Intelligent Planning Methods and Systems for Early Childhood Correction Programs

Abstract

This invention relates to the field of medical artificial intelligence and orthodontic auxiliary decision-making technology, and discloses an intelligent planning method and system for early childhood orthodontic treatment. The method collects cephalometric indicators and individual information of children to construct a growth status feature vector, predicts the growth trajectory of the jawbone through a multi-layer fully connected network, and constructs a multi-constraint bi-objective optimization problem with appliance type, wearing duration, force value and follow-up period as decision variables. An improved multi-objective genetic algorithm is used to search for a Pareto optimal solution set, and closed-loop monitoring of deviations is performed during the orthodontic process for adaptive adjustment, realizing intelligent planning of the entire process of prediction-optimization-monitoring-adjustment.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and orthodontic auxiliary decision-making technology, and in particular to an intelligent planning method and system for early childhood orthodontic treatment. Background Technology

[0002] Functional malocclusions in children during the mixed dentition period include types such as anterior crossbite, malocclusion, and deep overbite, which have a high incidence rate in clinical practice. Studies have shown that if timely intervention is not carried out during the active stage of jawbone development, these malocclusions often develop into skeletal deformities during the permanent dentition period and gradually solidify, eventually requiring orthognathic surgery for correction. This not only increases the treatment burden and physical trauma for children but also significantly increases medical costs. Therefore, the orthodontic community both domestically and internationally generally advocates for early orthodontic treatment in children during the mixed dentition period to fully utilize the jawbone growth potential at this stage to guide craniofacial development and eliminate malocclusions in their early stages.

[0003] However, designing early orthodontic treatment plans is a complex, multivariate decision-making task. When developing a plan, dentists need to consider multiple decision variables simultaneously, including appliance type selection, daily wearing time setting, force application control, force application frequency adjustment, and follow-up interval scheduling. These variables are closely coupled and mutually restrictive. For example, choosing anterior traction appliances usually requires a longer daily wearing time to ensure suture remodeling, but excessively long wearing times can reduce child compliance; larger force applications can shorten treatment duration, but may exceed the safe tolerance range of periodontal tissues. Faced with such multidimensional and complex constraints, dentists with insufficient clinical experience, especially young dentists in primary healthcare institutions, find it difficult to develop optimal individualized orthodontic plans that balance efficacy, safety, and feasibility for children at different stages of growth and development. This significantly limits the promotion and application of early childhood orthodontic techniques at the primary care level.

[0004] Chinese patent application CN117992912A discloses an artificial intelligence system for predicting dentofacial growth and development. This system integrates multimodal learning to gather biological information from multiple sources, including physical characteristics, secondary sexual characteristics, metabolite levels, bone age, and dental age. It then uses a multi-layer LSTM network to fuse features from each modality layer by layer and employs a softmax function to predict growth and development. This approach improves the accuracy of growth and development stage identification to some extent and reduces the professional knowledge required of operators. However, the above approach still has the following shortcomings: First, the system only achieves classification prediction of growth and development status, outputting discrete developmental stage labels rather than continuous numerical predictions for specific cephalometric indicators such as SNA angle, SNB angle, and ANB angle. Therefore, it cannot provide quantitative growth trend information for subsequent orthodontic treatment design, and physicians still lack numerical decision support when formulating treatment plans. Secondly, the system completely lacks a systematic optimization and selection of treatment parameters. After obtaining growth and development prediction results, dentists still need to rely entirely on their personal clinical experience and subjective judgment to determine key treatment parameters such as appliance type, wearing duration, and force adjustment plan. The decision-making process lacks systematicity, scientific rigor, and repeatability. Thirdly, the system lacks a dynamic monitoring and plan feedback adjustment mechanism during treatment. It cannot automatically suggest adjustments to treatment parameters based on the deviation between the child's actual growth response and the expected trajectory during long-term treatment. This makes it difficult to correct deviations in a timely manner once the initial plan does not match the actual progress, resulting in insufficient flexibility and individualized adaptation of the plan.

[0005] In summary, there is an urgent need for an intelligent planning technology for early childhood orthodontic treatment that can deeply integrate quantitative prediction of jawbone growth trajectory with multi-constraint collaborative optimization of orthodontic parameters, and achieve closed-loop deviation monitoring and adaptive adjustment of the treatment plan throughout the entire treatment process. This technology would help dentists quickly generate individualized optimal treatment plans while ensuring safety. Although there have been attempts to apply artificial intelligence to orthodontics in existing technologies, most have focused on the automatic identification of cephalometric landmarks or the classification and diagnosis of maxillofacial deformities. There is still no publicly available literature or patent that organically integrates continuous numerical prediction of jawbone growth, Pareto multi-objective optimization of multivariate orthodontic parameters, and closed-loop dynamic adjustment during the treatment process into a complete technical solution. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides an intelligent planning method and system for early childhood orthodontic treatment programs. This method deeply couples quantitative prediction of jawbone growth trajectory with multi-constraint collaborative optimization of orthodontic parameters, and establishes a closed-loop deviation monitoring and adaptive parameter adjustment mechanism during the orthodontic execution process, thereby achieving intelligent planning throughout the entire process from data acquisition, growth prediction, program optimization to dynamic adjustment.

[0007] This invention discloses an intelligent planning method for early orthodontic treatment in children. The method first collects lateral cephalometric radiographs of children in the mixed dentition stage and extracts multiple cephalometric parameters, including the SNA angle, SNB angle, ANB angle, mandibular plane angle, anterior to posterior height ratio, and inclination of the upper and lower incisors. Simultaneously, it acquires individual information about the child, including their current age, sex, bone age assessment, and family genetic facial type classification. All the cephalometric parameters and the individual information are then concatenated into a unified-dimensional growth status feature vector. Subsequently, this growth status feature vector is input into a jawbone growth prediction model based on a multilayer fully connected network. This model, trained on a large-sample longitudinal growth database, outputs a sequence of predicted natural growth trends for each of the cephalometric parameters within a preset prediction time window, thus providing a quantitative baseline for individual growth trends in subsequent optimization.

[0008] Based on growth prediction, a multi-constraint orthodontic parameter co-optimization problem is constructed using appliance type coding, recommended daily wearing time, applied force value, and follow-up adjustment cycle as decision variables. The dual objective function is to minimize the weighted combination of the deviation of the ANB angle from the ideal value at the end of treatment and the length of the treatment course. Inequality constraints are set for the maximum acceptable daily wearing time for patient compliance and the periodontal tissue safety applied force threshold. Furthermore, an improved multi-objective genetic algorithm is used to search for Pareto optimal solutions to this co-optimization problem. During the fitness evaluation phase, a jawbone growth prediction model is invoked for each candidate solution. The predicted value of the natural growth trend is superimposed with the correction amount of the orthodontic force effect generated by the candidate solution. The expected values ​​of each cephalometric index at the final treatment state are calculated and substituted into the dual objective function for evaluation. After multiple generations of iterative convergence, multiple alternative treatment solutions on the Pareto front are output. Each solution is labeled with the expected treatment course, final state index value, and compliance requirement level, allowing dentists and parents to jointly consult and choose the solution most suitable for the family's actual situation.

[0009] In addition, during the treatment, the actual cephalometric values ​​of the child are collected according to the preset review cycle. The actual measurement values ​​are compared with the predicted trajectory values ​​at the corresponding time of the selected plan. When the deviation exceeds the preset deviation threshold, the growth status feature vector is automatically updated with the current actual measurement value and the optimization search is re-executed. The adjusted treatment plan is then output, thereby realizing a closed-loop intelligent planning of prediction, optimization, monitoring and adjustment.

[0010] This invention also provides an intelligent planning system for early childhood orthodontic treatment. This system includes a cephalometric feature acquisition and encoding module, a jaw growth trajectory prediction module, an orthodontic parameter optimization problem construction module, a Pareto optimal solution search module, and an orthodontic trajectory deviation monitoring and adjustment module. Each module corresponds one-to-one with the steps of the aforementioned method. Specifically, the Pareto optimal solution search module deeply couples with the jaw growth trajectory prediction module during fitness assessment. When the deviation exceeds the limit, the orthodontic trajectory deviation monitoring and adjustment module triggers the re-execution of both the jaw growth trajectory prediction module and the Pareto optimal solution search module, forming a closed-loop feedback loop. This ensures that the various modules of the system form a closely coordinated closed-loop architecture.

[0011] The beneficial effects of this invention are as follows: It deeply couples jawbone growth prediction with multi-objective optimization of treatment parameters, enabling fitness assessment to fully consider the nonlinear impact of individualized growth trends on treatment outcomes, avoiding the disconnect between growth prediction and treatment plan design in traditional methods. Through Pareto optimality search, it outputs multiple balanced solutions for joint decision-making by both doctors and patients, achieving a scientific balance between optimal therapeutic effect and patient compliance. Through closed-loop deviation monitoring and adaptive parameter adjustment throughout the treatment process, it ensures that the treatment plan maintains dynamic optimality over long periods spanning months or even years, significantly improving the scientific rigor and individualization of clinical decision-making. Furthermore, this invention incorporates the actual degree of cooperation of children wearing orthodontic appliances into the optimization model through a compliance decay factor, significantly improving the feasibility of the output plan and effectively avoiding theoretically optimal but practically unfeasible plans that children cannot consistently adhere to—a problem that has not been systematically considered and addressed in previous orthodontic intelligent assisted decision-making research. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the intelligent planning method for early childhood intervention programs in this embodiment of the invention.

[0013] Figure 2 This is a schematic diagram of the architecture of the intelligent planning system for early childhood intervention programs in this embodiment of the invention. Detailed Implementation

[0014] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate specific implementations of the present invention and do not constitute a limitation on the scope of protection of the present invention. Unless otherwise specified, the features in the following embodiments can be combined with each other.

[0015] See Figure 1The intelligent planning method for early childhood intervention provided in this invention includes steps S1 to S5, forming a deeply coupled closed-loop collaborative architecture. Step S4 directly calls the jawbone growth prediction model from step S2 during fitness assessment, constituting feedforward deep coupling. Step S5 triggers the re-execution of steps S2 and S4 when the deviation exceeds the limit, forming a closed-loop feedback loop. Each step is described in detail below.

[0016] Step S1: Multidimensional Cephalometric Feature Acquisition and Growth Status Encoding. In this step, a standard lateral cephalometric radiograph is first taken of the child during the mixed dentition period. Preferably, the distance from the focal point of the X-ray cephalometric machine to the center of the film is set to 1.524m. The child's head is fixed in the cephalometric machine with an ear rod and nose pad to ensure that the midsagittal plane is parallel to the film. In one embodiment of the present invention, a trained operator locates marker points on the acquired lateral cephalometric radiograph and extracts a total of 12 cephalometric indicators. The specific definitions and clinical significance of these 12 indicators are as follows.

[0017] The first six items are angular indicators: the SNA angle is the angle formed by the center of the sella turcica (S), the nasal root (N), and the upper alveolar seat (A), with a normal value of approximately 82°, reflecting the sagittal position of the maxilla relative to the anterior cranial base; the SNB angle is the angle formed by the center of the sella turcica (S), the nasal root (N), and the lower alveolar seat (B), with a normal value of approximately 80°, representing the sagittal position of the mandible relative to the anterior cranial base; the ANB angle is the difference between the SNA angle and the SNB angle, with a normal value of approximately 2°, and is the most crucial indicator for assessing the sagittal relationship between the maxilla and mandible; the mandibular plane... The MP-FH angle is defined by the angle between the mandibular plane and the orbitoauricular plane, with a normal value of approximately 26°. It reflects the rotational trend of the mandible and the vertical characteristics of the facial features. The Y-axis angle is defined by the angle between the line connecting the center of the sella turcica to the apex of the chin and the orbitoauricular plane, with a normal value of approximately 59.4°. It comprehensively assesses the direction of facial growth. The Wits value is defined as the distance between the projections of the upper and lower alveolar points on the functional occlusal plane. The normal value is approximately -1 mm for males and approximately 0 mm for females. It is an important supplementary indicator for independently assessing the sagittal relationship of the mandible and maxilla after excluding the influence of the skull base.

[0018] The last six items are line distance and ratio indicators: the normal value of the anterior skull base length SN is about 71 mm, which serves as the baseline length for craniofacial analysis; maxillary protrusion, which is the distance from point A to the perpendicular line N, reflects the degree of maxillary protrusion; the mandibular body length Go-Gn is usually 65 mm to 75 mm during the mixed dentition period and is a key parameter for assessing the level of mandibular development; the mandibular ramus height Co-Go is about 45 mm to 55 mm in this age group; the ratio of anterior height N-Me to posterior height S-Go has a normal value of about 0.62 to 0.65, and a significant increase in this ratio suggests a tendency for a high-angle facial profile due to excessive vertical facial development; and the inclination of the upper and lower incisors, which are defined as the angles between the long axes of the upper and lower incisors and the anterior skull base plane or the mandibular plane, respectively.

[0019] After extracting the 12 cephalometric measurements, this step requires collecting four pieces of individual information about the child: current age recorded in months to the nearest whole number; gender represented by binary codes of 0 and 1; bone age assessment using the Fishman Bone Maturity Staging System's 11-level quantification standard, represented by integers 1 to 11; and familial genetic facial type classification represented by uniquely heated vector codes for three categories: convex, straight, and concave. In one embodiment of this invention, the 12 cephalometric measurements are Z-score standardized to ensure that each indicator has zero mean and unit variance when input to the network, avoiding the adverse effects of dimensional differences on subsequent model training. After standardization, the 12-dimensional measurements are concatenated with the 4-dimensional individual information to form a 16-dimensional growth state feature vector. This vector will serve as the sole input to the jawbone growth prediction model in the subsequent step S2.

[0020] It is worth noting that the feature acquisition and encoding steps in this process are not simply data transfers, but rather provide a standardized data foundation for all subsequent steps. Preferably, a consistency verification mechanism is introduced during the landmark location process: when different operators locate landmarks on the same lateral cephalometric radiograph, the system automatically calculates the difference between the two locations. If the coordinate deviation of any landmark exceeds 1.5mm, a relocation is prompted to ensure the reliability of the input data. Furthermore, for systematic measurement errors caused by slight deviations in the shooting angle, this invention introduces a compensation algorithm based on orbitoauricular plane orientation correction. This algorithm corrects the original measurements to the equivalent values ​​corresponding to the standard shooting posture before subsequent encoding, thereby reducing the interference of non-standard shooting operations on the final prediction results. In clinical practice, the above verification and compensation mechanisms effectively improve the robustness of the data acquisition process, reducing the skill requirements of the operators and making the system more suitable for widespread use in primary healthcare institutions.

[0021] Step S2: Jawbone growth trajectory prediction based on a multi-layer fully connected network. The core task of this step is to predict the natural growth trajectory of various cephalometric indicators of the child over the next 24 months based on the 16-dimensional growth state feature vector output from Step S1. In one embodiment of the present invention, the prediction time window is set to 24 months with a time step of 3 months. Therefore, the model outputs predicted values ​​of 12 indicators at 8 time points, with an output dimension of [missing information]. dimension.

[0022] The jawbone growth prediction model employs a multi-layer fully connected network architecture, specifically comprising one input layer, four hidden layers, and one output layer. The input layer receives a 16-dimensional growth state feature vector. The first hidden layer contains 128 neurons, the second hidden layer contains 256 neurons, the third hidden layer contains 256 neurons, and the fourth hidden layer contains 128 neurons. ReLU activation and Dropout random deactivation layers are applied sequentially between each hidden layer; in one embodiment of this invention, the Dropout probability is set to 0.3. The output layer contains 96 neurons, corresponding to the predicted values ​​of 12 indicators at each of the eight time points. No activation function is applied to the output layer to support continuous numerical regression.

[0023] Preferably, this invention introduces an age-decaying attention weight mechanism in the hidden layer to enhance the model's sensitivity to features of the adolescent growth spurt. Specifically, the age-decaying attention weight vector is defined as follows: ,in, For the first The feature vectors of the hidden layer have the same dimension as the number of neurons in that layer. The trainable attention projection matrix has a dimension equal to the number of neurons in the hidden layer multiplied by 12, mapping the hidden features to a 12-dimensional index space. It is the bias vector; The temperature coefficient is calculated as follows: ,in, The base temperature coefficient is set to 1.0 in one embodiment of the present invention; This is the decay rate parameter, with a value of 0.15, which is the reciprocal of the number of years. ; The current bone age assessment value of the child is expressed in Fishman stage equivalent age, in years; The reference age for the peak growth spurt during puberty is 13.5 years for males and 11.5 years for females. When the child's bone age approaches the peak growth period, the temperature coefficient... The larger temperature coefficient allows attention to be more focused on indicators sensitive to growth changes, thus enhancing the model's ability to capture rapid growth characteristics during peak periods. As bone age moves away from peak periods, the temperature coefficient decreases, attention distribution becomes more uniform, and the model treats the slow changes in various indicators in a relatively balanced manner.

[0024] For model training, in one embodiment of this invention, longitudinal cephalometric measurements of no fewer than 5,000 children aged 6 to 12 years with a follow-up period of no less than 24 months are used as the training dataset. The training dataset is divided into a training set, a validation set, and a test set in a 7:1.5:1.5 ratio. The loss function is mean squared error (MSE), the optimizer is Adam, and the initial learning rate is set to... The training process employs a cosine annealing learning rate scheduling strategy, with a total of 200 training rounds. The jawbone growth prediction model of this invention achieves a root mean square error (RMSE) of approximately 0.85° for the ANB angle and approximately 1.2 mm for the mandibular body length (Go-Gn) on the test set, demonstrating superior accuracy compared to the traditional Ricketts growth prediction method.

[0025] It is important to note that the construction of the aforementioned training dataset must adhere to a strict data quality control process. Preferably, the longitudinal cases included in the training dataset should meet the following conditions: the children did not receive any orthodontic treatment intervention during the follow-up period to ensure that the recorded cephalometric measurements all originated from natural growth; the interval between each follow-up visit did not exceed 6 months to ensure sufficient temporal resolution to support the model's refined learning of growth trends; all follow-up records were acquired using cephalometric equipment from the same institution under standardized operating procedures to eliminate systematic biases introduced by equipment differences. During the data preprocessing stage, outlier removal was performed on records with significant measurement anomalies, such as those where the difference between a measurement and subsequent measurements exceeded three times the normal annual growth rate of that indicator. Furthermore, considering the significant differences in jawbone growth patterns among children of different sexes and ethnicities, this invention incorporates gender information as an explicit input feature during model training, enabling the model to automatically learn the influence of gender differences on the growth rate of various indicators without requiring separate predictive models for male and female children. To address potential racial differences among children from different regions, the pre-trained model can be fine-tuned through transfer learning by collecting longitudinal data from the target population.

[0026] Step S3: Construction of a Co-optimization Problem for Multi-Constrained Orthodontic Parameters. Based on the predicted growth trends output in Step S2, this step transforms the design of early orthodontic treatment plans into a solvable multi-objective optimization problem. A core innovation of this invention lies in formalizing the clinically experience-dependent plan design process into a constrained mathematical optimization problem, enabling computers to systematically search for optimal combinations of solutions. This transformation process not only requires precise mathematical modeling of various variables in clinical decision-making but also necessitates converting implicit expert knowledge and rules of thumb in orthodontics into quantifiable constraints and objective functions. This is a crucial bridging link in achieving intelligent planning in this invention.

[0027] The set of decision variables is defined as follows ,in For the coding of orthodontic appliance type, one can choose one of three types: functional appliance, anterior traction appliance, and bow expander, using a unique hot vector form; The recommended daily wearing time is taken from 8 to 14 hours, with a step size of 0.5 hours. To determine the applied force, the value should be taken within the range of 50g to 400g, with a step size of 10g. To adjust the follow-up period, values ​​were taken within the range of 4 to 12 weeks, with a step size of 1 week.

[0028] The biobjective function is defined as: The first objective function is the deviation of the final ANB angle from the ideal value: ,

[0029] In decision variables End of treatment The expected value of the ANB angle is obtained by the fitness assessment process in step S4 by calling the growth prediction model in step S2 and superimposing the corrective force effect, and the unit is degrees (°). The ideal target value for the ANB angle is 2° in one embodiment of this invention. The second objective function is the length of the treatment course. , In decision variables The total number of months required to reach the termination criteria for treatment, expressed in months. The termination criteria are defined as the entry of the ANB angle deviation. The acceptable range, or reaching the preset maximum treatment duration of 24 months. The inequality constraints include two items. The first is a compliance constraint: , The first value represents the maximum acceptable daily wearing time for the child's current age, obtained through fitting clinical experience data. It is set at 10 hours for children aged 6-8, 12 hours for children aged 8-10, and 14 hours for children aged 10-12. The second value is periodontal safety constraints. , The recommended safe force thresholds for periodontal tissue application are as follows, depending on the type of orthodontic appliance: 300g for functional appliances, 400g for anterior traction appliances, and 350g for palpebral expanders. Exceeding these thresholds may lead to adverse consequences such as root resorption or periodontal ligament damage.

[0030] The aforementioned constraints reflect the dual consideration of patient safety and treatment feasibility in the optimization of the treatment plan in this invention. Preferably, in actual clinical application, dentists can further adjust the safe force threshold according to the specific periodontal condition of the child. For example, for children with significant root resorption of deciduous teeth or poor periodontal support tissue, the dentist can lower the safe threshold to 80% of the default value; for children with good periodontal condition and high bone density, it can be appropriately raised to 120% of the default value. This flexible adjustment mechanism based on clinical knowledge allows the optimization framework of this invention to adapt to actual clinical scenarios with significant individual differences. In addition, this invention also considers the mutual exclusion constraint between appliance types, that is, only one type of appliance can be selected in the same optimization plan, avoiding force system conflicts that may arise from the simultaneous use of different types of appliances. In some complex cases that require the sequential use of different appliances, the treatment process can be divided into multiple stages and optimization problems can be constructed for each stage separately, with the final state of the previous stage used as the initial state input for the optimization of each stage.

[0031] Step S4: Pareto Optimal Solution Search Based on Improved Multi-Objective Genetic Algorithm. This step applies the improved NSGA-II multi-objective genetic algorithm to search for Pareto optimal solutions to the multi-objective optimization problem constructed in Step S3. In one embodiment of the present invention, the population size is set to 100 to 300 individuals, where 100 individuals is the lower limit for maintaining population diversity; when the population size is less than 100, the Pareto front coverage of the multi-objective problem is insufficient. 300 individuals is the upper limit for balancing computational cost and convergence speed; when the population size exceeds 300, the computation time per generation increases significantly while the improvement in the Pareto front is not significant. In one embodiment of the present invention, 200 individuals are preferably selected, and the gene encoding of each individual corresponds to the decision variables defined in Step S3. Among them, the types of orthodontic appliances Recommended daily wearing time is encoded using integers. Increase strength value and follow-up period Real-number encoding is used. The crossover operation employs simulated binary crossover (SBX), with a crossover probability set between 0.8 and 0.95. 0.8 is the lower limit for effective transmission of genetic information in the population; below 0.8, offspring diversity decreases, leading to premature convergence. 0.95 is the upper limit to avoid excessive destruction of the genes of superior individuals; above 0.95, the information of elite individuals is difficult to retain stably. In one embodiment of this invention, 0.9 is preferred, and the distribution index is set to 20. The mutation operation employs polynomial mutation (PM), with a mutation probability set between 0.01 and 0.05. 0.01 is the lower limit for maintaining the population's exploration ability; below 0.01, the algorithm is prone to getting trapped in local optima. 0.05 is the upper limit to ensure the convergence stability of superior solutions; above 0.05, excessive population randomness leads to a decrease in convergence speed. In one embodiment of this invention, 0.02 is preferred, and the distribution index is set to 20. The maximum number of iterations is set to 200 to 500 generations, where 200 generations is the lower limit of the initial convergence of the Pareto front, and the algorithm has not yet fully converged when the number of generations is less than 200; 500 generations is the upper limit of the input-output ratio of computing resources, and the fitness improvement is less than 0.1% after 500 generations, which is not practically meaningful; in one embodiment of the present invention, 300 generations is preferred.

[0032] The key innovation of this step lies in the deep coupling between the fitness assessment process and the jawbone growth prediction model in step S2. Specifically, for each candidate scheme in the population... The fitness assessment is performed according to the following procedure: First, the jaw growth prediction model in step S2 is called to obtain the predicted values ​​of the natural growth trend of the child at each time point within the 24-month prediction time window. Then according to The correction amount for the corrective force effect is calculated based on the type of orthodontic appliance and the applied force value. Finally, the two values ​​are combined to obtain the expected values ​​of each indicator at the final state of treatment: ,in, For the natural growth prediction model at the corrective terminal state time The output is a 12-dimensional vector of predicted index values; In the first time step (In this embodiment, the incremental vector of corrective force effect generated by the candidate scheme within 3 months) depends on the type of orthodontic appliance. and increase strength value The values ​​were obtained through a pre-established mapping table of corrective force and bone remodeling effect, with units consistent with those of each indicator. The compliance decay factor is defined as:

[0033] ,in, This is the compliance and comfort threshold corresponding to the child's age. If the child's compliance is below this threshold, it is considered that they can wear the orthodontic appliance completely as required. This indicates that the corrective force is completely effective; The decay rate constant is 2.5 in one embodiment of this invention. This value is obtained based on a statistical fit of wearing time and actual wearing rate in clinical follow-up data and is dimensionless. When the recommended daily wearing time exceeds the comfort threshold, The exponential decay reflects the clinical fact that children voluntarily reduce the wearing time due to discomfort, leading to a decrease in the proportion of effective corrective force. In one embodiment of the invention, for children aged 6 to 8 years... 8 hours for children aged 8 to 10, 10 hours for children aged 10 to 12, and 11 hours for children aged 10 to 12.

[0034] After fitness evaluation, the algorithm performs elite retention selection on the population according to non-dominated sorting and crowding distance sorting. Non-dominated sorting divides all individuals in the population into multiple frontal layers based on Pareto dominance, with the first frontal layer consisting of all non-dominated optimal individuals. Crowding distance measures the distance between each individual within the same frontal layer and its neighbors in the target space; a larger crowding distance indicates greater isolation in the target space, representing a more unique trade-off on the Pareto front. The selection process prioritizes retaining individuals at the frontal layers and, within the same layer, prioritizes retaining individuals with larger crowding distances, thus maintaining the diversity of the solution set while promoting the overall convergence of the population towards the Pareto front.

[0035] Furthermore, this invention introduces a feasibility rule to address constraint violation issues. When a candidate solution violates compliance or periodontal safety constraints, the system calculates the amount of constraint violation and incorporates it as a penalty in the fitness assessment. Specifically, individuals with larger constraint violation amounts are at a disadvantage during the selection process, but they are not directly eliminated because their genetic information may contribute to offspring through crossover operations, thereby generating feasible high-quality solutions.

[0036] After 300 iterations, the algorithm converges to output a set of alternative solutions on the Pareto front. Preferably, 3 to 5 solutions with representative balance between efficacy and treatment duration are selected from the Pareto front as the final output. The selection strategy is as follows: first, select the extreme solutions at both ends, i.e., the solution with the best efficacy and the solution with the shortest treatment duration; then, select 1 to 3 balanced solutions in the middle region based on the principle of prioritizing the solution with the greatest crowding distance. Each output solution is associated with the following information: recommended appliance type, recommended daily wearing time, force applied, follow-up adjustment cycle, expected total treatment duration in months, expected final ANB angle value and other key indicator predictions, and a level label for the child's compliance requirements. The compliance level is divided into three levels: low requirement, medium requirement, and high requirement, corresponding to a recommended daily wearing time of no more than 10 hours, 10 to 12 hours, and more than 12 hours, respectively. Dentists and parents can negotiate and choose the most suitable solution from these alternatives based on the family's actual situation.

[0037] Step S5: Orthodontic Trajectory Deviation Monitoring and Adaptive Parameter Adjustment. After the dentist and parents select the orthodontic plan and begin its implementation, this step is responsible for continuously monitoring the consistency between the treatment progress and the predicted trajectory of the plan throughout the entire treatment process. When significant deviations occur, the plan is automatically re-optimized, thus forming a closed-loop feedback loop from step S5 back to steps S2 and S4. This closed-loop mechanism is one of the core innovative features that distinguishes this invention from existing technologies.

[0038] In one embodiment of the present invention, the preset follow-up period is 3 months, consistent with the time step of the growth prediction model in step S2. At each follow-up examination, a new lateral cephalometric radiograph is taken of the child, and the actual value vector of the 12 cephalometric indicators at the current moment is extracted according to the method in step S1. Simultaneously, the predicted trajectory value corresponding to the review time point is extracted from the selected scheme output in step S4. The deviation is calculated using weighted Euclidean distance: ,in, For the first During the second follow-up examination The actual measured value of the cephalic shadow measurement index; For the selected options at this time point The predicted value of each indicator; For the first The weight coefficient of each indicator reflects the importance of the indicator in the treatment goal. In one embodiment of the present invention, the weight of the ANB angle is set to 0.3, the weight of the SNA angle and the SNB angle are each set to 0.15, the weight of the mandibular plane angle is set to 0.1, and the remaining 8 indicators are evenly distributed with the remaining weight of 0.3. The sum of all weights is 1.0.

[0039] Deviation threshold The threshold is dynamically set according to the treatment stage. In the early stage (first 6 months after treatment begins), the deviation threshold is relatively lenient at 2.5, allowing for greater deviation fluctuations, as the child needs time to adapt to the appliance. In the middle stage (6-12 months), the threshold narrows to 2.0. In the late stage (after 12 months), the threshold is further narrowed to 1.5 to ensure the precise achievement of the treatment goals. This staged threshold strategy is based on the fact that deviations in the early stages are more due to adaptive fluctuations than to incompatibility with the treatment plan itself; premature re-optimization may lead to frequent plan changes, which is detrimental to treatment continuity. In the late stage, as the final goal is approached, more precise trajectory control is required.

[0040] When the deviation value is calculated during a certain review Exceeding the deviation threshold corresponding to the current stage At this point, the system automatically executes the following re-optimization process: First, it reconstructs the growth status feature vector using the 12 actual measurements from the current follow-up examination and the updated child's age and bone age information, replacing the initial output of step S1; then, it calls the jawbone growth prediction model from step S2, using the updated feature vector as input to re-predict the natural growth trend within the remaining prediction time window; next, in step S3, it reconstructs the optimization problem with the remaining treatment duration as a constraint. At this point, the range of decision variables may be adjusted due to treatment progress; for example, the determined type of orthodontic appliance generally remains unchanged, with adjustments mainly made to the wearing duration and the applied force; finally, in step S4, it re-executes the Pareto search, outputting the adjusted alternative solutions. The adjusted solutions also include information on expected efficacy and compliance requirements, which are evaluated by the physician to determine whether to adopt. It is worth noting that the re-optimization process is not a simple repetition of the initial optimization, but rather an incremental adjustment based on the inherited treatment progress information. Specifically, the search space during re-optimization is typically smaller than that during initial optimization because the appliance type remains largely unchanged. The search focuses primarily on adjusting wearing time and force, resulting in significantly lower computational overhead for re-optimization, usually completed within seconds. Furthermore, the population initialization strategy during re-optimization differs: the system retains the parameters of the current treatment plan as an elite individual within the population and generates initial individuals based on random perturbations within its neighborhood. This guides the search to quickly explore better adjustment directions near the current plan, avoiding drastic changes to the treatment plan that might result from a blind global search. This incremental re-optimization strategy ensures smooth and clinically acceptable treatment adjustments, preventing dentists and parents from facing the confusion of completely altered treatment plans after each follow-up visit.

[0041] Through the aforementioned closed-loop mechanism, this invention can continuously track treatment progress throughout the entire treatment process, promptly identify and correct deviations between the treatment plan and the actual outcome, ensuring that the treatment plan remains dynamically optimal throughout long-term treatment spanning months or even years. In contrast, traditional methods rely solely on physicians' subjective experience to determine whether adjustments to the treatment plan are needed during follow-up examinations, lacking quantitative deviation assessment standards and systematic re-optimization methods, which easily leads to problems such as delayed or insufficient adjustments. The closed-loop architecture of this invention transforms passive experience-based judgment into proactive data-driven decision-making, representing a fundamental upgrade from open-loop control to closed-loop control.

[0042] See Figure 2 The present invention also provides an intelligent planning system for early childhood orthodontic treatment. The system includes a cephalometric feature acquisition and encoding module, a jaw growth trajectory prediction module, an orthodontic parameter optimization problem construction module, a Pareto optimal solution search module, and an orthodontic trajectory deviation monitoring and adjustment module. The above five modules correspond one-to-one with steps S1 to S5 in the method embodiment, and together they form a closed-loop architecture of prediction-optimization-monitoring-adjustment.

[0043] The cephalometric feature acquisition and encoding module is configured to receive digitized images of the child's lateral cephalometric radiographs and clinically reported individual information. In one embodiment, this module integrates an automatic cephalometric landmark identification subunit and a feature standardization subunit. The automatic cephalometric landmark identification subunit uses a pre-trained deep convolutional neural network to automatically locate key landmarks such as the center of the sella turcica (S), the root of the nose (N), the upper alveolar fossa (A), and the lower alveolar fossa (B) on the digitized lateral cephalometric radiograph, and then automatically calculates the aforementioned 12 cephalometric index values ​​based on the landmark coordinates. In one embodiment of the invention, the convolutional network is a ResNet-50-based regression model, with an average error of no more than 1.2 mm on the landmark localization task, meeting the accuracy requirements of clinical cephalometric measurement. The feature standardization subunit performs Z-score standardization processing on the 12 index values ​​as described in the method embodiment, and concatenates them with the 4-dimensional individual information to form a 16-dimensional growth state feature vector, which is then output to the jawbone growth trajectory prediction module. This module's automated marker recognition capability significantly reduces reliance on manual tracing experience, enabling primary healthcare institutions to quickly obtain high-quality cephalometric measurement data. Preferably, the module also integrates an automatic measurement value verification function. When the system-identified marker position deviates from the normal anatomical range, it automatically prompts the operator for manual review, ensuring that outliers do not enter subsequent prediction and optimization processes. This protective mechanism is crucial for improving the system's reliability in real clinical environments.

[0044] The jawbone growth trajectory prediction module is configured to receive the aforementioned 16-dimensional growth state feature vector and output a sequence of predicted natural growth trends for 12 indicators at 8 time points within a 24-month prediction time window through a pre-trained multi-layer fully connected network jawbone growth prediction model. The network architecture, age-decay attention weight mechanism, and training method of this module are consistent with those described in step S2 of the method embodiment. During system deployment, this module is embedded in the system as an inference model with offline training and fixed weights, supporting CPU-based operation on ordinary clinical workstations. A single inference iteration takes no more than 50ms, meeting the needs of real-time clinical interaction. Preferably, this module also maintains a growth prediction result cache. When the fitness evaluation in step S4 is called multiple times, the natural growth baseline data in the cache can be directly read, avoiding the computational overhead of repeated inference.

[0045] The orthodontic parameter optimization problem construction module is configured to automatically construct a bi-objective optimization problem defined in step S3 of the method embodiment based on clinical configuration parameters. This module internally maintains a knowledge base of orthodontic appliance parameter constraints, storing parameters such as the force range, recommended wearing time intervals, and safety thresholds for various orthodontic appliances. Dentists can manually modify the default constraint range on the system interface based on actual clinical judgment; for example, they can appropriately relax the upper limit of wearing time for children with particularly good compliance, or tighten the upper limit of force for children with poor periodontal conditions. This module merges the manually adjusted parameters with the default parameters to generate the final optimization problem description, which is then passed to the Pareto optimal solution search module.

[0046] The Pareto optimal solution search module is configured to execute the improved NSGA-II multi-objective genetic algorithm described in step S4 of the method embodiment. Preferably, this module concurrently calls the jawbone growth trajectory prediction module to obtain the natural growth baseline during the fitness assessment process, and works in conjunction with the built-in orthodontic force-bone remodeling effect mapping engine to complete the calculation of the final state indices for all candidate solutions. This mapping engine is built based on finite element analysis results and clinical retrospective data, storing a table of expected changes in the cephalometric indices for different appliance types under different applied force values, supporting interpolation queries. In specific implementation, the mapping engine organizes the data in the form of a three-dimensional lookup table, with the three dimensions corresponding to the appliance type index, applied force value interval, and action time step, respectively. Each element in the table is a 12-dimensional vector representing the single-step change increment of the 12 cephalometric indices under that condition combination. When the queried applied force value falls between two predefined intervals, the engine automatically performs linear interpolation to obtain continuous incremental estimates. The data sources for this mapping engine consist of two parts: one part comes from biomechanical simulation calculations based on a three-dimensional craniofacial finite element model, which obtains theoretical incremental values ​​by simulating the strain distribution and bone remodeling response generated by orthodontic forces of different directions and magnitudes in the jawbone suture region; the other part comes from statistical analysis of cephalometric changes before and after treatment in clinical retrospective data, used to calibrate and correct the theoretical values, making the mapping results closer to real clinical scenarios. After the search is completed, the module extracts 3 to 5 representative alternatives from the Pareto front, generates a structured protocol report, including protocol parameters, expected treatment duration, predicted final state indicators, and compliance level labels, and presents them to dentists and parents in the form of protocol comparison cards on a graphical user interface, supporting interactive filtering and protocol selection. Preferably, the user interface also provides a multi-dimensional indicator visualization comparison function in the form of radar charts, displaying the predicted final state indicators of each alternative protocol as a radar chart overlay, enabling dentists to intuitively compare the advantages and disadvantages of different protocols in various dimensions, assisting in making more scientific selection decisions.

[0047] The orthodontic trajectory deviation monitoring and adjustment module is configured to automatically receive cephalometric data collected during each of the next scheduled follow-up examinations at a preset 3-month interval during treatment. It then performs the weighted Euclidean distance deviation calculation and phased threshold judgment as described in step S5 of the method embodiment. When the deviation exceeds the limit, the module automatically triggers the jawbone growth trajectory prediction module and the Pareto optimal solution search module to re-execute prediction and optimization using the current actual data as input. The module also notifies the dentist of the deviation status and adjustment suggestions via a warning pop-up window on the user interface. Preferably, this module also maintains a historical orthodontic trajectory database, recording the actual measurement values, deviation values, and whether re-optimization was triggered for each follow-up examination. This provides complete data support for the dentist to review the entire orthodontic process and also provides real-world data accumulation for future model iteration training.

[0048] The information flow and collaborative relationships among the five modules constitute the core architectural advantage of this invention's system. Specifically, the cephalometric feature acquisition and encoding module provides input features to the jawbone growth trajectory prediction module, forming a data-driven feedforward pathway; the jawbone growth trajectory prediction module provides a growth baseline to the orthodontic parameter optimization problem construction module, enabling the optimization objective to be based on individualized predictions rather than population averages; the Pareto optimal solution search module calls back to the jawbone growth trajectory prediction module during fitness evaluation, forming a deep coupling between prediction and optimization; the orthodontic trajectory deviation monitoring and adjustment module reactivates the prediction and search modules when the deviation exceeds the limit, forming a complete closed-loop feedback loop. This deeply coupled and closed-loop collaborative architecture makes the overall system performance far exceed the effect of simply superimposing the modules, demonstrating significant nonlinear synergistic gains.

[0049] To verify the effectiveness of the method and system of the present invention, a retrospective verification experiment was conducted in one embodiment. The experimental data came from the medical records of children with functional malocclusion during the mixed dentition period treated at a tertiary-level dental hospital between 2018 and 2023, and a total of 612 cases meeting the inclusion criteria were selected. The inclusion criteria included: age between 6 and 12 years old, diagnosed with at least one of the following functional malocclusions: anterior crossbite, deep overbite, or malocclusion; complete lateral cephalometric radiographs before and after treatment; and complete follow-up records for at least 18 months.

[0050] The 612 cases were divided into a training / validation set and a test set in an 8:2 ratio. 490 cases in the training / validation set were used for training the jawbone growth prediction model and hyperparameter tuning, while 122 cases in the test set were used for final effect evaluation. Evaluation indicators included three aspects: growth prediction accuracy, evaluated using the RMSE and mean absolute error (MAE) of each cephalometric index; plan optimization quality, comparing the recommended plan output by this invention with the actual clinically implemented plan in terms of final ANB angle deviation and treatment duration; and closed-loop adjustment effectiveness, statistically analyzing the improvement in final efficacy after re-optimization of the plan in the subset with deviation exceeding limits.

[0051] Regarding growth prediction accuracy, the jaw growth prediction model of this invention achieved an RMSE of 0.83° for the ANB angle, 0.91° for the SNB angle, and 1.15 mm for the mandibular body length on the test set. In comparison, the traditional Ricketts prediction method achieved an RMSE of 1.52° for the ANB angle and 1.78° for the SNB angle on the same test set. This demonstrates that the prediction accuracy of the model in this invention is approximately 40% to 50% higher than that of traditional methods. This improvement is primarily attributed to the age-decaying attention weighting mechanism's ability to refine the capture of adolescent growth characteristics.

[0052] Regarding the quality of the treatment plan optimization, a comparison was made with the actual treatment plans developed by clinical experts. In 122 test cases, the Pareto optimal plan group recommended by this invention always included a plan with a final ANB angle deviation of less than 1°, resulting in a 100% usability rate. For the 89 cases that actually completed treatment, the average deviation of the final ANB angle after simulation using the parameters of this invention's recommended plan from the ideal value was 0.72°, while the corresponding average deviation of the actual clinical plan was 1.35°, indicating that this invention's plan improved treatment accuracy by approximately 47%. In terms of treatment duration, the average treatment duration of the recommended plan was 14.6 months, while the average clinical treatment duration was 17.8 months, a reduction of approximately 18%.

[0053] Regarding the effectiveness of closed-loop adjustments, 34 out of 89 cases that actually completed treatment experienced deviation exceeding limits during the treatment process. In these 34 cases, the system automatically suggested parameter adjustment plans. Retrospective analysis showed that if adjustments were made according to the system's suggestions, the final ANB angle deviation could be further reduced by 0.4° to 0.6°, while avoiding unnecessary extension of the treatment course. Further analysis of the causes of deviation exceeding limits revealed that approximately 41% of deviation exceeding limits were due to the child's actual compliance with the appliance being lower than expected, approximately 35% were due to the child entering puberty and experiencing accelerated growth, resulting in a natural growth rate exceeding the model's prediction, and approximately 24% were due to external factors such as improper appliance wearing or aging and deformation of the appliance itself. The above analysis shows that the closed-loop deviation monitoring mechanism of the present invention can not only provide timely warnings and corrections when deviations occur, but also provide dentists with clues for analyzing the causes of deviations, helping them to take targeted clinical intervention measures, such as strengthening guidance and education on wearing the device for children with poor compliance, or adjusting the frequency of follow-up examinations to once every two months for children entering the growth acceleration period in order to more closely track growth changes.

[0054] Furthermore, this invention also evaluated the system's operational efficiency. On a typical clinical workstation equipped with a 12th-generation Intel Core processor and 16GB of memory, the average time for a single inference iteration of the jawbone growth prediction model was 38ms, the average time for a 300-generation genetic algorithm optimization search with a population of 200 individuals was 45s, and the time for calculating deviations in a single review was no more than 10ms. These time performance indicators demonstrate that the system provided by this invention can fully meet the needs of real-time clinical interaction, allowing dentists to obtain intelligent planning results during the child's visit without waiting for time-consuming offline calculations.

[0055] Based on the above verification results, the intelligent planning method and system for early childhood orthodontic treatment provided by this invention has achieved significantly better results than traditional methods in three dimensions: growth prediction accuracy, plan optimization quality, and closed-loop adjustment effectiveness. It can effectively assist dentists, especially young dentists in primary healthcare institutions, in formulating scientific and individualized early childhood orthodontic treatment plans.

[0056] To further evaluate the applicability of this invention in different malocclusion types, 122 cases in the test set were grouped and statistically analyzed according to three malocclusion types: anterior crossbite, deep overbite, and malocclusion. In the anterior crossbite group (52 cases), the average deviation of the final ANB angle of the recommended treatment plan was 0.68°, significantly better than the 1.41° of the actual clinical plan, and the average treatment duration was shortened by approximately 2.8 months. In the deep overbite group (43 cases), the average deviation of the final ANB angle was 0.79°, also better than the 1.27° of the clinical plan. Furthermore, this invention showed particularly good optimization effects in controlling overbite depth, and all Pareto treatment plans included at least one feasible approach to reduce overbite depth to the normal range. In the malocclusion group (27 cases), due to the lateral asymmetry involved in malocclusion correction, the limitations of relying solely on sagittal cephalometric measurements for optimization became apparent. The improvement in final deviation was approximately 35%, lower than the 45% to 52% improvement in the first two groups, suggesting that future versions could consider incorporating orthogonal radiograph data to enhance the optimization capability for lateral malocclusions.

[0057] The above group validation results confirm the significant clinical application value of this invention in the sagittal-dominant functional malocclusion correction. Simultaneously, these results provide clear data support for the iterative improvement of this invention, demonstrating the scientific rigor and scalability of the system during continuous optimization, and laying a solid data and technical foundation for subsequent system upgrades and iterations. In terms of clinical application, the hardware requirements for this invention are standard cephalometric radiography equipment and a general clinical workstation, without relying on high-end computing servers. This makes its deployment in county-level and above medical institutions highly feasible, helping to narrow the gap in early orthodontic treatment plan design capabilities between primary healthcare institutions and large dental hospitals.

[0058] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. An intelligent planning method for early childhood intervention programs, characterized in that, Includes the following steps: Step S1: Collect lateral cephalometric radiographs of children in the mixed dentition period and extract multiple cephalometric index values, including SNA angle, SNB angle, ANB angle and mandibular plane angle. Combine these with the individual information of the child, which consists of the child's age, gender, bone age assessment value and family genetic facial type classification information, and splice them into a growth status feature vector. Step S2: Input the growth status feature vector into a jawbone growth prediction model pre-trained based on a large sample longitudinal growth database. The jawbone growth prediction model contains multiple hidden layers and sets nonlinear activation functions and random deactivation layers between each hidden layer. Output the natural growth trend prediction value sequence of each of the cephalometric index values ​​of the child within a preset prediction time window. Step S3: Using appliance type coding, recommended daily wearing time, applied force value, and follow-up adjustment cycle as decision variables, minimizing the weighted combination of the final ANB angle deviation and the treatment course as the dual objective function, and using the maximum wearing time for compliance and the periodontal safe applied force threshold as constraints, a collaborative optimization problem is constructed. Step S4: Perform a multi-objective genetic algorithm search on the collaborative optimization problem. During the fitness evaluation stage, call the jawbone growth prediction model for each candidate solution, add the predicted value of the natural growth trend to the correction force effect correction amount of the candidate solution to calculate the final state index value of the correction and substitute it into the bi-objective function. After iterative convergence, output multiple alternative solutions on the Pareto front. Step S5: During the treatment, collect actual cephalometric index values ​​according to the preset review cycle, calculate the deviation between the actual cephalometric index values ​​and the predicted trajectory values ​​at the corresponding time of the selected plan, and update the growth state feature vector with the current actual measurement value when the deviation exceeds the preset deviation threshold, and re-execute steps S3 and S4 to output the adjusted treatment plan.

2. The method according to claim 1, characterized in that, In step S1, the cephalometric index values ​​also include anterior skull base length, maxillary protrusion, mandibular body length, mandibular ramus height, Y-axis angle, and Wits value. The number of cephalometric index values ​​is 12, and the dimension of the growth state feature vector is 16, of which 12 dimensions correspond to the cephalometric index values ​​and 4 dimensions correspond to the individual information of the child.

3. The method according to claim 1, characterized in that, In step S2, the preset prediction time window is 24 months, the jaw growth prediction model outputs the predicted values ​​of the cephalometric indexes at 8 time points with a time step of 3 months, and the large sample longitudinal growth database contains no less than 5,000 longitudinal cephalometric follow-up records of children aged 6 to 12 years.

4. The method according to claim 1, characterized in that, In step S3, the orthodontic type coding adopts one-hot vector representation, and the coding space covers functional orthodontic appliances, anterior traction devices, and bow expanders. The recommended daily wearing time ranges from 8h to 14h, the applied force ranges from 50g to 400g, and the follow-up adjustment cycle ranges from 4 weeks to 12 weeks.

5. The method according to claim 1, characterized in that, In step S2, age decay attention weights are introduced into each hidden layer of the jawbone growth prediction model. The age decay attention weights dynamically adjust the contribution of each cephalometric index value in the feature transformation of the hidden layer according to the distance between the child's current bone age and the peak of puberty, so that the indexes close to the peak of puberty obtain higher feature response gains.

6. The method according to claim 1, characterized in that, In step S4, the corrective force effect correction amount is adjusted by a compliance decay factor. The compliance decay factor is determined based on the functional relationship between the recommended daily wearing time and the child's age. When the recommended daily wearing time exceeds the compliance comfort threshold corresponding to the child's age, the compliance decay factor reduces the effective proportion of the corrective force effect correction amount according to an exponential decay model.

7. The method according to claim 1, characterized in that, In step S4, the multi-objective genetic algorithm adopts an elite retention strategy based on crowding distance, with a population size of 100 to 300 individuals, a crossover probability of 0.8 to 0.95, a mutation probability of 0.01 to 0.05, and an iteration number of 200 to 500 generations. In each generation fitness evaluation, the jawbone growth prediction model is called in parallel to complete the prediction of the final state index of all candidate treatment schemes.

8. The method according to claim 1, characterized in that, The number of alternative treatment options on the Pareto front output in step S4 is 3 to 5. Each alternative treatment option is associated with expected treatment course information, the numerical values ​​of each of the cephalometric indicators of the expected final state of treatment, and a level label of the required level of cooperation from the child.

9. The method according to claim 1, characterized in that, In step S5, the preset review period is 3 months, the deviation calculation uses the weighted Euclidean distance between the actual value and the predicted trajectory value of each of the cephalometric index values, and the preset deviation threshold is dynamically set according to the correction stage. The preset deviation threshold in the early stage of correction is greater than the preset deviation threshold in the later stage of correction.

10. An intelligent planning system for early childhood intervention programs, used to implement the method described in any one of claims 1-9, characterized in that, include: The cephalometric feature acquisition and encoding module is configured to acquire lateral cephalometric radiographs of children in the mixed dentition period and extract multiple cephalometric index values. Combined with the child's age, gender, bone age assessment value and family genetic facial pattern classification information, a growth status feature vector is generated. The jawbone growth trajectory prediction module is configured to receive the growth state feature vector and output a sequence of natural growth trend prediction values ​​of each of the cephalometric index values ​​within a preset prediction time window based on a pre-trained multi-layer fully connected network jawbone growth prediction model. The orthodontic parameter optimization problem construction module is configured to use the orthodontic appliance type code, recommended daily wearing time, applied force value and follow-up adjustment cycle as decision variables, minimize the ANB angle deviation at the end of treatment and the treatment course as dual objective functions, and use the child's compliance constraint and periodontal safety constraint as inequality constraints to construct a collaborative optimization problem. The Pareto optimal solution search module is configured to perform a multi-objective genetic algorithm search on the collaborative optimization problem. During the fitness evaluation stage, the jawbone growth trajectory prediction module is called to superimpose the natural growth trend prediction value with the correction force effect correction to calculate the final state index value of the correction. After iterative convergence, multiple alternative correction solutions on the Pareto front are output. The correction trajectory deviation monitoring and adjustment module is configured to collect actual cephalometric index values ​​at a preset review cycle during the correction period, calculate the deviation between the actual values ​​and the predicted trajectory values ​​at the corresponding time of the selected plan, and trigger the jawbone growth trajectory prediction module and the Pareto optimal plan search module to re-execute when the deviation exceeds the preset deviation threshold, and output the adjusted correction plan.

Images