereby achieving highly connected and unbroken precise segmentation of millimeter-narrow root canals, providing a reliable spatial structure model for subsequent precise drug delivery and zonal demand assessment. Based on the binary mask of the pulp cavity, the probability map of pulp injury, and the root canal centerline, a functional zoning map is constructed. The functional zoning map is input into a type-aware graph neural network, and an attention mechanism based on node type consistency is introduced for information aggregation and feature learning to obtain the final feature representation of the nodes. Based on the final feature representation of the nodes, a zonal multi-task regression head is used for prediction to obtain the pulp regeneration factor demand vector, thereby achieving precise quantification and spatial decoupling assessment of the regeneration demand of different functional areas in the apex, middle, and coronal regions of the pulp cavity, overcoming the limitation of treating the pulp as a homogeneous whole in traditional methods; based on the regeneration factor demand... The vector, through a dual-headed hybrid proximal strategy optimization agent, generates corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, determining the targeted exovesicle delivery plan. The abstract regeneration factor demand vector is automatically transformed into an executable targeted exovesicle delivery plan, enabling precision drug manufacturing. This solves the problem of fixed treatment plans and inability to personalize treatment in traditional methods. Patients are treated based on the targeted exovesicle delivery plan, and postoperative monitoring data is collected on days 7 and 14 after treatment. A dynamic regeneration score is calculated based on the postoperative monitoring data. When the dynamic regeneration score falls below a preset threshold, a new targeted exovesicle delivery plan is generated. This forms an intelligent adaptive closed-loop control of pulp regeneration, encompassing assessment, decision-making, treatment, reassessment, and feedback re-intervention, significantly improving treatment success rate and controllability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1A schematic flowchart illustrating an intelligent pulp regeneration control method based on targeted external vesicles provided in this application embodiment; Figure 2 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0022] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0023] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0024] In view of the current problems in the research on pulp regeneration regulation in the interdisciplinary field of artificial intelligence and regenerative medicine, this application provides an intelligent pulp regeneration regulation method based on targeted external vesicles.

[0025] The technical solution of this application is described below, starting with the method embodiments.

[0026] Please refer to Figure 1It shows a flowchart of an intelligent pulp regeneration regulation method based on targeted external vesicles provided in an embodiment of this application, such as... Figure 1 As shown, the method includes at least the following steps S110 to S150.

[0027] Step S110: Collect the patient's preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data. Preprocess the multimodal medical images to obtain dual-channel fused images, wherein the dual-channel fused images are constructed by stitching together CBCT images and MRI images along the channel dimension.

[0028] In this embodiment, multi-source heterogeneous data of the patient is collected. Specifically, before surgery, multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data are collected simultaneously. Multimodal medical images include CBCT images, MRI images, and OCT images. Molecular biomarker data includes inflammatory factors, regeneration factors (IL-6, TNF-α, VEGF, BMP-2, etc.), and exosome-derived miRNAs detected through gingival crevicular fluid and serum samples. Clinical data includes age, medical history, smoking status, and tooth position. Furthermore, to form a closed loop, OCT images, gingival crevicular fluid, and serum data are collected again on days 7 and 14 after TEV treatment as postoperative monitoring data. Further, the multimodal medical images are preprocessed. CBCT images are resampled and normalized, and MRI images are registered to the CBCT image space and normalized. The preprocessed CBCT and MRI images are stitched together along the channel dimension to construct a dual-channel fused image, where the first channel is the CBCT image channel and the second channel is the MRI image channel.

[0029] Step S120: Generate a priori map of the root canal centerline based on the CBCT image, and combine it with the dual-channel fused image to form network input data; input the network input data into the 3D U-Net++ network for segmentation, and output the pulp chamber binary mask and pulp damage probability map.

[0030] In this embodiment, a root canal centerline prior map is generated based on CBCT images and combined with dual-channel fused images to form the input data for a 3D U-Net++ network. Specifically, the network input data includes dual-channel fused images and a root canal centerline prior map. The root canal centerline prior map is obtained through the following process: From the first channel of the dual-channel fused images, i.e., the CBCT image channel, the CBCT image is coarsely segmented using a fast thresholding method to obtain an initial pulp mask, and the root canal centerline is extracted from the initial pulp mask using a vascular modeling toolkit such as VMTK; In the three-dimensional voxel coordinate space defined by the CBCT image, the Euclidean distance from each voxel to all points on the root canal centerline is calculated, and the minimum value is taken to obtain a distance transformation map. The distance transformation map is then normalized to the [0,1] interval using Min-Max to obtain the root canal centerline prior map. This root canal centerline prior map is used as the third channel and stitched with the dual-channel fused images to form three-channel network input data, providing high-quality data input for the subsequent 3D U-Net++ network.

[0031] In this embodiment, network input data is input into a 3D U-Net++ network for segmentation, outputting a binary mask of the pulp cavity and a pulp damage probability map, thereby achieving accurate segmentation of the pulp cavity and damaged areas. Specifically, the 3D U-Net++ network adopts a 4-layer nested dense U-shaped structure, including a shared encoder-decoder backbone and two independent output heads. Network input data is input into the 3D U-Net++ network for forward propagation. The first output head outputs the pulp cavity probability map, which characterizes the confidence level of each voxel belonging to healthy or necrotic pulp tissue. The second output head outputs the pulp damage probability map, which characterizes the probability of each voxel exhibiting damage such as necrosis or inflammation. Furthermore, since standard loss cannot guarantee a single connected channel in the segmentation result, often resulting in breakpoints and fragmentation that do not conform to the physiological structure of the pulp, the loss function is improved to ensure the path connectivity and lack of breaks in the segmentation result, and to ensure that the subsequent TEV can be delivered to the pulp apex. A composite loss function containing a topology-aware regularization term is constructed. .in, The Dice and cross-entropy combined loss is used for pulp chamber segmentation. For Focal loss in the damaged area, The loss is topology-aware, specifically a curvature regularization term based on the pulp cavity probability map. To encourage the formation of smoothly connected tubular boundaries, The topology loss weights are set to a specific range of 0.1 to 0.3, as setting them to the conventional parameters of 0.5-1.0 can lead to apical fracture. Therefore, for the small-scale structure of the pulp cavity, the topology loss weights are set to a specific range of 0.1 to 0.3 to ensure high connectivity in the segmentation results within narrow root canals. The 3D U-Net++ network is trained using the AdamW optimizer with an early stopping strategy. Furthermore, the pulp cavity probability map is binarized by setting a threshold, generating the final pulp cavity binary mask. The pulp injury probability map remains as continuous probability values. Based on the pulp cavity binary mask and the pulp injury probability map, a direct data foundation is provided for subsequent regeneration demand prediction.

[0032] Step S130: Based on the binary mask of the pulp cavity, the pulp injury probability map, and the root canal centerline, a functional partition map is constructed; the functional partition map is input into a type-aware graph neural network, and an attention mechanism based on node type consistency is introduced to perform information aggregation and feature learning to obtain the final feature representation of the node; based on the final feature representation of the node, a partitioned multi-task regression head is used to predict and obtain the pulp regeneration factor demand vector.

[0033] In this embodiment, a functional zoning map is constructed based on a binary mask of the pulp cavity, a pulp injury probability map, and the root canal centerline. Specifically, the pulp cavity space is divided into three functional zones along the root canal centerline: the apical zone, the central zone, and the coronal zone. Each voxel with a value of 1 in the binary mask of the pulp cavity is defined as a graph node. Based on all the graph nodes, a node set in the functional zoning map can be constructed. According to the spatial position of each node, a type label corresponding to the functional zone is assigned to each node. Based on the three-dimensional 26-neighborhood rule, a connection edge is established between each node and its adjacent voxel nodes. Based on all the connection edges, an edge set in the functional zoning map can be constructed. Finally, an initial feature vector is constructed for each node. This initial feature vector includes the pulp injury probability value corresponding to the node determined according to the pulp injury probability map, the local curvature representing the local geometry, the normalized distance to the apex, and the type embedding feature. The type embedding feature is determined based on the type label. The pulp injury probability value corresponding to the node is determined according to the pulp injury probability map, the local curvature is mainly used to represent the local geometry of the node, and the type embedding feature is determined based on the type label.

[0034] In this embodiment, after the functional partitioning graph is constructed, it is input into a type-aware graph neural network for information aggregation and feature learning to obtain the final feature representation of the nodes. Specifically, the type-aware graph neural network adopts a multi-layer GraphSAGE architecture. This network performs information aggregation and feature learning on the functional partitioning graph. The neighbor information aggregation process of each layer's nodes is based on a node type consistency-based attention mechanism. This node type consistency-based attention mechanism ensures that when calculating the attention weight between a node and its neighboring nodes, it not only relies on the features of the current node and its neighboring nodes but also directly introduces an indicator function as one of the input features. The value of this indicator function is determined based on the consistency between the type label of the current node and the type labels of its neighboring nodes. When the type label of the current node is the same as the type label of its neighboring nodes, i.e., the current node and its neighboring nodes belong to the same functional area, a first predetermined value is output as the indicator function. The value of the number is used to determine the output value of the indicator function. When the type label of the current node is different from that of its neighboring nodes, i.e., the current node and its neighboring nodes belong to different functional areas, a second predetermined value is output as the value of the indicator function. At this time, the output value of the indicator function is 0, which makes the attention weight between nodes belonging to the same functional area higher than the attention weight between nodes not belonging to the same functional area. This makes the information of nodes in the same functional area given priority during aggregation, realizing information focus within the functional area. The features of neighboring nodes are weighted and aggregated according to the attention weight, and combined with the node's own features, the feature representation of the node is updated through linear transformation and nonlinear activation function. After multiple iterations, the final feature representation of each node is obtained.

[0035] In this embodiment, based on the final feature representation of each node, a partitioned multi-task regression head is used for prediction to obtain a pulp regeneration factor demand vector, thereby predicting the regeneration demand for spatial heterogeneity. Specifically, independent regression prediction modules, such as MLP, are configured for the apical, mid-region, and coronal regions. Based on the type label of each node, the corresponding final feature representation is input into the corresponding regression prediction module. Specifically, the regression prediction module for the apical region outputs predicted demand values ​​for anti-inflammatory factors, including IL-6 inhibition demand and TNF-α inhibition demand; the regression prediction module for the mid-region outputs predicted demand values ​​for pro-angiogenesis factors, including VEGF promotion demand and SDF-1 promotion demand; and the regression prediction module for the coronal region outputs predicted demand values ​​for odontoblast-related factors, including BMP-2 expression potential and DSPP expression potential. Finally, based on the predicted demand values ​​of all nodes within each functional zone, a pulp regeneration factor demand vector is generated through aggregation operations. For example, the mean of the predicted demand values ​​of all nodes within each functional zone is calculated, and then the pulp regeneration factor demand vector is determined based on the mean of the predicted demand values ​​corresponding to each functional zone. This pulp regeneration factor demand vector is a 6-dimensional vector, which accurately quantifies the differentiated regulatory needs of different functional zones, providing a quantitative basis for the subsequent generation of targeted external vesicle solutions.

[0036] Step S140: Based on the regeneration factor demand vector, the agent is optimized by using a dual-headed hybrid action proximal strategy to generate the corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, thereby determining the targeted exovesicle delivery scheme.

[0037] In this embodiment, a targeted exovesicle delivery scheme is determined using a dual-headed hybrid action proximal strategy optimization agent based on the regeneration factor demand vector. Specifically, the regeneration factor demand vector is standardized to form an input state representation. The trained dual-headed hybrid action proximal strategy optimization agent generates corresponding targeted exovesicle formulation parameters and target ligand amino acid sequences, thereby determining the targeted exovesicle delivery scheme. The core architecture of this dual-headed hybrid action proximal strategy optimization agent consists of a shared encoder, a continuous action head, and a discrete action head. The shared encoder is connected to both the continuous and discrete action heads. The shared encoder (e.g., a 2-layer MLP) encodes the input state representation into a high-dimensional implicit semantic vector, which is then input to... In the continuous action head and discrete action head, after receiving the implicit semantic vector, the continuous action head outputs a continuous action vector, which corresponds to the targeted exovesicle formulation parameters, including key physicochemical parameters such as particle size, zeta potential, drug loading ratio of various regeneration factors, ligand density, total exovesicle concentration, and hyaluronic acid concentration. Among these, the sum of the drug loading ratios of various regeneration factors is 1. After receiving the implicit semantic vector, the discrete action head outputs a discrete action sequence consisting of 7 amino acids in an autoregressive manner. This discrete action sequence corresponds to the target ligand amino acid sequence, which is sampled from a vocabulary of 20 standard amino acids.

[0038] In this embodiment, the training of the dual-head hybrid action proximal policy optimization agent relies on a composite reward function that integrates multi-objectives and prior knowledge. This composite reward function is expressed by the following formula: ; In the formula, This represents a factor-matching reward system. By substituting the formulation parameters in the action into a dose-response function pre-calibrated through in vitro cell experiments, the predicted biological effect is calculated and compared with the target demand, thereby encouraging the generation of formulations that match the demand. The weighting coefficients represent the factor matching rewards; This represents a sequence stability reward, which assesses the natural stability of the generated ligand sequence based on a protein language model. The weighting coefficients representing the sequence stability reward; This indicates a safety penalty, imposed for formulation parameters exceeding safe limits or for ligand sequences containing known immunogenic motifs. The weight coefficients represent the safety penalty. A model-based meta-reinforcement learning strategy is used to train a dual-headed hybrid action proximal policy optimization agent to address the feedback delay issue in real clinical rewards based on postoperative 14-day efficacy. Specifically, a simulation environment based on intracanal fluid dynamics and pharmacodynamics models is constructed. The dual-headed hybrid action proximal policy optimization agent interacts with the simulation environment according to the composite reward function and the simulation environment, collecting a state-action-reward trajectory dataset. A training sample set is constructed, including postoperative monitoring features and actual long-term regeneration scores from historical patient data at 7 and 14 days. A lightweight value prediction network is trained using this training sample set, outputting a predicted long-term regeneration score. The loss values ​​for the predicted and actual long-term regeneration scores are calculated, and the parameters of the value prediction network are optimized using these loss values ​​until convergence, resulting in the trained value prediction network. Based on the states in the state-action-reward trajectory dataset collected during the initial training of the dual-headed hybrid action proximal policy optimization agent, the corresponding predicted long-term regeneration score is determined through the trained value prediction network. This predicted long-term regeneration score is used as the agent reward to accelerate policy convergence. Combined with a composite reward function, the policy network parameters of the dual-headed hybrid action proximal policy optimization agent are updated through the proximal policy optimization algorithm. After obtaining the true long-term regeneration score, the true reward is calculated based on the true long-term regeneration score using the composite reward function. Then, the policy network parameters are fine-tuned using this true reward root until the policy converges, resulting in the trained dual-headed hybrid action proximal policy optimization agent.

[0039] Furthermore, based on the trained dual-head hybrid action proximal strategy optimization agent, after performing a single forward propagation for a new patient condition, personalized targeted exovesicle formulation parameters and target ligand amino acid sequences can be simultaneously output to determine the targeted exovesicle delivery protocol. This protocol can directly guide the laboratory in the preparation of engineered exovesicles, for example, using HEK293 cell lines expressing specific target peptides and loading drugs via electroporation according to the formulation parameters, providing clear auxiliary basis for clinicians to perform precise intracanal injection.

[0040] Step S150: The patient is treated based on the targeted external vesicle delivery protocol, and postoperative monitoring data is collected on the 7th and 14th days after treatment. A dynamic regeneration score is calculated based on the postoperative monitoring data. When the dynamic regeneration score is lower than a preset threshold, the targeted external vesicle delivery protocol is regenerated.

[0041] In this embodiment, after treatment with a targeted external vesicle delivery protocol, the targeted external vesicles are mixed with hyaluronic acid hydrogel and injected into the pulp cavity via a microinjector for dynamic efficacy evaluation and control. Specifically, postoperative monitoring data is collected on days 7 and 14 after treatment, including obtaining postoperative OCT images using an OCT probe, collecting gingival crevicular fluid and serum samples to detect the concentration of specific biomarkers, and calculating a multi-dimensional dynamic regeneration score based on this postoperative monitoring data. This dynamic regeneration score consists of an anti-inflammatory effect score, a vascularization progression score, and a dentin formation potential score. The anti-inflammatory effect score is determined by comparing the concentration of inflammatory factors (such as IL-6 and TNF-α) in the gingival crevicular fluid on day 7 postoperatively with the average baseline concentration before treatment. If inflammation worsens, the score is 0. The vascularization progression score is determined by weighted summation of OCT blood flow signals at 7 and 14 days post-surgery. For example, the weight of the OCT blood flow signal at 7 days post-surgery is 0.3, and the weight of the OCT blood flow signal at 14 days post-surgery is 0.7, thereby assessing the persistence of neovascularization. The odontogenic potential score is determined based on the concentration of odontogenic markers (such as BMP-2 and DSPP) in the gingival crevicular fluid at 14 days post-surgery and an individualized ideal threshold. The individualized ideal threshold is dynamically mapped based on the required predicted values ​​of BMP-2 expression potential and DSPP expression potential, thereby achieving a precise match between the assessment criteria and the individual's expected regeneration potential. Finally, the anti-inflammatory effect score, vascularization progression score, and odontogenic potential score are added together to obtain the dynamic regeneration score.

[0042] Furthermore, a tiered decision-making process is performed based on the dynamic regeneration score. The dynamic regeneration score is compared with a preset threshold, which is determined by ROC curve analysis of historical experimental data. The upper limit of the preset threshold is 0.85, and the lower limit is 0.65. When the dynamic regeneration score is greater than or equal to 0.85, pulp regeneration is considered successful, an evaluation report is generated, and the intervention ends. When the dynamic regeneration score is between 0.65 and 0.84, three sub-scores are analyzed: anti-inflammatory effect score, vascularization progression score, and odontogenic potential score. For the deficient dimension (such as insufficient anti-inflammatory effect), a micro-adjustment instruction for local supplemental drug administration is triggered. When the dynamic regeneration score is less than 0.65, the current targeted external vesicle delivery protocol is considered ineffective. At this time, the protocol redesign process is automatically triggered. Based on the current efficacy feedback data and the current patient status, the dual-head hybrid proximal strategy optimization agent is re-driven to generate an optimized second-generation targeted external vesicle delivery protocol, thereby realizing intelligent adaptive closed-loop control of pulp regeneration through evaluation, decision-making, feedback, and re-intervention.

[0043] In summary, the intelligent pulp regeneration regulation method based on targeted external vesicles provided in this application collects preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data from patients. The multimodal medical images are preprocessed to obtain a dual-channel fused image. This dual-channel fused image is constructed by stitching CBCT and MRI images along the channel dimension, providing a comprehensive and accurate data foundation for subsequent intelligent analysis and solving the problems of single information and insufficient support in traditional methods. A priori map of the root canal centerline is generated based on the CBCT images and combined with the dual-channel fused image to form network input data. The network input data is then input into a 3D... The segmentation is performed in the U-Net++ network, outputting a binary mask of the pulp cavity and a probability map of pulp injury, thereby achieving highly connected and unbroken precise segmentation of millimeter-narrow root canals, providing a reliable spatial structure model for subsequent precise drug delivery and zonal demand assessment. Based on the binary mask of the pulp cavity, the probability map of pulp injury, and the root canal centerline, a functional zoning map is constructed. The functional zoning map is input into a type-aware graph neural network, and an attention mechanism based on node type consistency is introduced for information aggregation and feature learning to obtain the final feature representation of the nodes. Based on the final feature representation of the nodes, a zonal multi-task regression head is used for prediction to obtain the pulp regeneration factor demand vector, thereby achieving precise quantification and spatial decoupling assessment of the regeneration demand of different functional areas in the apex, middle, and coronal regions of the pulp cavity, overcoming the limitation of treating the pulp as a homogeneous whole in traditional methods; based on the regeneration factor demand... The vector, through a dual-headed hybrid proximal strategy optimization agent, generates corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, determining the targeted exovesicle delivery plan. The abstract regeneration factor demand vector is automatically transformed into an executable targeted exovesicle delivery plan, enabling precision drug manufacturing. This solves the problem of fixed treatment plans and inability to personalize treatment in traditional methods. Patients are treated based on the targeted exovesicle delivery plan, and postoperative monitoring data is collected on days 7 and 14 after treatment. A dynamic regeneration score is calculated based on the postoperative monitoring data. When the dynamic regeneration score falls below a preset threshold, a new targeted exovesicle delivery plan is generated. This forms an intelligent adaptive closed-loop control of pulp regeneration, encompassing assessment, decision-making, treatment, reassessment, and feedback re-intervention, significantly improving treatment success rate and controllability.

[0044] It should be noted that, in the embodiments of this application, if the above-mentioned intelligent pulp regeneration regulation method based on targeted external vesicles is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a magnetic disk, or an optical disk. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0045] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps in any of the above embodiments of an intelligent pulp regeneration control method based on targeted external vesicles. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in any of the above embodiments of an intelligent pulp regeneration control method based on targeted external vesicles.

[0046] Based on the same technical concept, this application provides an electronic device for implementing the intelligent pulp regeneration control method based on targeted external vesicles described in the above method embodiments. Figure 2 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application, such as... Figure 2 As shown, the electronic device 200 includes a memory 210 and a processor 220. The memory 210 stores a computer program that can run on the processor 220. When the processor 220 executes the program, it implements the steps in any of the intelligent pulp regeneration control methods based on targeted external vesicles described in the embodiments of this application.

[0047] The memory 210 is configured to store instructions and applications executable by the processor 220, and can also cache data to be processed or already processed by the processor 220 and various modules in the electronic device (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0048] When the processor 220 executes the program, it implements the steps of an intelligent pulp regeneration control method based on targeted external vesicles, as described above. The processor 220 typically controls the overall operation of the electronic device 200.

[0049] The aforementioned processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that other electronic devices can also implement the functions of the aforementioned processor, and this application does not specifically limit the specific implementation.

[0050] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0051] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0052] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0053] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0054] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0055] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0056] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0057] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0058] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0059] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0060] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent regulation of dental pulp regeneration based on targeted external vesicles, characterized in that, The method includes: The patient's preoperative multimodal medical images, molecular biomarker data, clinical data, and postoperative monitoring data are collected. The multimodal medical images are preprocessed to obtain dual-channel fused images, which are constructed by stitching CBCT images and MRI images along the channel dimension. A priori map of the root canal centerline is generated based on the CBCT images, and network input data is formed by combining the dual-channel fused images. The network input data is then input into a 3D U-Net++ network for segmentation, and a binary mask of the pulp chamber and a probability map of pulp damage are output. Based on the binary mask of the pulp cavity, the probability map of pulp injury, and the root canal centerline, a functional zoning map is constructed. The functional zoning map is input into a type-aware graph neural network, and an attention mechanism based on node type consistency is introduced for information aggregation and feature learning to obtain the final feature representation of the nodes. Based on the final feature representation of the nodes, a zoning multi-task regression head is used for prediction to obtain the pulp regeneration factor demand vector. Based on the regeneration factor demand vector, the corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences are generated by optimizing the agent through a dual-head hybrid action proximal strategy, thereby determining the targeted exovesicle delivery scheme. The patient was treated using the targeted extravesicular vesicle delivery protocol, and postoperative monitoring data were collected on the 7th and 14th days after treatment. A dynamic regeneration score was calculated based on the postoperative monitoring data. When the dynamic regeneration score was lower than a preset threshold, the targeted extravesicular vesicle delivery protocol was regenerated.

2. The method according to claim 1, characterized in that, The step of generating a priori root canal centerline map based on the CBCT images and combining it with the dual-channel fused images to form network input data includes: The CBCT image is coarsely segmented to obtain an initial pulp mask, and the root canal centerline is extracted from the initial pulp mask; In the three-dimensional voxel coordinate space defined by the CBCT image, the Euclidean distance from each voxel to the root canal centerline is calculated, a distance transformation map is generated, and the distance transformation map is normalized to obtain the root canal centerline prior map. The prior image of the root canal centerline is used as the third channel and stitched together with the dual-channel fused image to form three-channel network input data.

3. The method according to claim 2, characterized in that, The 3D U-Net++ network includes a shared encoder-decoder backbone and two independent output heads. The input data to the 3D U-Net++ network is segmented to output a binary mask of the pulp cavity and a pulp injury probability map, including: The network input data is input into a 3D U-Net++ network for forward propagation. A pulp cavity probability map, representing the confidence level of each voxel belonging to the dental pulp tissue, is output through the first output head. The pulp cavity probability map is thresholded to generate a pulp cavity binary mask. A pulp injury probability map, representing the probability of damage to each voxel, is output through the second output head.

4. The method according to claim 3, characterized in that, The construction of a functional zoning map based on the binary mask of the pulp cavity, the pulp injury probability map, and the root canal centerline includes: The pulp chamber space is divided into three functional zones along the root canal centerline: the apical zone, the middle zone, and the coronal zone. Each voxel in the binary mask of the pulp cavity is defined as a graph node, a node set is constructed, and a type label of the functional partition to which each node belongs is assigned according to the spatial position of each node; Based on the three-dimensional neighborhood rule, a connection edge is established between each node and its neighboring voxel nodes to construct an edge set; An initial feature vector is constructed for each node. The initial feature vector includes the pulp injury probability value, local curvature, normalized distance to the root apex, and type embedding feature corresponding to the node. The type embedding feature is determined based on the type label.

5. The method according to claim 1, characterized in that, The step of inputting the functional partitioning map into a type-aware graph neural network, introducing an attention mechanism based on node type consistency for information aggregation and feature learning, and obtaining the final feature representation of the nodes includes: The functional partitioning graph is processed by a multi-layer type-aware graph neural network for information aggregation and feature learning. In the process of neighbor information aggregation at each layer, the attention weight between a node and its neighbors is determined by an attention mechanism based on the consistency between the node's type label and the type labels of its neighbors. When calculating the attention weight between a node and its neighbors, the attention mechanism introduces an indicator function as one of the input features. The indicator function outputs a first predetermined value when the type label of a node is the same as the type label of its neighbors, and outputs a second predetermined value when they are different, so that the attention weight between nodes belonging to the same functional partition is higher than the attention weight between nodes that do not belong to the same functional partition. The features of neighboring nodes are weighted and aggregated according to attention weights, and combined with the node's own features, the feature representation of the node is updated through linear transformation and nonlinear activation function. After iterative update, the final feature representation of each node is obtained.

6. The method according to claim 5, characterized in that, The step of predicting the pulp regeneration factor demand vector based on the final feature representation of the node using a partitioned multi-task regression head includes: Independent regression prediction modules are configured for the root tip zone, middle zone, and crown zone, respectively; Based on the type label of each node, the final feature representation is input into the corresponding regression prediction module. Specifically, the regression prediction module corresponding to the apical region outputs the required predicted values ​​of anti-inflammatory factors, the regression prediction module corresponding to the middle region outputs the required predicted values ​​of angiogenesis-related factors, and the regression prediction module corresponding to the coronal region outputs the required predicted values ​​of odontoblast-related factors. Based on the demand forecast values ​​of all nodes within each functional zone, a pulp regeneration factor demand vector is generated through aggregation operations.

7. The method according to claim 5, characterized in that, The dual-head hybrid action proximal strategy optimization agent includes a shared encoder, a continuous action head, and a discrete action head. The shared encoder is connected to the continuous action head and the discrete action head, respectively. Based on the regeneration factor demand vector, the dual-head hybrid action proximal strategy optimization agent generates corresponding targeted exovesicle formulation parameters and targeted ligand amino acid sequences, including: The regeneration factor demand vector is encoded into an implicit semantic vector using the shared encoder. The implicit semantic vector is input into the continuous action head and the discrete action head, and a continuous action vector is output. The continuous action vector corresponds to the targeted exovesicle formulation parameters, which include particle size, zeta potential, drug loading, ligand density, and concentration parameters. The discrete action head outputs a discrete action sequence in an autoregressive manner, and the discrete action sequence corresponds to the amino acid sequence of the targeted ligand.

8. The method according to claim 7, characterized in that, The dual-head hybrid action proximal policy optimization agent is trained through the following process: A composite reward function including a dose-response function and a simulation environment based on an intracanal fluid dynamics and pharmacodynamics model are constructed. Based on the composite reward function and the simulation environment, the dual-head hybrid action proximal strategy optimization agent interacts with the simulation environment to collect a state-action-reward trajectory dataset. A training sample set is constructed, which includes postoperative monitoring features and actual long-term regeneration scores from historical patient data. A value prediction network is trained using the training sample set, and a predicted long-term regeneration score is output. The loss value of the predicted long-term regeneration score and the actual long-term regeneration score is calculated, and the parameters of the value prediction network are optimized using the loss value until the loss value converges to obtain the trained value prediction network. Based on the states in the state-action-reward trajectory dataset collected in the early stage of training, the corresponding predicted long-term regeneration score is determined as the agent reward through the trained value prediction network, and the policy network parameters of the dual-head hybrid action proximal policy optimization agent are updated through the proximal policy optimization algorithm in combination with the composite reward function. Once the true long-term regeneration score is obtained, the true reward is calculated based on the true long-term regeneration score using the composite reward function, and the policy network parameters are updated using the true reward until the policy converges, thus obtaining the trained dual-head hybrid action proximal policy optimization agent.

9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 8.

Images