Dynamic follow-up visit system after maxillofacial deformity orthognathic operation based on intelligent nursing platform

Abstract

The invention discloses a maxillofacial deformity orthognathic post-operation dynamic follow-up visit system based on an intelligent nursing platform, and the system comprises an image input and preprocessing module which is used for collecting images and generating a continuous time sequence image sequence; the geometric flow modeling module is used for constructing a non-parametric partial differential geometric flow network model; the feature coding module is used for extracting features of the structure parameters; the parameter optimization module is used for executing a Kelvin-snake group resonance optimization algorithm; the reasoning and scoring module is used for executing reasoning, outputting geometric flow evolution representation and generating recovery state scores and abnormal prompt information; and the intelligent nursing platform integration module is used for deploying a final non-parametric partial differential geometric flow network model, receiving a patient image, and completing recovery monitoring, score output, abnormity marking and follow-up information pushing. According to the invention, automation of postoperative recovery modeling is realized, evaluation precision and follow-up visit efficiency are improved, and precise management of intelligent nursing is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of medical artificial intelligence, and particularly to a dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform. Background Art

[0002] Currently, in the rehabilitation management after orthognathic surgery for maxillofacial deformities, the traditional follow-up mode mainly relies on means such as manual regular follow-up visits, subjective judgment of doctors, and image comparison and analysis to monitor and guide the postoperative recovery of patients. This method has obvious limitations such as long cycle, dependence on experience, and lagged intervention. Especially in the early postoperative recovery stage, the facial structure changes dynamically and frequently, and it is difficult for traditional means to achieve continuous modeling of the recovery process and regional differential management. Some hospitals and rehabilitation centers have tried to introduce digital platforms to assist in follow-up, such as electronic questionnaire systems and image upload interfaces, but they can only achieve data collection and archiving, lacking intelligent evaluation and recovery trend analysis functions, and it is difficult to meet the refined and personalized rehabilitation needs.

[0003] In terms of modeling methods, existing research mostly uses static deep neural networks or time series prediction models to roughly classify the recovery status through discriminant classifiers. These methods usually ignore the structural evolution process of facial regions in the time series dimension and fail to establish the coupling and correlation relationship between regions, resulting in the lack of continuity and spatial pertinence of the model when processing facial data at different times. In addition, the dynamic changes in the facial recovery process have obvious regional heterogeneity. For example, the postoperative recovery speeds and evolution directions of the upper and lower jaws, cheekbones, and facial soft tissues are significantly different, and existing technologies mostly use a unified parameter modeling method, making it difficult to capture the time response differences between regions.

[0004] In terms of structural parameter optimization, traditional models often use grid search or gradient-based optimization algorithms. These methods are prone to falling into local optima in high-dimensional spaces. Especially when dealing with non-linear evolution features and complex facial curvature changes, their optimization effects are unstable, the convergence speed is slow, and the model generalization ability is poor. In addition, existing optimization strategies usually do not have an adaptive perturbation mechanism and cannot dynamically adjust the optimization step size according to the current state of the model, resulting in the convergence result of structural parameters being highly sensitive to the initial value, increasing the model training difficulty and dependence on manual intervention.

[0005] To sum up the above problems, the existing technologies generally have the following deficiencies in postoperative recovery monitoring: First, there is a lack of continuous modeling ability for the dynamic process after maxillofacial surgery; second, it is impossible to establish a differential recovery parameter system according to different facial regions; third, the optimization effect is poor and the robustness is lacking in the high-dimensional parameter space. Therefore, there is an urgent need for an intelligent follow-up system that can fuse time series images and regional features, construct an interpretable geometric evolution model, and have an adaptive optimization ability to improve the accuracy and efficiency of postoperative recovery assessment.

[0006] Therefore, how to provide a dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on non-parametric partial differential geometric flow modeling and Kelvin-snake swarm resonance optimization algorithm. The present invention fully combines continuous time-series image modeling, facial region structure parameterization and bio-inspired optimization methods, and details the whole process of continuously predicting the postoperative facial recovery state and personalized scoring, and has the advantages of visualizing the recovery trend, automatically identifying abnormal regions and high evaluation accuracy.

[0008] According to an embodiment of the present invention, a dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform includes: An image input and preprocessing module, configured to collect preoperative images and postoperative images, perform region calibration and standardization processing, and generate a continuous time-series image sequence; A geometric flow modeling module, configured to construct a non-parametric partial differential geometric flow network model, receive the continuous time-series image sequence and extract region structure parameters; A feature encoding module, configured to extract features of the region structure parameters, and generate a region weight parameter, an evolution direction parameter and a time scale regulation factor; A parameter optimization module, configured to execute the Kelvin-snake swarm resonance optimization algorithm to optimize the region structure parameters; An inference and scoring module, configured to perform inference using the optimized non-parametric partial differential geometric flow network model, output a geometric flow evolution representation, and generate a recovery state score and an abnormal prompt message; An intelligent nursing platform integration module, configured to deploy the final non-parametric partial differential geometric flow network model, receive postoperative images of patients, and complete recovery monitoring, score output, abnormal marking and follow-up information push.

[0009] Optionally, the modules are implemented by the following method: S1. Obtain preoperative images and images at several postoperative time periods, perform region calibration and standardization processing, and generate a continuous time-series image sequence; S2. Based on the continuous time-series image sequence, construct a non-parametric partial differential geometric flow network model, and the non-parametric partial differential geometric flow network model includes a region weight parameter, an evolution direction parameter and a time scale regulation factor; S3. Input the continuous time-series image sequence into the non-parametric partial differential geometric flow network model, encode the maxillofacial region through a local geometric kernel function, and generate a geometric flow evolution representation sequence including a region weight parameter, an evolution direction parameter and a time scale regulation factor; S4. Construct a Kelvin-snake swarm resonance optimization algorithm and initialize the population parameter set containing the multi-frequency perturbation mechanism and the curvature adjustment factor; S5. Use the Kelvin-snake swarm resonance optimization algorithm to jump out of the local extreme point under the action of the multi-frequency perturbation mechanism, adaptively adjust the step size under the control of the curvature adjustment factor, and iteratively optimize the geometric flow evolution representation sequence to generate a non-parametric partial differential geometric flow network model with optimized structure; S6. Deploy the non-parametric partial differential geometric flow network model with optimized structure to the intelligent nursing platform, receive the postoperative images of the patient, perform geometric flow inference, and output the recovery status score and abnormal prompt information; S7. Generate a follow-up feedback information package based on the recovery status score and abnormal prompt information, and complete archiving and pushing.

[0010] Optionally, the specific content of S2 includes: S21. Construct a non-parametric partial differential geometric flow network model based on the continuous time-series image sequence. The continuous time-series image sequence is a set of preoperative images and two-dimensional facial images at several postoperative time periods arranged in chronological order, denoted as , where represents the two-dimensional facial image submanifold at the th time step, and is the total number of time steps; S22. Divide facial regions on each two-dimensional facial image submanifold , and define the regional weight parameter set , where represents the weight coefficient of the th region at time step ; S23. Define the evolution direction parameter set , where represents the two-dimensional evolution direction vector of the th region at time step , which is used to represent the spatial orientation characteristics of the morphological changes of the facial region; S24. Define the time scale regulation factor set , where represents the time response factor of the th facial region during the recovery process, which is used to control the difference in the geometric evolution speed between different regions; S25. Construct a non-parametric partial differential evolution based on the regional weight parameter , the evolution direction parameter and the time scale regulation factor , and set , As the structural parameters that can be trained for the non-parametric partial differential geometric flow network model, the initial structure of the model is set.

[0011] Optionally, the S3 specifically includes: S31. Perform regional division on each two-dimensional facial image sub-manifold to obtain a facial region set , where represents the time step at the th facial sub-region; S32. Construct a local geometric kernel function on each facial sub-region . The local geometric kernel function is an embedded computing unit that extracts geometric structure features such as curvature change, boundary trend, and texture density within the region, and is used to map the two-dimensional facial image sub-manifold into a structural feature vector. The generated structural feature vector is denoted as ; S33. Initialize the region weight parameter , the evolution direction parameter , and the time scale regulation factor in the non-parametric partial differential geometric flow network model according to the structural feature vector , where: The region weight parameter is estimated by the mean value of the geometric gradient modulus in the structural feature vector . The gradient modulus reflects the local curvature and deformation intensity of the region; The evolution direction parameter is calculated from the principal direction feature component in the structural feature vector , and the direction is taken as the vector direction corresponding to the principal component; The time scale regulation factor is calculated from the deformation degree of the th region of the face in the full-time sequence image, and the standard deviation of the structural feature changes in the region at different time steps is used as the estimated value of the response factor; S34. Based on , and , construct a geometric flow evolution representation , where represents the geometric evolution intensity of the th region at the time step , which is used to characterize the local morphological changes of the face during the recovery process. The geometric flow evolution representation set is used as the structural output of the non-parametric partial differential geometric flow network model.

[0012] Optionally, the S4 specifically includes: S41. Construct a Kelvin - snake swarm resonance optimization algorithm to optimize the regional weight parameters, evolution direction parameters and time - scale regulation factors in the non - parametric partial differential geometric flow network model, and achieve structural perturbation and convergent search by simulating the collaborative action behavior of snakes in a multi - frequency resonance field; S42. Set the optimized population parameter set , where is the population size, and each individual contains a structural parameter ; S43. Assign an independent multi - frequency perturbation vector to each individual , where each component represents the resonance frequency corresponding to the th structural parameter dimension, and the perturbation frequency is obtained by randomly sampling within a preset frequency interval; S44. Introduce a curvature adjustment factor set in each structural parameter dimension. The curvature adjustment factor is calculated based on the gradient change rate in the geometric flow evolution representation of the previous stage, and dynamically adjusts the perturbation amplitude of each parameter dimension; S45. Construct a multi - frequency resonance perturbation expression. Let the perturbation vector of the th individual in the th structural parameter dimension be ; S46. Bind the structural parameter of each initial individual with the perturbation vector to form a perturbed structural parameter , which is used as the initial population input of the Kelvin - snake swarm resonance optimization algorithm.

[0013] Optionally, the specific steps of S5 are as follows: S51. Based on the initial population input, in each iteration, periodically perturb the structural parameters according to the multi - frequency perturbation vector corresponding to the initial individual, so that the structural parameters jump out of the local extreme region. At the same time, according to the curvature adjustment factor of the structural parameter dimension, adaptively adjust the perturbation amplitude, take a preset small step size in the changing region and a preset large step size in the smooth region, and dynamically control the update amplitude of the structural parameters to generate new candidate individuals of the structural parameters; S52. Input the new candidate individuals of the structural parameters into the non - parametric partial differential geometric flow network model, execute the inference process, and obtain the geometric flow evolution results of each region of the face at each time step; ​S53. Compare the geometric flow evolution results output by the model with the true recovery state, calculate the fitting effect, and assign a fitness score. S54. Rank all individuals in the population, select the individual with the optimal fitness for retention in the next round, and at the same time replace the individuals with low fitness based on the multi-frequency perturbation mechanism to generate new parameter combinations and expand the search space. S55. Repeat the process of parameter update, model inference, and population iteration until the preset convergence condition is met or the maximum number of rounds is reached. Finally, extract the regional weight parameters, evolution direction parameters, and time scale adjustment factors of the individual with the optimal fitness to form a non-parametric partial differential geometric flow network model with optimized structure.

[0014] Optionally, the recovery state score in step S6 is generated based on the evolution amplitude and trend stability of each region, and the anomaly prompt information is used to identify the region location and time period with local recovery anomalies.

[0015] Optionally, the follow-up feedback information package in step S7 includes a recovery trend graph, rehabilitation advice text, and interaction instructions, and archiving and pushing are completed based on the follow-up feedback information package.

[0016] The beneficial effects of the present invention are as follows: First, by constructing a non-parametric partial differential geometric flow network model, the present invention can simulate the continuous geometric evolution process of the facial region after maxillofacial surgery in the time dimension, realize the dynamic modeling of the postoperative recovery situation, overcome the defect that traditional static neural networks cannot capture time series changes, and effectively improve the continuity of the recovery process modeling and the regional expression accuracy.

[0017] Second, the present invention introduces three types of structural parameters, namely regional weight parameters, evolution direction parameters, and time scale adjustment factors, independently models the differential recovery characteristics of each anatomical region of the face, realizes multi-dimensional control of the recovery speed, dominant deformation direction, and influence strength, enhances the model's ability to depict facial heterogeneity and regional differences, and meets the refined evaluation requirements.

[0018] In addition, the present invention uses the Kelvin-snake swarm resonance optimization algorithm to optimize the model structure parameters. Combining the multi-frequency perturbation mechanism and the curvature adjustment factor, it can effectively jump out of the local extreme point and dynamically adjust the optimization step size, improving the search efficiency and convergence stability in the high-dimensional structure space, and solving the problem that traditional optimization methods are prone to falling into local optima. Description of the Drawings

[0019] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings: Figure 1Schematic diagram of the module composition of the dynamic follow-up system for maxillofacial deformity orthognathic surgery based on the intelligent nursing platform proposed by the present invention; Figure 2 Method flowchart of the dynamic follow-up system for maxillofacial deformity orthognathic surgery based on the intelligent nursing platform proposed by the present invention; Figure 3 Initialization structure diagram of the Kelvin-snake swarm resonance optimization algorithm for the dynamic follow-up system for maxillofacial deformity orthognathic surgery based on the intelligent nursing platform proposed by the present invention. Specific implementation manners

[0020] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0021] Refer to Figures 1 - 3 , the dynamic follow-up system for maxillofacial deformity orthognathic surgery based on the intelligent nursing platform includes: An image input and preprocessing module, which is used to collect preoperative images and postoperative images, perform regional calibration and standardization processing, and generate a continuous time-series image sequence; A geometric flow modeling module, which is used to construct a non-parametric partial differential geometric flow network model, receive the continuous time-series image sequence and extract regional structure parameters; A feature encoding module, which is used to extract the features of the regional structure parameters, and generate regional weight parameters, evolution direction parameters and time scale control factors; A parameter optimization module, which is used to execute the Kelvin-snake swarm resonance optimization algorithm to optimize the regional structure parameters; An inference and scoring module, which is used to perform inference using the optimized non-parametric partial differential geometric flow network model, output the geometric flow evolution representation, and generate a recovery status score and an abnormal prompt message; An intelligent nursing platform integration module, which is used to deploy the final non-parametric partial differential geometric flow network model, receive the patient's postoperative images, and complete recovery monitoring, score output, abnormal marking and follow-up information push.

[0022] The present invention clearly divides functional units such as image input, modeling, feature extraction, parameter optimization, inference scoring and platform integration. The logic between modules is clear and the data is in a closed loop, which is convenient for the actual system deployment and function expansion. Compared with the existing closed algorithm flow scheme, it has more advantages in terms of being structured, integratable and engineering implementable.

[0023] In this embodiment, the modules are implemented through the following methods: S1. Obtain preoperative images and images at several postoperative time periods, perform regional calibration and standardization processing to generate a continuous time-series image sequence, specifically including: collecting preoperative images of the patient and images at multiple key time nodes after surgery to ensure coverage of the structural evolution process throughout the recovery period. After image acquisition, the system automatically calibrates the facial region, including the positioning of key anatomical regions such as the zygomatic bone, mandibular angle, maxilla, and chin. Subsequently, the image data will undergo processing such as size normalization, brightness standardization, and pose correction to eliminate interference caused by shooting conditions and individual differences, ensuring the comparability of images at different time points. After processing, the image sequence is constructed into a continuous time-series image sequence in a unified format according to the time order; S2. Based on the continuous time-series image sequence, construct a non-parametric partial differential geometric flow network model, which includes regional weight parameters, evolution direction parameters, and time-scale control factors; S3. Input the continuous time-series image sequence into the non-parametric partial differential geometric flow network model, encode the maxillofacial region through a local geometric kernel function, and generate a geometric flow evolution representation sequence containing regional weight parameters, evolution direction parameters, and time-scale control factors; S4. Construct a Kelvin-snake swarm resonance optimization algorithm and initialize a population parameter set containing a multi-frequency perturbation mechanism and a curvature adjustment factor; S5. Use the Kelvin-snake swarm resonance optimization algorithm to jump out of local extreme points under the action of the multi-frequency perturbation mechanism, adaptively adjust the step size under the control of the curvature adjustment factor, and iteratively optimize the geometric flow evolution representation sequence to generate a non-parametric partial differential geometric flow network model with optimized structure; S6. Deploy the non-parametric partial differential geometric flow network model with optimized structure to the intelligent nursing platform, receive postoperative images of the patient, perform geometric flow inference, and output a recovery status score and abnormal prompt information; S7. Generate a follow-up feedback information package based on the recovery status score and abnormal prompt information, and complete archiving and pushing.

[0024] The present invention constructs an intelligent follow-up system based on non-parametric partial differential geometric flow modeling and Kelvin-snake swarm resonance optimization algorithm, breaking the limitations of traditional postoperative rehabilitation methods that rely on static image comparison and manual experience evaluation, realizing continuous modeling, regional difference analysis, and parameter adaptive optimization of the postoperative recovery process, effectively improving the accuracy and efficiency of recovery prediction, and having the technical advantages of high modeling flexibility, strong regional resolution, and intelligent system response.

[0025] In this embodiment, the specific content of S2 includes: S21. Construct a non-parametric partial differential geometric flow network model based on a continuous time-series image sequence, where the continuous time-series image sequence is a set of preoperative images and two-dimensional facial images at several postoperative time periods arranged in chronological order, denoted as , where represents the two-dimensional facial image submanifold at the -th time step, is the total number of time steps; S22. Divide facial regions on each two-dimensional facial image submanifold , and define a set of regional weight parameters , where represents the weight coefficient of the -th region at time step ; S23. Define a set of evolution direction parameters , where represents the two-dimensional evolution direction vector of the -th region at time step , which is used to represent the spatial guiding features of the morphological changes of facial regions; S24. Define a set of time scale regulation factors , where represents the time response factor of the -th facial region during the recovery process, which is used to control the difference in the geometric evolution speed between different regions; S25. Construct a non-parametric partial differential evolution based on the regional weight parameter , the evolution direction parameter , and the time scale regulation factor , and take , as the trainable structure parameters of the non-parametric partial differential geometric flow network model, and complete the setting of the initial structure of the model. The expression is as follows: ; where represents the local geometric morphological change rate of the -th facial region at time step , is the divergence value of the evolution direction parameter , which is used to control the trend of regional expansion or contraction.

[0026] In the present invention, the time-scale regulation factor regional weight parameter and the divergence term of the evolution direction are coupled into the driving core of the facial region shape change rate through this formula, so that the dynamic change of each region during the recovery process not only has spatial direction guidance, but also has time regulation and importance weight adjustment capabilities. Compared with traditional models that only consider image pixel changes or unified evolution methods, this structure has the advantages of strong interpretability, fine control granularity, and obvious regional response differentiation, and can more realistically reflect the expansion or contraction trend during the postoperative recovery process of facial regions, thereby significantly improving the robustness and accuracy of the model in time series prediction and anomaly recognition. Especially in scenarios with strong regional heterogeneity and asynchronous recovery rhythms, this formula provides an accurate, controllable, and dynamic evolution modeling method, with strong clinical adaptability and intelligent analysis value.

[0027] The present invention innovatively introduces regional weight parameters, evolution direction parameters, and time-scale regulation factors to construct a multi-dimensional structure expression for describing the evolution process of maxillofacial recovery. Different from the unified weight modeling strategy of existing models, it can achieve regional differential modeling and the controllability of the recovery rhythm, thereby enhancing the expression ability and prediction ability of the model for complex facial deformation structures.

[0028] In this embodiment, S3 specifically includes: S31. Perform regional division on each two-dimensional facial image submanifold Specifically, first extract the standard feature points in the image through a face key point detection algorithm, such as the eyebrow bone, the corner of the eye, the nose wing, the corner of the lip, the zygomatic arch, the mandibular margin, etc.; then construct a facial geometric topology map based on these key points, and divide the entire image into several facial regions with physiological significance, such as the zygomatic region, the perinasal region, the mandibular angle region, and the lower lip region, etc., to obtain the facial region set where represents the th facial sub-region at time step S32. In each facial sub-region Construct a local geometric kernel function thereon; constructing a local geometric kernel function means generating a kernel function mapping method for feature extraction within each divided facial sub-region by utilizing the local geometric structure information of the image; this process first extracts low-dimensional geometric descriptors representing the regional morphology based on geometric elements such as boundary curvature, gray-scale gradient, and texture direction within the region, and then maps these descriptors to a high-dimensional feature space using the kernel method. Common kernels include Gaussian kernel, Laplace kernel, or a custom kernel based on the structure tensor, with the aim of magnifying regional detail differences; the constructed local geometric kernel function has the ability of non-linear mapping and can capture the minute deformations and structural changes of the facial region in different temporal states; finally, each kernel function forms a set of structural feature vectors within the corresponding region, serving as the basic input for subsequent structure parameter generation, thus realizing the structured conversion from the image to the parameter space; the local geometric kernel function is an embedded computing unit for extracting geometric structure features such as curvature changes, boundary trends, and texture densities within the region, used to map the two-dimensional facial image sub-manifold into structural feature vectors, and the generated structural feature vectors are denoted as ; S33. According to the structural feature vectors Initialize the regional weight parameter , the evolution direction parameter and the time-scale regulation factor , where: The regional weight parameter is estimated by the mean value of the geometric gradient modulus in the structural feature vectors . Specifically, it includes: first, extracting the gradient vector of each pixel within the region, for example, by calculating through the Sobel operator or the structure tensor, then taking the modulus of each gradient vector to obtain the gradient intensity distribution, and finally calculating the mean value of the gradient modulus of all pixels in this region. The gradient modulus reflects the local curvature and deformation intensity of this region; The evolution direction parameter is calculated from the main direction feature component in the structural feature vectors . Specifically, it includes: first, constructing a structure tensor matrix within the region, fusing the horizontal and vertical gradient information, then performing eigenvalue decomposition on this tensor, and extracting the eigenvector corresponding to the maximum eigenvalue as the main direction vector component of this region, and the direction is taken as the vector direction corresponding to the principal component; The time-scale regulation factor is from the Calculating the deformation degree of a region in the full-time series image, specifically including: First, registering and tracking the same sub-region in the continuous time series image, and extracting its key geometric feature points; then calculating the spatial displacement or contour change amplitude of these feature points at different time points; finally, normalizing the change values in the entire time series process to obtain the cumulative deformation degree index of the region, and using the standard deviation of the structural feature changes of the region in different time steps as the response factor estimation value; S34. Based on , and construct a geometric flow evolution representation , where represents the geometric evolution intensity of the -th region at the time step , which is used to characterize the local morphological changes of the face during the recovery process. The set of geometric flow evolution representations is used as the structural output of the non-parametric partial differential geometric flow network model.

[0029] The present invention constructs a regional feature vector through a local geometric kernel function, and generates three types of structural parameters, namely regional weight, evolution direction, and time regulation, based on this vector to form an initial structural representation for evolutionary modeling, solving the problems of insufficient expressiveness of traditional image features in the time dimension and inability to support time series recovery reasoning, and significantly improving the regional-level dynamic prediction effect.

[0030] In this embodiment, the specific content of S4 includes: S41. Construct a Kelvin-snake swarm resonance optimization algorithm to optimize the regional weight parameter , evolution direction parameter and time scale regulation factor in the non-parametric partial differential geometric flow network model, and realize structural perturbation and convergence search by simulating the cooperative behavior of snakes in a multi-frequency resonance field; S42. Set the optimized population parameter set , where is the population size, and each individual contains a structural parameter ; S43. Assign an independent multi-frequency perturbation vector to each individual , where each component represents the resonance frequency corresponding to the -th structural parameter dimension, and the perturbation frequency is obtained by randomly sampling within a preset frequency range; S44. Introduce a curvature adjustment factor set in each structural parameter dimension. The curvature adjustment factor Dynamically adjust the perturbation amplitude of each parameter dimension according to the calculation of the gradient change rate in the geometric flow evolution representation in the previous stage; S45. Construct a multi-frequency resonance perturbation expression. Let the perturbation vector of the th individual in the th structural parameter dimension be : ; Among them, is the base amplitude, is pi, is the iteration step, is the phase offset, is the curvature adjustment factor, represents the perturbation step size of the corresponding dimension; In the present invention, by constructing a multi-frequency resonance perturbation expression and combining the base amplitude, multi-frequency perturbation frequency, phase offset, and curvature adjustment factor, dynamic perturbation control of model parameters is achieved during the structure optimization process. This formula not only allows each parameter dimension to perform periodic perturbations in an independent frequency space but also introduces a curvature adjustment mechanism, enabling the perturbation amplitude to be flexibly adjusted according to the gradient change of the current model state. Compared with traditional fixed-step or single-frequency perturbation optimization algorithms, this mechanism significantly improves the flexibility of parameter space exploration and the ability to jump out of local optima, effectively enhancing the optimization convergence speed and structural adaptability, and providing a refined adjustment basis for the efficient optimization of high-dimensional geometric parameters.

[0031] S46. Bind the structural parameters of each initial individual with the perturbation vector to form the perturbed structural parameter , which is used as the initial population input of the Kelvin-snake swarm resonance optimization algorithm.

[0032] In the present invention, by introducing a multi-frequency perturbation mechanism and a curvature adjustment factor, intelligent perturbation control during the search process of model structural parameters is achieved. This method can effectively avoid the problems of traditional optimization algorithms falling into local optima or having rigid parameter updates, enhance the jumpiness and convergence stability of high-dimensional parameter space search, and has obvious performance advantages over genetic algorithms or particle swarms in structure optimization tasks.

[0033] In this embodiment, the specific steps of S5 include: S51. Based on the input of the initial population, in each iteration, the structural parameters are periodically perturbed according to the multi-frequency perturbation vectors corresponding to the initial individuals, so that the structural parameters jump out of the local extreme region. At the same time, according to the curvature adjustment factor of the dimension of the structural parameters, the perturbation amplitude is adaptively adjusted. A preset small step size is adopted in the changing region and a preset large step size is adopted in the smooth region to dynamically control the update amplitude of the structural parameters, and new candidate individuals of the structural parameters are generated; S52. Input the new candidate individuals of the structural parameters into the non-parametric partial differential geometric flow network model, perform the inference process, and obtain the geometric flow evolution results of each region of the face at each time step; S53. Compare the geometric flow evolution results output by the model with the true recovery state, calculate the fitting effect, and assign a fitness score. The specific process is as follows: First, extract the structural parameters of the key regions in the reconstructed image or feature curve output by the model; then perform spatial alignment and feature correspondence with the true temporal image at the same time point, and calculate the differences between the two in terms of contour position, gradient direction or geometric deformation amplitude; finally, quantify these differences into fitting error indicators to reflect the model's ability to restore the postoperative facial changes at the current stage, and provide a basis for further scoring and optimization; S54. Score and sort all individuals in the population, select the individual with the optimal fitness for retention in the next round, and at the same time replace the individuals with low fitness based on the multi-frequency perturbation mechanism to generate new parameter combinations to expand the search space; S55. Repeat the parameter update, model inference and population iteration process until the preset convergence condition is met or the maximum number of rounds is reached. Finally, extract the regional weight parameters, evolution direction parameters and time scale regulation factors of the individual with the optimal fitness to form a non-parametric partial differential geometric flow network model with optimized structure.

[0034] The present invention constructs a parameter update process based on perturbation control and differential feedback evaluation. By dynamically adjusting the parameters under the action of the multi-frequency perturbation and adaptive adjustment mechanism, it realizes the fine-tuning of the model structure and the improvement of the fitting ability required for high-precision recovery state scoring, and solves the problem of insufficient model accuracy caused by fixed parameters or single training path in the prior art.

[0035] In this embodiment, the recovery status score of step S6 is generated based on the evolution amplitude and trend stability of each region, and the abnormal prompt information is used to identify the regional position and time period of local recovery abnormalities, specifically including: firstly, analyzing the deformation trajectory of each facial sub-region in the full time series image, extracting its evolution amplitude change value and the change trend in the time series, the system calculates the morphological change size of each region at different time points, and evaluates the continuity and stability of these changes. When an abnormal mutation occurs in a certain area, the recovery speed is abnormally slow or fluctuates violently, the system marks the area and the corresponding time period as an abnormal point, and generates abnormal prompt information to prompt clinical attention.

[0036] In this embodiment, the follow-up feedback information package of step S7 includes a recovery trend chart, rehabilitation advice text and interactive instructions, and archiving and pushing are completed based on the follow-up feedback information package, specifically including: after the follow-up is completed, the collected facial recovery images, scoring results and abnormal prompt information are uniformly processed. The system generates a visual recovery trend chart based on the recovery scores and trend changes of each area; combines the postoperative recovery rules and model evaluation results to automatically generate personalized rehabilitation advice text; and extracts interactive instructions based on user interactive behaviors to guide the next nursing operation. Finally, this information is packaged into a follow-up feedback information package, automatically archived in the user's nursing file, and pushed to medical staff and patients through the platform to achieve closed-loop follow-up management.

[0037] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention was applied to the remote rehabilitation management project after orthognathic surgery carried out by the maxillofacial surgery department of a tertiary comprehensive hospital. The project aims to improve the efficiency of postoperative recovery follow-up and the accuracy of recovery status assessment of patients, and alleviate the problems of tight manual follow-up resources, strong subjectivity in scoring, and delayed abnormal identification. In actual use, 60 postoperative patients who underwent correction of mandibular protrusion or maxillary underdevelopment were selected. All patients uploaded a total of about 3,400 facial image data from the first to the 90th day after surgery, covering daily changes in the key recovery stage after surgery.

[0038] Before the system was deployed, the traditional follow-up method used offline consultations and postoperative satisfaction questionnaires every two weeks combined with the doctor's visual judgment to evaluate recovery. The scoring was mainly based on facial swelling, mouth cleft and subjective bite feedback. On average, each patient's follow-up required 22 minutes of manual processing time, and the subjective scoring consistency rate was only 73%. After the system of the present invention is deployed, patients can automatically upload facial images through the mobile app. After the images are standardized by the image input and preprocessing modules, they are sent to the constructed non-parametric partial differential geometric flow network model. The system generates the recovery evolution intensity score of each region in real time based on the regional weight parameters, evolution direction parameters and time scale control factors output by the model.

[0039] In the initial stage of system deployment, expert-annotated samples are introduced to construct initial evolution reference values. The Kelvin-Snake swarm resonance optimization algorithm is used to automatically optimize the network structure parameters, and stable convergence is achieved through about 80 rounds of optimization iterations. The mean error of the recovery score of the trained model on real test data is ±3.7 points (out of 100 full marks), which is better than the ±9.1 points of traditional questionnaires and doctor judgments. Among the samples with large fluctuations in the recovery score within 14 days after surgery, the system automatically marked 37 abnormal areas, and 32 of them were confirmed by doctors during reexamination as real problems of unresolved early edema, local infection, or occlusal deviation, with an abnormal recognition accuracy rate of 86.5%. For the 8 cases where patients reported discomfort but there were no abnormalities in the traditional score, the system could also identify the deviation of the evolution trend in the zygomatic arch area, effectively assisting doctors in making early interventions.

[0040] The recovery trend curve graphs, score reports, and abnormal prompt information automatically generated by the platform are uniformly summarized and pushed to the mobile terminals of attending physicians and patients, and archived in the hospital rehabilitation system. The actual deployment data shows that the average scoring time of the system for each patient does not exceed 1.8 seconds, the auxiliary follow-up efficiency is increased by nearly 12 times, and the consistency rate of medical staff scoring is increased to 92.6%. In terms of the questionnaire recovery rate and patient satisfaction, after system tracking, the active cooperation rate of patients within 90 days has increased from the original 58% to 91%, and the average number of times patients actively upload images is 47 times per person, which is 1.9 times higher than the traditional mode.

[0041] In practical applications, the present invention has significantly improved the efficiency of postoperative follow-up, the accuracy of scoring, and the timeliness of abnormal prompts, verifying its practical feasibility and promotion value in postoperative dynamic rehabilitation management.

[0042] The following table shows the statistical data of the key data collected and evaluated during the implementation of the present invention: Table 1: Statistical Table of Dynamic Recovery Modeling and Scoring Performance Evaluation after Orthognathic Surgery

[0043] During the implementation of the orthognathic postoperative dynamic follow-up system, a "Statistical Table of Dynamic Recovery Modeling and Scoring Performance Evaluation after Orthognathic Surgery" was used to conduct a comparative analysis of the traditional follow-up method and the system proposed in the present invention. In terms of the average manual time required for each patient's follow-up, the traditional method is 22 minutes on average, while the present system only requires 1.8 seconds, indicating that the system has an order-of-magnitude improvement in follow-up efficiency and greatly reduces the workload of medical staff. In terms of the accuracy of recovery scoring, the traditional scoring error is about ±9.1 points, while the present system reduces the error to ±3.7 points through model evolution output, effectively improving the accuracy of quantitative evaluation.

[0044] In terms of anomaly recognition, the accuracy rate of this system is as high as 86.5%, significantly superior to the 38.2% of traditional manual recognition, which reflects the significant role of geometric flow modeling and structural parameter optimization. At the same time, the consistency rate of medical staff scoring has increased from the original 73.0% to 92.6%, indicating that the system output results have higher professionalism and consistency, and are more reliable in assisting doctors' decision-making. The active cooperation rate of patients has also increased from the original 58.0% to 91.0%, reflecting that the intelligent platform is more attractive and convenient to operate in terms of user experience and feedback process design.

[0045] In terms of the frequency of patients uploading images, the average number of uploads under the guidance of this system reaches 47 times per person, much higher than the 25 times per person in the traditional method, ensuring the quality of sequential input of the model and the continuity of recovery trend modeling. In addition, during the model training and deployment process, an average of 80 rounds of structural optimization iterations are completed, achieving fine-tuning of the model structure while ensuring the reasonable use of computing resources, and supporting the efficient operation of the final scoring and prompting module. The generation time of the system's automatic scoring report is controlled within 3 seconds, providing a strong guarantee for the real-time feedback of the remote care platform.

[0046] Overall, the data in this table clearly demonstrate the various technical advantages of this invention in the postoperative recovery assessment process, being comprehensively superior to traditional means in terms of scoring accuracy, processing speed, user participation, and anomaly detection ability, verifying the practical implementation value of this system in the dynamic follow-up scenario after maxillofacial deformity surgery.

[0047] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform, characterized in that, Including: An image input and preprocessing module, which is used to collect preoperative images and postoperative images, perform regional calibration and normalization processing, and generate a continuous time-series image sequence; A geometric flow modeling module, which is used to construct a non-parametric partial differential geometric flow network model, receive the continuous time-series image sequence, and extract regional structure parameters; A feature encoding module, which is used to extract the features of the regional structure parameters, and generate regional weight parameters, evolution direction parameters, and time-scale regulation factors; A parameter optimization module, which is used to execute the Kelvin-snake swarm resonance optimization algorithm to optimize the regional structure parameters; An inference and scoring module, which is used to perform inference using the optimized non-parametric partial differential geometric flow network model, output the geometric flow evolution representation, and generate a recovery status score and an anomaly prompt message; An intelligent nursing platform integration module, which is used to deploy the final non-parametric partial differential geometric flow network model, receive the patient's postoperative images, and complete recovery monitoring, score output, anomaly marking, and follow-up information push.

2. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on the intelligent nursing platform according to claim 1, wherein The modules are implemented through the following methods: S1. Obtain preoperative images and images at several postoperative time periods, perform regional calibration and normalization processing, and generate a continuous time-series image sequence; S2. Based on the continuous time-series image sequence, construct a non-parametric partial differential geometric flow network model, and the non-parametric partial differential geometric flow network model includes regional weight parameters, evolution direction parameters, and time-scale regulation factors; S3. Input the continuous time-series image sequence into the non-parametric partial differential geometric flow network model, encode the maxillofacial region through a local geometric kernel function, and generate a geometric flow evolution representation sequence containing regional weight parameters, evolution direction parameters, and time-scale regulation factors; S4. Construct a Kelvin-snake swarm resonance optimization algorithm, and initialize a population parameter set including a multi-frequency perturbation mechanism and a curvature adjustment factor; S5. Use the Kelvin-snake swarm resonance optimization algorithm to jump out of the local extreme point under the action of the multi-frequency perturbation mechanism, adaptively adjust the step size under the control of the curvature adjustment factor, iteratively optimize the geometric flow evolution representation sequence, and generate a non-parametric partial differential geometric flow network model with optimized structure; S6. Deploy the non-parametric partial differential geometric flow network model with optimized structure to the intelligent nursing platform, receive the patient's postoperative images, perform geometric flow inference, and output a recovery status score and an anomaly prompt message; S7. Generate a follow-up feedback information package based on the recovery status score and the anomaly prompt message, and complete archiving and pushing.

3. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on the intelligent nursing platform according to claim 2, characterized in that, The specific content of S2 includes: S21. Construct a non-parametric partial differential geometric flow network model based on a continuous time-series image sequence, where the continuous time-series image sequence is a set of preoperative images and two-dimensional facial images at several postoperative time periods arranged in chronological order, denoted as , where represents the two-dimensional facial image submanifold at the -th time step, and is the total number of time steps. S22. On each two-dimensional facial image submanifold partition facial regions, and define a set of region weight parameters , where represents the time step and is the weight coefficient of the S23. Define the set of evolution direction parameters , where represents the time step at time the two-dimensional evolution direction vector of the -th region, which is used to represent the spatial guiding features of the morphological changes of the facial region; S24. Define a set of time-scale regulation factors , where represents the time response factor of the th facial region during the recovery process, which is used to control the difference in the geometric evolution speed between different regions; S25. Based on the regional weight parameter , the evolution direction parameter and the time scale regulation factor are used to construct a non-parametric partial differential evolution, and , is used as the trainable structural parameter of the non-parametric partial differential geometric flow network model to complete the setting of the initial structure of the model.

4. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform according to claim 3, characterized in that The specific content of S3 includes: S31. Perform regional division on each two-dimensional facial image submanifold to obtain a facial region set , where represents the time step at the th facial sub-region; S32. On each facial sub-region construct a local geometric kernel function, which is an embedded computing unit for extracting geometric structure features such as curvature change, boundary trend, and texture density within a region, and is used to map a two-dimensional facial image sub-manifold into a structural feature vector. The generated structural feature vector is denoted as ; S33. Initialize the regional weight parameters, evolution direction parameters and time scale regulation factors in the non-parametric partial differential geometric flow network model according to the structural feature vector, where: Region weight parameter Estimated from the mean value of the geometric gradient magnitude in the structural feature vector wherein the gradient magnitude reflects the local curvature and deformation intensity of the region Evolution direction parameter Calculated from the main direction feature component in the structural feature vector The direction is taken as the vector direction corresponding to the principal component; Time scale regulator Calculated from the deformation degree of the region in the full-time series image, and the standard deviation of the structural feature changes of the region in different time steps is used as the estimated value of the response factor; S34. Based on , and construct a geometric flow evolution representation , where represents the geometric evolution intensity of the th region at time step is used to characterize the local morphological changes of the face during the recovery process, and the geometric flow evolution representation set is used as the structural output of the non-parametric partial differential geometric flow network model.

5. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on the intelligent nursing platform according to claim 4, wherein The specific content of S4 includes: S41. Construct a Kelvin-snake swarm resonance optimization algorithm to optimize the regional weight parameters, evolution direction parameters and time scale regulation factors in the non-parametric partial differential geometric flow network model, and achieve structural perturbation and convergence search by simulating the collaborative action behavior of snakes in a multi-frequency resonance field; S42. Set the optimized population parameter set , where is the population size, and each individual contains a structural parameter ; S43. For each individual allocate an independent multi-frequency perturbation vector , where each component represents the resonance frequency corresponding to the th structural parameter dimension, and the perturbation frequency is obtained by randomly sampling within a preset frequency range; S44. Introduce a set of curvature adjustment factors in each structural parameter dimension , the curvature adjustment factor is calculated based on the gradient change rate in the geometric flow evolution representation of the previous stage, and dynamically adjusts the perturbation amplitude of each parameter dimension; S45. Construct a multi-frequency resonance perturbation expression. Let the perturbation vector of the th individual in the th structural parameter dimension be ; S46. Bind the structural parameters of each initial individual with the perturbation vector to form the structural parameters with perturbation , which are used as the input of the initial population of the Kelvin-snake swarm resonance optimization algorithm.

6. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on the intelligent nursing platform according to claim 5, wherein The specific content of S5 includes: S51. Based on the input of the initial population, in each iteration, periodically perturb the structure parameters according to the multi-frequency perturbation vector corresponding to the initial individual, so that the structure parameters jump out of the local extreme region. At the same time, according to the curvature adjustment factor of the dimension of the structure parameters, adaptively adjust the perturbation amplitude, adopt a preset small step size in the changing region, and adopt a preset large step size in the smooth region, and dynamically control the update amplitude of the structure parameters to generate a new candidate individual of the structure parameters; S52. Input the new candidate individual of the structure parameters into the non-parametric partial differential geometric flow network model, perform the inference process, and obtain the geometric flow evolution results of each region of the face at each time step. S53. Compare the geometric flow evolution result output by the model with the true recovery state, calculate the fitting effect, and assign a fitness score; S54. Rank all individuals in the population, select the individual with the optimal fitness for retention in the next round, and at the same time replace the individuals with low fitness based on the multi-frequency perturbation mechanism to generate new parameter combinations to expand the search space; S55. Repeat the parameter update, model inference, and population iteration processes until the preset convergence condition is met or the maximum number of rounds is reached. Finally, extract the regional weight parameters, evolution direction parameters, and time scale regulation factors of the individual with the optimal fitness to form a non-parametric partial differential geometric flow network model with optimized structure.

7. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on the intelligent nursing platform according to claim 6, wherein The recovery state score in step S6 is generated based on the evolution amplitude and trend stability of each region, and the anomaly prompt information is used to identify the regional location and time period of local recovery anomalies.

8. The dynamic follow-up system for orthognathic surgery of maxillofacial deformities based on an intelligent nursing platform according to claim 7, wherein The follow-up feedback information package in step S7 includes a recovery trend graph, rehabilitation advice text, and interaction instructions, and archiving and pushing are completed based on the follow-up feedback information package.

Images