Fracture healing process monitoring method and system based on callus growth simulation

Abstract

The invention discloses a fracture healing process monitoring method and system based on callus growth simulation, and relates to the field of intelligent medicines.The method comprises the steps that callus growth paradigm data of fracture healing are collected through a big data method, and a combined training data set is constructed based on the callus growth paradigm data; initializing a regression model, and training the regression model according to the joint training data set in combination with a knowledge distillation thought to obtain a growth rate mapping model; acquiring online disease information of a target patient, and initializing a configuration stage reference callus model according to the online disease information; performing callus growth simulation based on a dynamic growth rate in combination with a preset time step length to obtain a stage prediction callus model; and iteratively monitoring the fracture healing process of the target patient according to the stage prediction callus model. The problems that in the prior art, callus growth real-time prediction is inaccurate, healing prediction lacks adaptability, and reliability is insufficient are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent healthcare, specifically to a method and system for monitoring the fracture healing process based on callus growth simulation. Background Technology

[0002] Bone callus formation is a complex biological process involving the coupled effects of multiple factors, including mechanical stimulation, cellular activity, angiogenesis, and bone remodeling. Accurate monitoring and prediction of the healing process are crucial for optimizing clinical treatment plans, providing rehabilitation guidance, and early warning of complications.

[0003] Currently, monitoring of callus growth mainly relies on qualitative comparisons of a series of medical images and clinical experience, lacking objective and quantitative dynamic assessment methods. Existing research, based on quantitative analysis using biomechanical simulations, establishes cell activity models, which can reveal the deep mechanisms of healing. However, these models are complex to construct, computationally expensive, and time-consuming, and heavily depend on expert parameter settings. Consequently, they cannot predict callus growth in real time, and healing predictions lack adaptability and reliability.

[0004] Therefore, an innovative approach is needed that integrates prior biomechanical knowledge, clinical big data, and efficient computational frameworks to achieve rapid, accurate, and individualized dynamic monitoring and prediction of the fracture healing process. Summary of the Invention

[0005] This application provides a method and system for monitoring the fracture healing process based on callus growth simulation, which addresses the problems of inaccurate real-time prediction of callus growth and insufficient adaptability and reliability of healing prediction in the prior art.

[0006] In view of the above problems, this application provides a method for monitoring the fracture healing process based on callus growth simulation.

[0007] In a first aspect, this application provides a method for monitoring the fracture healing process based on callus growth simulation, the method comprising: Using big data methods, we collect case data on callus growth for fracture healing and construct a joint training dataset based on the case data. The joint training dataset includes a first training dataset and a second training dataset. Initialize the regression model, and train the regression model based on the joint training dataset and the knowledge distillation concept to obtain the growth rate mapping model; Obtain online symptom information of the target patient, and initialize the configuration stage baseline callus model based on the online symptom information; Based on the stage baseline callus model and the growth rate mapping model, and combined with a preset time step, callus growth simulation based on dynamic growth rate is performed to obtain the stage prediction callus model. Based on the stage-predicted callus model, the fracture healing process of the target patient is monitored iteratively.

[0008] Secondly, the present invention provides a fracture healing process monitoring system based on callus growth simulation, comprising: The data acquisition and processing module is used to collect callus growth example data for fracture healing through big data methods, and to construct a joint training dataset based on the callus growth example data, wherein the joint training dataset includes a first training dataset and a second training dataset. The regression model building module is used to initialize the regression model and train the regression model based on the joint training dataset, combined with the knowledge distillation idea, to obtain the growth rate mapping model. The baseline model construction module is used to acquire online symptom information of the target patient and initialize the configuration stage baseline callus model based on the online symptom information. The prediction model construction module is used to simulate callus growth based on dynamic growth rate by combining the stage baseline callus model and the growth rate mapping model with a preset time step, and obtain the stage prediction callus model. The discrimination module is used to predict the callus model based on the stage and iteratively monitor the fracture healing process of the target patient.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes a method for monitoring fracture healing based on callus growth simulation. First, it collects callus growth paradigm data, including simulation cases and clinical images, using big data methods to construct a joint training dataset. This breaks the limitations of traditional methods that rely on a single data source, integrating high-fidelity simulation data and real clinical data to provide a comprehensive, multimodal data foundation for subsequent model training. Second, it initializes a regression model and uses the joint training dataset, combined with the concept of knowledge distillation, to train a growth rate mapping model. This transforms the biological laws in the complex prior simulation model into the regression model, achieving an accurate mapping from patient status to local callus growth rate and reducing computational costs. Third, it acquires online disease information of the target patient and initializes a configuration stage baseline callus model, achieving precise digital modeling of the individual patient's fracture condition, providing a foundation for subsequent simulations and ensuring the targeting and accuracy of the monitoring process.

[0010] Simultaneously, based on the stage-based baseline callus model and growth rate mapping model, and combined with a preset time step, callus growth simulation based on dynamic growth rate is performed to obtain a stage-predictive callus model. This model predicts macroscopic morphological changes in the callus, reveals the spatiotemporal distribution of internal growth rate, and meticulously assesses the progress and quality of the healing process, providing a deeper basis for intervention decisions. Finally, based on the iterative monitoring of the fracture healing process of target patients using the stage-predictive callus model, a closed-loop monitoring system is constructed to objectively assess the healing status, provide quantitative decision support for clinicians, and realize the dynamic management of individualized healing processes. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the fracture healing process monitoring method based on callus growth simulation proposed in this application.

[0013] Figure 2 This is a schematic diagram of the fracture healing process monitoring system based on callus growth simulation proposed in this application.

[0014] In the attached diagram, the components represented by each number are as follows: Data acquisition and processing module 11; regression model construction module 12; benchmark model construction module 13; prediction model construction module 14; operation and discrimination module 15. Detailed Implementation

[0015] This application provides a method for monitoring the fracture healing process based on callus growth simulation, which solves the problems of inaccurate real-time prediction of callus growth and lack of adaptability and reliability in healing prediction.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0018] The present invention will now be described in detail with reference to the accompanying drawings.

[0019] Example 1, as Figure 1 As shown, this application provides a method for monitoring the fracture healing process based on callus growth simulation, the method comprising: S10: Using big data methods, collect callus growth example data for fracture healing, and construct a joint training dataset based on the callus growth example data, wherein the joint training dataset includes a first training dataset and a second training dataset. In this embodiment of the application, the big data method refers to the technology of collecting, cleaning and integrating large-scale, multi-dimensional and heterogeneous callus growth-related data through distributed storage, parallel computing, data mining and other technologies; callus growth example data refers to data that includes the entire cycle of callus formation, growth and remodeling during fracture healing, specifically including simulation case data and clinical imaging data; the joint training dataset refers to a training data set that integrates annotation information from different sources and of different types.

[0020] By collecting and integrating multi-source data, the problem of lack of clinical realism in single simulation data and high cost of labeling single clinical data is solved, providing a reliable data source for subsequent model training. A labeling method combining soft and hard targets is introduced, which utilizes the refined features of simulation data while anchoring to the real results of clinical data, thereby improving the information density of the training dataset. The joint training dataset covers simulation and clinical scenarios, enabling the subsequently trained model to not only match the biomechanical laws of simulation but also adapt to individual differences in clinical practice, laying the foundation for the practical application of the model.

[0021] In step S10 of the method provided in this application embodiment, the callus growth example data includes simulation case data and clinical imaging data.

[0022] In this embodiment, clinical data of fracture patients and corresponding fracture types and treatment plans from the simulation model library are collected. Postoperative CT images are acquired for each patient, and the image data is stored. The actual volume, morphology, and spatial distribution characteristics of callus at each time point are extracted for each patient. The image feature data are correlated and labeled with the patient's basic information, clinical examination data, treatment plan, and healing outcome to form a complete clinical image dataset. The standardized simulation case dataset is merged with the clinical image dataset to provide a foundation for the construction of a subsequent joint training dataset.

[0023] Step S10 of the method provided in this application embodiment, which involves constructing a joint training dataset based on the callus growth example data, includes: Based on the simulation case data, a first training dataset is constructed by combining the corresponding prior simulation model and clinical validation image data, wherein the first training dataset includes soft targets; Based on clinical imaging data, bimodal analysis is performed using the prior simulation model to construct a second training dataset, which includes soft targets and hard targets. The first training dataset and the second training dataset are merged to output the joint training dataset.

[0024] In this embodiment, a first training dataset is first constructed based on simulation case data, combined with the corresponding prior simulation model and clinically validated imaging data. The first training dataset includes soft targets. The prior simulation model refers to a clinically validated, high-precision simulation model of callus growth that couples cellular activity and biomechanics, capable of outputting virtual biofield data during callus growth. The soft targets refer to intermediate feature data before the output layer of the prior simulation model, which cannot be directly obtained through clinical testing and is used to assist the regression model in learning the knowledge of the simulation model.

[0025] From the simulation case data of callus growth paradigms, simulation data matching the clinical validation image data scenarios are selected and input into the prior simulation model. After running the simulation calculation, the virtual biofield data before the model output layer is extracted as the soft target. The output results of the simulation case data are compared with the corresponding clinical validation image data, and simulation cases with excessive deviations are eliminated. Finally, the sample state features of the simulation cases are associated with the corresponding soft targets to form the first training dataset.

[0026] Secondly, based on clinical imaging data, a bimodal analysis was performed using a prior simulation model to construct a second training dataset. This second training dataset includes soft and hard targets. Hard targets refer to the actual callus growth indicators directly calculated from clinical imaging data, serving as the final anchoring standard for the regression model's predictions and ensuring that the model output closely matches clinical reality. Bimodal analysis involves simultaneously analyzing the actual callus characteristics from clinical imaging data and the virtual biofield characteristics output from the prior simulation model, achieving the fusion of real clinical data and simulation mechanism data, and providing soft and hard dual-target annotations for the second training dataset.

[0027] Clinical imaging data from callus growth paradigms are segmented and reconstructed to extract the first callus state features, such as actual morphology, volume, and spatial distribution. Neighborhood temporal change analysis is performed on clinical imaging data of the same patient at different time points to calculate the true indices of the callus body as hard targets. The sample state features are input into a priori simulation model, and after simulation calculation, virtual biofield data before the output layer is extracted as soft targets. Finally, the sample state features, hard targets, and soft targets are correlated to form a second training dataset. Each sample contains an annotation structure of input features, hard targets, and soft targets, achieving a two-dimensional annotation of clinical real results and simulation mechanisms.

[0028] Finally, the first and second training datasets are merged to output a joint training dataset. The input feature dimensions and data types of the first and second training datasets are standardized to the same standard. The first and second training datasets are then merged to form the final joint training dataset, which serves as the basis for training the subsequent regression model.

[0029] Step S10 of the method provided in this application embodiment, which involves constructing a first training dataset based on the simulation case data, combined with the corresponding prior simulation model and clinical validation image data, includes: Based on the simulation case data, extract the first sample state features corresponding to each simulation case; By combining the simulation case data with the prior simulation model corresponding to the simulation case data, the virtual biofield data before the output layer is matched and extracted, and the output is the first soft target. The first sample state features are merged with the first soft target to obtain the first training dataset.

[0030] In this embodiment, firstly, based on the simulation case data, the first sample state features corresponding to each simulation case are extracted. The first sample state features refer to the set of key parameters extracted from the simulation case data that characterize the initial conditions and environment of callus growth; the virtual biofield data before the output layer is intermediate feature data generated by the prior simulation model during calculation, without being mapped by the output layer; the first soft target refers to the virtual biofield data extracted from the prior simulation model that matches the simulation case data, serving as the target value for the first training dataset.

[0031] Based on the influencing factors of callus growth, the dimensions of the first sample state features are determined; for each set of simulation case data, sample state features are extracted, where categorical variables are encoded using one-hot encoding and continuous variables are normalized; the encoded first sample state features are stored according to the simulation case number to form a feature matrix, ensuring that each set of simulation case data corresponds to a unique feature vector.

[0032] For example, features were extracted and coded from a set of simulated closed tibial fracture case data: fracture location [1,0,0], fixation method [1,0,0], displacement degree 8mm (normalized to 0.53), bone mineral density T-score -1.5 (normalized to 0.5), age 50 years (normalized to 0.6), postoperative day 7 (normalized to 0.06), and initial callus volume 50mm². 3 After normalization, the initial stress value is 10 MPa, and after normalization, it is 0.33, forming the first sample state feature vector {[1,0,0],[1,0,0],0.53,0.5,0.6,0.06,0.1,0.33}.

[0033] Secondly, by combining the simulation case data with the corresponding prior simulation model, the virtual biofield data before the output layer is matched and extracted, and the output is used as the first soft target. The initial parameters of each set of simulation case data are input into the prior simulation model, and the callus growth simulation calculation is run. The second-to-last layer before the output layer of the prior simulation model is located, and the virtual biofield data is extracted, including: stress field distribution, nutrient factor concentration field, and osteoblast activity field. The selected virtual biofield data is used as the first soft target to form a target matrix.

[0034] For example, the simulation case data is input into the prior simulation model, and after running the simulation calculation, the virtual biofield data before the output layer is extracted: stress field distribution: stress values ​​of 100 grids, ranging from 5 to 18 MPa; nutrient factor concentration field: ranging from 0.2 to 0.8 mol / L; osteoblast activity field: ranging from 0.1 to 0.9. The vectors are merged and used as the first soft target.

[0035] Finally, the first sample state features and the first soft target are merged to obtain the first training dataset. The number of cases for the first sample state features is ensured to be exactly the same as the number of cases for the first soft target, with each case number corresponding to the others. The first sample state feature vector of each group of cases is concatenated with the first soft target vector to form sample data. The assembled sample data is then converted into a format recognizable by the machine learning model, divided proportionally into training and validation subsets, and labeled to form the complete first training dataset.

[0036] For example, the first sample state feature vector is concatenated with the soft target vector to form sample data; the same operation is repeated for simulation cases to form the first training dataset of samples, the format is converted, and it is divided into a training subset and a validation subset in a 7:3 ratio. After labeling the metadata, the construction of the first training dataset is completed.

[0037] In step S10 of the method provided in this application embodiment, the method performs bimodal analysis based on clinical imaging data and the prior simulation model to construct a second training dataset. The second training dataset includes soft targets and hard targets, including: The clinical image data is segmented and analyzed using machine vision methods to obtain the first callus state characteristics. Based on the first callus state characteristics, a neighborhood temporal change analysis is performed to calculate and obtain the true generation rate as the second hard target; The second sample state features corresponding to the clinical imaging data are obtained, and the second sample state features are input into the prior simulation model to extract the virtual biofield data before the output layer as the second soft target. The second training dataset is constructed by combining the second sample state features, the second hard target, and the second soft target.

[0038] In this embodiment, clinical image data is first segmented and analyzed using machine vision methods to obtain first callus state features. The machine vision method refers to deep learning-based medical image analysis technology, which uses a semantic segmentation model to automatically identify, segment, and extract features from callus regions in clinical images. The first callus state features refer to the external morphological features of the callus directly extracted from clinical images, reflecting the actual growth state of the callus.

[0039] Clinical image data undergoes format conversion, noise reduction, and grayscale normalization to unify slice thickness and field of view. Callus region segmentation is performed using a pre-trained model. The model input is the grayscale matrix of the CT image, and the output is a binary mask of the callus region. Based on the segmented callus mask, the first callus state features are calculated, including: callus volume, surface area, density distribution, and fracture end distance. Callus volume = number of mask pixels × voxel volume; surface area is calculated using a 3D surface reconstruction algorithm; density distribution is calculated as the mean / standard deviation of CT values ​​within the callus region; and fracture end distance is the shortest distance from the callus centroid to the fracture end. These features are quantified into numerical vectors to obtain the characteristic vector of the first callus state features.

[0040] For example, the first callus state feature extraction: After preprocessing the CT images of a 50-year-old male patient one week post-surgery, the images are input into a U-Net segmentation model to obtain a callus region mask; the callus volume of 50 mm² is calculated based on the mask. 3 Surface area 120mm 2 The average CT value is 350 HU and the distance between the fracture ends is 2 mm, forming the first callus state feature vector [50, 120, 350, 2].

[0041] Secondly, based on the characteristics of the first callus state, a neighborhood temporal change analysis was performed to calculate the true generation rate as the second hard target. Neighborhood temporal change analysis refers to comparing adjacent time windows of callus regions in clinical images of the same patient at different time points, calculating the rate of change of callus characteristics over time. The hard target refers to the true indicators of callus growth obtained through clinical image temporal analysis, which is the final anchoring standard for the regression model's prediction results and possesses the authenticity of clinical measurement.

[0042] Image registration was performed on consecutive time points of the same patient to ensure complete spatial alignment of the fracture site across different images, avoiding analytical errors caused by changes in patient position. Neighborhood feature comparison was conducted to extract the first callus state features at adjacent time points after pairing, and the feature change amount and rate of change were calculated. The callus volume growth rate was used as the core second hard target, supplemented by the density change rate as an auxiliary hard target, forming a hard target vector to ensure the hard targets possess the realism of clinical measurements. Specifically, the feature change amount is the difference in the first callus state features at adjacent time points; the feature change rate is the ratio of the change amount to time.

[0043] For example, CT images of the patient at 1 week and 2 weeks post-surgery were registered, and the callus volume of 120 mm² at 2 weeks post-surgery was extracted. 3 With a mean CT value of 380 HU, the calculated volume growth rate is (120-50) / 50 = 140% / week, and the density change rate is (380-350) / 350 ≈ 8.57% / week, forming the second hard target vector [140%, 8.57%].

[0044] Next, the second sample state features corresponding to the clinical imaging data are obtained and input into the prior simulation model. The virtual biofield data before the output layer is extracted as the second soft target. The second hard target is the true indicator of callus growth obtained through time-series analysis of clinical images, which is the final anchoring standard for the regression model's prediction results. The second sample state features refer to the individual patient characteristics and clinical treatment information corresponding to the clinical imaging data, which are the basic parameters input into the prior simulation model to generate the soft target. The second soft target refers to the virtual biofield data before the model's output layer extracted after inputting the second sample state features into the prior simulation model.

[0045] Patient information corresponding to clinical imaging data is obtained from the hospital system. Second sample state features are extracted, including: patient age, gender, bone mineral density T-score, fracture type, fixation method, postoperative time point, degree of fracture displacement, and initial callus volume. Continuous variables are normalized. Then, a priori simulation model is invoked, and the normalized second sample state features are input into the priori simulation model, with simulation parameters consistent with the clinical case. Subsequently, soft target extraction is performed, and the virtual biofield data before the output layer of the priori simulation model is extracted as the second soft target.

[0046] For example, the patient's second sample status characteristics were extracted as follows: age 55 years (normalized 0.7), bone mineral density T-score -1.5 (normalized 0.5), fracture type closed tibial fracture ([1,0,0]), fixation method intramedullary nail ([1,0,0]), postoperative time 6 days (0.05), displacement degree 6 mm (0.4), and initial callus volume 50 mm². 3 The initial stress value is 13MPa, and after normalization it is 0.4 to obtain the second sample state feature vector: {[1,0,0],[1,0,0],0.4,0.5,0.7,0.05,0.1,0.4}. After normalization, it is input into the prior simulation model; after running the simulation, the virtual biofield data before the output layer is extracted.

[0047] Finally, a second training dataset is constructed by combining the second sample state features, the second hard target, and the second soft target. Data dimension alignment is performed to ensure that the number of cases for the second sample state features, the second hard target, and the second soft target are completely consistent, corresponding one-to-one by patient ID. Then, the second sample state feature vector, the second hard target vector, and the second soft target vector for each group of cases are concatenated. The assembled sample data is then converted to a different format, divided proportionally into training and validation subsets, and metadata annotations are added, ultimately forming the complete second training dataset.

[0048] For example, the second sample state features, the second hard target, and the second soft target of the patient are concatenated, and the operation is repeated for all clinical patients to form a second training dataset. The format is converted and divided into 350 training subsets and 150 validation subsets in a 7:3 ratio to complete the construction of the second training dataset.

[0049] In this embodiment, a joint training dataset is constructed by integrating simulation case data and clinical imaging data. A lightweight growth rate mapping model is then trained using the knowledge distillation concept to achieve synergistic effects. By fusing the deep biophysical laws revealed by high-fidelity biomechanical simulation with real clinical imaging data, the actual healing situation is realistically reflected, addressing the problem of weak generalization ability of single-data-source models. Mapping growth rates from computationally intensive prior simulation models and encapsulating the mapping process into a regression model enables accurate prediction from the patient's current state to the local callus growth rate, providing a highly efficient quantitative prediction tool for clinical practice.

[0050] S20: Initialize the regression model and train it using the joint training dataset and the knowledge distillation concept to obtain the growth rate mapping model; In this embodiment, the regression model refers to a prediction model that takes the sample state characteristics of callus growth as input and the callus growth rate as output; the knowledge distillation idea refers to the technique of transferring the knowledge of complex prior simulation models to lightweight regression models; the growth rate mapping model refers to a prediction model that quickly outputs the growth rate of each region of the callus based on the input patient disease information after training.

[0051] By employing the concept of knowledge distillation, knowledge from prior simulation models is transferred to lightweight regression models, thereby improving the inference speed of the growth rate mapping model compared to prior simulation models. Through weighted fusion of multiple loss terms, the model learns the biomechanical laws of simulation data while remaining grounded in the realities of clinical data, addressing the disconnect between simulation models and clinical practice. The trained growth rate mapping model can then be deployed on clinical terminals without relying on high-performance simulation servers, reducing the hardware costs for technology implementation and enhancing the convenience of clinical application.

[0052] Step S20 in the method provided in this application embodiment includes: Construct a training loss function, which includes: Distillation loss term, wherein the distillation loss term is the loss of the regression model on the soft target of the first training dataset; The soft target loss term is the loss of the regression model on soft targets in the second training dataset. Hard target loss term, wherein the hard target loss term is the loss of the regression model on hard targets in the second training dataset; The distillation loss term, the soft target loss term, and the hard target loss term are weighted and fused together by coefficients that sum to 1 to form the training loss function.

[0053] In this embodiment, a distillation loss term is obtained, wherein the distillation loss term is the loss of the regression model on the soft target in the first training dataset. The training loss function is a mathematical function used to measure the difference between the predicted output of the regression model and the true target. It is the core basis for parameter optimization during model training, and the smaller the loss value, the higher the prediction accuracy of the model. The distillation loss term refers to the difference between the predicted value of the regression model on the soft target in the first training dataset and the soft target output of the prior simulation model.

[0054] Training distillation loss term: The mean squared error (MSE) is used to calculate the difference between the regression model's predicted value and the true soft target on the first training dataset. For each sample in the dataset, the difference between its true value y1 and the model's predicted value y1 is calculated. n1The difference is calculated, and the squared difference is then summed. Finally, this sum is divided by the total number of samples N1. Here, N1 is the number of samples in the first training dataset, y1 is the predicted value of the model for the soft target in the sample, and y... n1 The first training dataset contains the true soft targets.

[0055] For example, the first training dataset has 580 samples, and the second training dataset has 500 samples. Taking a certain sample, the model's predicted value for the soft target in the first training dataset is [15MPa, 0.8mol / L, 0.9], while the actual soft target value is [14.5MPa, 0.78mol / L, 0.88]. Then, the distillation loss component for this sample is (1 / 580)×(15-14.5). 2 +(0.8-0.78) 2 +(0.9-0.88) 2 ≈0.0017.

[0056] Obtain the soft target loss term, which is the loss of the regression model on the soft targets in the second training dataset. The soft target loss term refers to the difference between the predicted value of the regression model on the soft targets in the second training dataset and the soft targets output by the prior simulation model. It is used to strengthen the student model's learning of simulation mechanisms in clinical scenarios and ensure the adaptability of simulation knowledge to clinical data.

[0057] Similarly, MSE is used to calculate the difference between the regression model's predicted value and the true soft target value on the second training dataset. For each sample in the dataset, the difference between its true value y2 and the model's predicted value y2 is calculated. n2 The difference is calculated, and the squared difference is squared. The sum of the squared errors of all N2 samples is then calculated. Finally, this sum is divided by the total number of samples N2, where N2 is the number of samples in the second training dataset, and y2 is the predicted value of the model for the soft target of the sample. n2 The second training dataset contains the true soft targets.

[0058] For example, the model's predicted value for soft targets in the second training dataset is [15.2 MPa, 0.82 mol / L, 0.91], while the actual soft target value is [14.8 MPa, 0.79 mol / L, 0.89]. The soft target loss term component is (1 / 500) × (15.2 - 14.8). 2 +(0.82-0.79) 2 +(0.91-0.89) 2 ≈0.0003.

[0059] Obtain the hard target loss term, which is the loss of the regression model on hard targets in the second training dataset. The hard target loss term refers to the difference between the regression model's predicted value on hard targets in the second training dataset and the clinically measured hard targets. It is used to anchor the clinical realism of the model's prediction results and avoid the model deviating from reality.

[0060] The difference between the regression model's predicted values ​​on the hard targets of the second training dataset and the clinically measured hard targets was calculated using MSE. For each sample in the dataset, the difference between its true value y3 and the model's predicted value y3 was calculated. n3 The difference is calculated, and the squared difference is then summed. Finally, this sum is divided by the total number of samples N3, where N3 is the number of samples in the second training dataset, y3 is the predicted value of the model for the hard target of the sample, and y n3 This represents the callus growth rate measured in clinical practice.

[0061] For example, the predicted value of hard targets in the second training dataset is 135% / week, the volume growth rate is 140% / week, and the hard target loss component is (1 / 500)×(135-140). 2 =0.05.

[0062] The distillation loss term, soft target loss term, and hard target loss term are weighted and fused together using coefficients that sum to 1 to form the training loss function. Weighted fusion refers to assigning different weight coefficients to the three loss terms and forming the final training loss function through linear combination, thus achieving a balance between simulated knowledge learning and clinical real-world anchoring.

[0063] Based on the principle of balancing simulation knowledge learning with clinical reality anchoring, and combined with the results of preliminary experiments, the following coefficients were determined: distillation loss term coefficient, soft target loss term coefficient, and hard target loss term coefficient. The coefficients determined by the preliminary experiments were: distillation loss term coefficient + soft target loss term coefficient + hard target loss term coefficient = 1. The rationality of the coefficients was verified by the control variable method. If the hard target coefficient is too low, the model prediction results will deviate too much from the clinical measured values. If the distillation loss term coefficient is too low, the model will not be able to learn the biomechanical mechanism of callus growth, and the generalization ability will decrease. Finally, the above combination of coefficients was determined to be the optimal solution.

[0064] Finally, the three loss terms are weighted and summed according to their coefficients to construct the final training loss function, where the training loss function L = distillation loss term × distillation loss term coefficient + hard target loss term × hard target loss term coefficient + soft target loss term × soft target loss term coefficient. Samples from the joint training dataset are input into the initial regression model to calculate the initial loss value, verifying the computability of the function and the effectiveness of gradient propagation, ensuring that the loss function can guide the optimization of model parameters.

[0065] For example, the coefficients of the distillation loss term, the soft target loss term, and the hard target loss term are 0.3, 0.2, and 0.5, respectively. The total loss component for this sample is 0.3×0.0017+0.2×0.0003+0.5×0.05≈0.26. The average loss value is calculated by iterating through all samples, resulting in an initial total loss L=0.85. During model training, parameters are optimized using gradient descent, ultimately reducing the total loss to below 0.05 and keeping the prediction error within 5%.

[0066] In this embodiment, a weighted training loss function integrating distillation loss, soft target loss, and hard target loss is constructed to balance the granularity of model training, ensuring that the lightweight regression model learns while simultaneously monitoring the true growth rate. This prevents bias or overfitting caused by excessive reliance on a single target during model training, improves the prediction accuracy of the final growth rate mapping model, and enhances the reliability of the model output.

[0067] S30: Obtain online symptom information of the target patient and initialize the configuration stage baseline callus model based on the online symptom information; In this embodiment of the application, online disease information refers to the real-time clinical data of the target patient, including imaging data of the fracture site, basic patient information, clinical examination data, treatment plan, etc.; the stage benchmark callus model refers to the three-dimensional model of the initial state of the callus at the fracture site constructed in digital space based on the online disease information of the target patient. It is the starting model for callus growth simulation and corresponds to the actual state of the callus at a specific time point.

[0068] Fracture images of target patients are obtained through the hospital system, along with the patient's basic information, clinical examination data, and treatment plan. Based on the clinical data, initial biomechanical parameters are mapped to each grid to form the final stage benchmark callus model.

[0069] Step S30 in the method provided in this application embodiment includes: Based on the online medical information, a physical model of the fracture site is established in the digital space; Based on the shape characteristics of the physical model, the physical model is divided into spatial grids to obtain the stage reference callus model.

[0070] In this embodiment, a physical model of the fracture site is first established in the digital space based on online symptom information. The online symptom information refers to the clinical diagnostic data, medical imaging data, and treatment-related information of the target patient, which is the core basis for constructing the physical model. The physical model in the digital space refers to the digital geometric model of the fracture site constructed by three-dimensional reconstruction technology based on the patient's online symptom information, which is a precise digital replica of the patient's fracture site.

[0071] After preprocessing the patient's CT image data by denoising and slice thickness normalization, three-dimensional reconstruction is performed to generate a surface mesh model of the fracture site, accurately restoring the anatomical structure of the bone, the morphology of the fracture ends, and the position of the fixation device. Then, geometric parameter calibration is performed. Combined with the patient's clinical diagnostic data, the reconstructed surface mesh model is calibrated to adjust the bone density properties and the relative positions of the fracture ends, ensuring that the physical model is consistent with the patient's actual fracture state. If the patient already has initial callus, the callus region in the CT image is segmented using an image segmentation model, and the three-dimensional morphology of the callus is integrated into the physical model. The initial callus is labeled to form a complete physical model of the fracture site.

[0072] Noise reduction aims to effectively suppress random noise and artifacts introduced during image acquisition, improving the signal-to-noise ratio of the image while maintaining the clarity of tissue edges and fine structures. Slice thickness normalization unifies image sequences with different inter-slice resolutions acquired by different devices to the same spatial resolution through interpolation or resampling techniques. This ensures accurate registration between image data from different periods or sources.

[0073] Secondly, based on the shape characteristics of the physical model, spatial meshing is performed on the physical model to obtain a stage-based baseline callus model. Spatial meshing is the process of discretizing a continuous three-dimensional object or geometric region into a set of interconnected, regularly or irregularly shaped small units, which can be directly manipulated by numerical methods or used as the computational framework for the digital model. Therefore, the size and number of mesh units directly affect the accuracy of the calculation results.

[0074] Based on the geometric complexity of the fracture site, tetrahedral meshes were selected as the basic unit to balance computational accuracy and efficiency. An adaptive mesh generation strategy was adopted to set high-density meshes for key areas such as the fracture ends and callus regions. The quality of the segmented mesh model was checked to ensure the quality of the mesh units. The segmented and validated mesh model was then linked with the patient's clinical information, and the spatial coordinates, geometric parameters, and mechanical properties of each mesh unit were labeled to form a stage-based baseline callus model for subsequent callus growth simulation.

[0075] For example, 3D reconstruction is performed on the patient's CT images to generate a surface mesh model of the tibia. The fracture fragment displacement is calibrated to 8 mm, the bone elastic modulus is set to 18 GPa, corresponding to a T-value of -1.5, and the segmented initial callus is fused with a volume of 50 mm². 3A physical model of the initial callus was formed. The physical model was then adaptively segmented into tetrahedral meshes, with a mesh size of 1 mm for the fracture ends and callus region, and 5 mm for the normal tibia region, for a total of approximately 250,000 elements. The mesh quality was checked, with an average warpage of 8°, an average aspect ratio of 3:1, and no deformed meshes. The mesh model was then associated with patient information, and the coordinates, elastic modulus, and other parameters of each element were labeled to form a stage-based baseline callus model, which was then stored for subsequent growth simulation.

[0076] In this embodiment, a physical model of the fracture site is established and gridded in digital space using the patient's specific online symptom information. This enables the personalized and precise initialization of the stage benchmark callus model, providing a digital model that strictly matches the patient's actual anatomical structure, fracture type, and initial condition for subsequent growth simulation. This ensures that the entire monitoring process has highly individualized characteristics and improves the accuracy of prediction.

[0077] S40: Based on the stage baseline callus model and the growth rate mapping model, and combined with a preset time step, perform callus growth simulation based on dynamic growth rate to obtain a stage prediction callus model. In this embodiment, the time step refers to the time interval of the callus growth simulation, which is set according to the clinical pattern of callus growth to ensure that the simulation time resolution matches the clinical monitoring cycle; the dynamic growth rate refers to the rate value of the dynamic change of the growth rate of different areas of the callus with time, spatial location and individual patient characteristics, which is output in real time by the growth rate mapping model according to the callus status at each stage; the stage prediction callus model refers to the callus model obtained after a growth simulation of one time step, corresponding to the predicted callus status of the target patient at the next monitoring time point.

[0078] The information from the stage baseline callus model is input into the growth rate mapping model, and the results are output. Based on a preset time step, the newly added callus volume of each grid is superimposed onto the corresponding grid in the stage baseline callus model, updating and reconstructing the grid to form the final stage prediction callus model.

[0079] Step S40 in the method provided in this application embodiment includes: By combining the growth rate mapping model with the online symptom information of the target patient, the growth rate is predicted by traversing each spatial grid region of the stage benchmark callus model to obtain the growth rate of multiple units. The growth rates of multiple units are traversed, and time integration is performed along a preset growth direction to obtain the newly added bone tissue information of each spatial grid region. The newly added bone tissue information is aggregated, and the stage baseline callus model is updated using a geometry engine method to obtain the stage predicted callus model.

[0080] In this embodiment, the growth rate is first predicted by combining the growth rate mapping model with the online symptom information of the target patient, and then traversing each spatial grid region of the stage benchmark callus model to obtain the growth rate of multiple units.

[0081] For each grid cell of the stage baseline callus model, an input feature vector is constructed, with the dimension consistent with the input dimension of the growth rate mapping model. The feature vector of each grid cell is input into the trained growth rate mapping model, and the model outputs the callus growth rate of that cell. The predicted cell growth rate is validated for rationality, and outliers exceeding the physiological range are removed to ensure that the rate conforms to the biological laws of callus growth. Finally, a set of cell growth rates corresponding to each grid cell is formed.

[0082] For example, in a 50-year-old male patient, the staged baseline callus model one week after a closed tibial fracture surgery contained 250,000 mesh cells; a 12-dimensional feature vector was constructed for each cell and input into the growth rate mapping model; the predicted growth rate of the core region at the fracture ends was 2 mm. 3 / day, edge area is 0.5mm 3 / day, the normal bone area is 0, 20 abnormal rate values ​​are removed to form a unit growth rate set.

[0083] Secondly, the growth rates of multiple units are traversed, and time integration is performed along the preset growth direction to obtain the information on newly added bone tissue in each spatial grid region.

[0084] Based on clinical monitoring needs, a preset time step is established, and monitoring frequency and efficiency are calculated. Then, the growth direction is determined. Based on the biological laws of callus growth, the growth direction is set towards the fracture ends, prioritizing filling the fracture gaps, while non-gap areas grow radially along the bone surface. Next, an integral calculation is performed, integrating the growth rate of each grid cell over time using the formula: V = v × Δt × cosθ, where v is the cell growth rate, Δt is the time step, and θ is the angle between the growth direction and the cell normal direction, calculating the volume of newly added bone tissue in each cell. Simultaneously, combined with the density change rate output from the growth rate mapping model, the density of the newly added bone tissue is calculated, forming complete information on the newly added bone tissue.

[0085] For example, the preset time step is 1 week, and the growth direction is towards the fracture end; the integral calculation of the core unit at the fracture end is: V = 2 × 7 × cos0° = 14 mm 3 Edge element: V = 0.5 × 7 × cos30° ≈ 3 mm 3 This yields the volume of newly added bone tissue in each unit.

[0086] Finally, the newly added bone tissue information is aggregated, and the stage baseline callus model is updated using a geometric engine method to obtain the stage prediction callus model. The newly added bone tissue is aggregated, with information from each grid cell based on its spatial location, forming the overall distribution of callus growth. Then, the geometric model is updated by calling the geometric engine method to adjust the geometric dimensions of the corresponding grid cells according to the volume of the newly added bone tissue, merging the topology of adjacent cells, assigning values ​​to density changes, and updating the mechanical properties of the cells. Subsequently, the updated model is validated for geometric continuity and physical rationality, ensuring that there are no overlapping cells and that the mechanical properties conform to the physiological characteristics of bone tissue. Finally, the stage prediction callus model is output.

[0087] For example, by aggregating newly added bone tissue information, 140,000 mm of new bone tissue was added to the core area of ​​the fracture ends. 3 An additional 150,000 mm of material was added to the edge area. 3 The mesh element volume was updated, the topology of adjacent elements was merged, and the mechanical properties were updated, with the core element's elastic modulus increasing from 5 GPa to 8 GPa. After verifying that the model had no abnormalities, a stage prediction callus model was output at 2 weeks post-surgery, showing that the callus volume increased from 50 mm. 3 Increased to 290,000 mm 3 The gap filling rate at the fracture ends reaches 30%.

[0088] In this embodiment, efficient numerical simulation of the callus growth process based on dynamic rate is achieved by traversing grid cells to predict the growth rate, integrating over time, and updating the geometric model. The prediction of the callus growth rate is transformed into a dynamic prediction of the callus morphology. Simulating the complete dynamic process of the callus allows for the prediction of callus morphology at different time points, providing sufficient decision-making basis for more accurate rehabilitation decisions.

[0089] S50: Based on the predicted callus model of the stage, iteratively monitor the fracture healing process of the target patient.

[0090] In this embodiment of the application, iterative monitoring refers to the cyclical process of using the stage-predicted callus model as the new stage-based baseline callus model, obtaining the callus prediction model for subsequent time steps, and comparing and verifying it with the patient's actual clinical images; fracture healing process monitoring refers to the monitoring behavior of assessing issues such as callus growth by comparing the stage-predicted callus model with the patient's actual clinical images.

[0091] Based on the parameters of the stage-predicted callus model, the callus areas of the virtual images and the actual clinical images are compared. The actual callus model is then used as the new stage-based baseline callus model to enter the simulation of the next time step, and this process is repeated until the patient's fracture heals.

[0092] Step S50 in the method provided in this application embodiment includes: Each time the stage prediction callus model is obtained in the iteration, the comparative image data of the target patient is updated based on the preset time step. Based on the image acquisition parameters of the predicted callus model and the comparative image data, virtual images are generated to obtain reference image data. The reference image data is compared with the comparison image data and the residual is calculated. If the residual is greater than or equal to a preset residual constraint, a healing anomaly is reported, and the distribution of the residual is reported simultaneously.

[0093] In this embodiment, during each iteration of acquiring the stage-predicted callus model, the comparative image data of the target patient is updated based on a preset time step. The comparative image data refers to the clinical images of the fracture site of the target patient acquired at the preset time step node. These are real data reflecting the actual healing status of the patient and are used for comparison with the virtual images generated by the prediction model.

[0094] The generation of the stage prediction callus model is based on a preset time step and iterative time node matching is performed. After each iteration, the postoperative CT images of the patient are collected simultaneously at the clinical end to ensure that the prediction time and the actual image acquisition time are completely matched. Subsequently, standardized image acquisition is performed, and parameters are obtained according to a unified acquisition specification. The acquired comparison images are processed by noise reduction, grayscale normalization, and format conversion to remove motion artifacts and metal artifacts, ensuring that the images clearly reflect the actual growth state of the callus.

[0095] For example, after generating a staged predicted callus model for the patient 2 weeks post-surgery, CT images of the patient 2 weeks post-surgery were simultaneously acquired with a slice thickness of 1 mm and a tube voltage of 120 kV. After preprocessing, the actual callus growth status was clearly displayed: the callus volume on the right side of the fracture end was 10 mm². 3 There was no obvious callus on the left side.

[0096] Secondly, based on the image acquisition parameters of the stage-predicted callus model and the comparative image data, virtual images are generated to obtain reference image data. The image acquisition parameters refer to the equipment parameters and scanning conditions used when acquiring the comparative images; these are the key basis for generating the reference images, ensuring the comparability of the virtual images and the comparative images. The reference image data refers to the virtual clinical images generated based on the stage-predicted callus model, simulating the acquisition parameters of the comparative images. Its format, resolution, and imaging characteristics are consistent with the comparative images, serving as a bridge connecting the digital model and the clinical images.

[0097] The medical imaging simulation engine is invoked, and the acquisition parameters of the stage-predicted callus model and the comparison image are input. The imaging principle of the CT equipment is simulated to generate a virtual image with the same format as the comparison image. Subsequently, based on the callus prediction parameters, the X-ray attenuation coefficient of different tissues is simulated to make the gray value and contrast of the callus area in the reference image consistent with the real CT image. If the patient has implanted a fixation device, the morphology and distribution of metal artifacts are simulated in the reference image to ensure that the imaging characteristics of the virtual image are completely matched with the comparison image and to eliminate the comparison deviation caused by artifacts.

[0098] For example, the predicted callus volume at the 2-week postoperative stage was 14 mm on both sides of the fracture site. 3 Based on the CT acquisition parameters, a medical imaging simulation engine is called to generate a reference image to simulate the metal artifacts of the intramedullary nail. The CT value of the callus region is set to 350 HU.

[0099] Finally, the reference image data and the contrast image data are compared and the residuals are calculated. If the residual is greater than or equal to the preset residual constraint, an abnormal healing is reported, and the distribution of the residuals is also reported simultaneously. The residual refers to the pixel-level difference between the reference image and the contrast image in the callus region, quantified by the mean squared error index, reflecting the degree of deviation between the predicted model and the actual healing state. The residual constraint is a preset deviation threshold; exceeding the threshold indicates an abnormal healing. An abnormal healing refers to a state deviating from the normal healing process, such as a callus growth rate significantly lower or higher than the normal range, uneven callus distribution, or failure of callus filling at the fracture ends.

[0100] Bone callus region registration was performed using a rigid registration algorithm to spatially align the fracture sites of the reference and contrast images, ensuring a one-to-one pixel correspondence within the callus region and avoiding registration errors caused by changes in patient position. Residual quantization was then performed, calculating the mean squared error (MSE) between the reference and contrast images within the registered callus region. The grayscale values ​​of pixels at the same position (i,j) in I1 and I2 were subtracted to obtain the error at that point. This error value was then squared, and the squared errors over all H×W pixels were summed. This sum was divided by the total number of pixels in the image (H×W) to obtain the mean squared error. Here, I1 represents the pixel value of the reference image, I2 represents the pixel value of the contrast image, and H and W represent the height and width of the callus region. If the MSE ≥ a preset residual constraint, healing abnormality was identified. The distribution of residuals was visualized using a heatmap to identify the location and degree of deviation of abnormal areas. The abnormal results and residual heatmap were simultaneously fed back to clinicians to assist in developing intervention plans.

[0101] For example, the callus region was registered between the reference image and the contrast image. The calculated MSE was 35 on the right side and 90 on the left side of the fracture end. The residual constraint was 50, and the overall MSE was 62.5 ≥ 50, which was judged as abnormal healing. A residual heat map was generated, with the left side being the high residual red area and the right side being the low residual green area. This was fed back to the clinician, indicating that the callus growth on the left side of the fracture end had stopped.

[0102] In this embodiment, a closed-loop online dynamic monitoring system is constructed by iteratively comparing the virtual images generated by the stage-predicted callus model with the patient's actual follow-up images, and calculating and analyzing the residuals, to objectively evaluate the actual healing trajectory. If healing is abnormal, timely warnings can be issued and the spatial area where the abnormality occurs can be located, providing alarm information to doctors and realizing intelligent management of the patient's healing status.

[0103] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, a joint training dataset is first constructed by integrating simulation case data and clinical imaging data. A lightweight growth rate mapping model is then trained using the knowledge distillation concept to achieve synergistic effects. By fusing the deep biophysical laws revealed by high-fidelity biomechanical simulation with real clinical imaging data, the actual healing situation is realistically reflected, addressing the problem of weak generalization ability of single-data-source models. Furthermore, by encapsulating the mapping process into a regression model using the growth rate mapping knowledge in the computationally intensive prior simulation model, accurate prediction from the patient's current state to the local callus growth rate is achieved, providing a highly efficient quantitative prediction tool for clinical practice.

[0104] Secondly, a weighted training loss function integrating distillation loss, soft-target loss, and hard-target loss is constructed to balance the granularity of model training, ensuring that the lightweight regression model monitors the true growth rate simultaneously during learning. This prevents bias or overfitting caused by excessive reliance on a single target during model training, improves the prediction accuracy of the final growth rate mapping model, and enhances the reliability of the model output.

[0105] Furthermore, by using the patient's specific online symptom information, a physical model of the fracture site is established and gridded in the digital space, realizing the personalized and precise initialization of the stage benchmark callus model. This provides a digital model that strictly matches the patient's actual anatomical structure, fracture type, and initial condition for subsequent growth simulation, ensuring that the entire monitoring process has highly individualized characteristics and improving the accuracy of prediction.

[0106] Simultaneously, by traversing grid cells to predict growth rates, integrating over time, and updating the geometric model, efficient numerical simulation of the callus growth process based on dynamic rates is achieved. The prediction of callus growth rate is transformed into an intuitive prediction of the evolution of the macroscopic morphology of the callus over time. The complete dynamic process of callus growth from initial growth to later remodeling is simulated, enabling the prediction of callus morphology at different time points and providing sufficient decision-making basis for more accurate rehabilitation decisions.

[0107] Ultimately, by iteratively comparing virtual images generated from the stage-predicted callus model with actual follow-up images of the patient and calculating and analyzing the residuals, a closed-loop online dynamic monitoring system is constructed to objectively assess the actual healing trajectory. If healing abnormalities are detected, timely warnings can be issued and the spatial area of ​​the abnormality can be located, providing alert information to doctors and enabling intelligent management of the patient's healing status.

[0108] In summary, existing technologies are time-consuming and heavily reliant on expert experience, making them difficult to integrate into clinical workflows. This application utilizes the concept of knowledge distillation to transform the complex biophysical laws inherent in historical high-fidelity simulation cases into a lightweight growth rate mapping model, thus shortening computation time. Secondly, existing methods based on pure data-driven or pure mechanism simulation suffer from weak generalization ability. This application constructs a joint training dataset that integrates simulation soft objectives and clinical hard objectives. By designing a multi-task weighted loss function to train the regression model, the model learns both simulation logic and clinical patterns simultaneously, significantly improving the model's generalization ability and the reliability of prediction results. Finally, a dynamic closed-loop intelligent monitoring and early warning system is constructed to achieve dynamic management from static assessment to callus morphology prediction.

[0109] Example 2, as Figure 2 As shown, based on the same inventive concept as the fracture healing process monitoring method based on callus growth simulation provided in Embodiment 1, this embodiment of the invention also provides a fracture healing process monitoring system based on callus growth simulation, including: The data acquisition and processing module 11 is used to collect callus growth example data for fracture healing through big data methods, and to construct a joint training dataset based on the callus growth example data, wherein the joint training dataset includes a first training dataset and a second training dataset. The regression model construction module 12 is used to initialize the regression model and train the regression model based on the joint training dataset and the knowledge distillation idea to obtain the growth rate mapping model. The baseline model construction module 13 is used to acquire online symptom information of the target patient and initialize the configuration stage baseline callus model based on the online symptom information. The prediction model construction module 14 is used to perform callus growth simulation based on dynamic growth rate according to the stage benchmark callus model and the growth rate mapping model, combined with a preset time step, to obtain the stage prediction callus model. Run the discrimination module 15 to predict the callus model based on the stage and iteratively monitor the fracture healing process of the target patient.

[0110] In one embodiment, the data acquisition and processing module 11 is used for: The data on callus growth examples includes simulation case data and clinical imaging data.

[0111] In one embodiment, the data acquisition and processing module 11 is further configured to: Based on the simulation case data, a first training dataset is constructed by combining the corresponding prior simulation model and clinical validation image data, wherein the first training dataset includes soft targets; Based on clinical imaging data, bimodal analysis is performed using the prior simulation model to construct a second training dataset, which includes soft targets and hard targets. The first training dataset and the second training dataset are merged to output the joint training dataset.

[0112] The first training dataset is constructed based on the simulation case data, combined with the corresponding prior simulation model and clinical validation image data, and includes: Based on the simulation case data, extract the first sample state features corresponding to each simulation case; By combining the simulation case data with the prior simulation model corresponding to the simulation case data, the virtual biofield data before the output layer is matched and extracted, and the output is the first soft target. The first sample state features are merged with the first soft target to obtain the first training dataset.

[0113] Specifically, based on clinical imaging data, a bimodal analysis is performed using the prior simulation model to construct a second training dataset. This second training dataset includes soft and hard targets, including: The clinical image data is segmented and analyzed using machine vision methods to obtain the first callus state characteristics. Based on the first callus state characteristics, a neighborhood temporal change analysis is performed to calculate and obtain the true generation rate as the second hard target; The second sample state features corresponding to the clinical imaging data are obtained, and the second sample state features are input into the prior simulation model to extract the virtual biofield data before the output layer as the second soft target. The second training dataset is constructed by combining the second sample state features, the second hard target, and the second soft target.

[0114] In one embodiment, the regression model building module 12 is used for: Construct a training loss function, which includes: Distillation loss term, wherein the distillation loss term is the loss of the regression model on the soft target of the first training dataset; The soft target loss term is the loss of the regression model on soft targets in the second training dataset. Hard target loss term, wherein the hard target loss term is the loss of the regression model on hard targets in the second training dataset; The distillation loss term, the soft target loss term, and the hard target loss term are weighted and fused together by coefficients that sum to 1 to form the training loss function.

[0115] In one embodiment, the baseline model building module 13 is used for: Based on the online medical information, a physical model of the fracture site is established in the digital space; Based on the shape characteristics of the physical model, the physical model is divided into spatial grids to obtain the stage reference callus model.

[0116] In one embodiment, the prediction model building module 14 is used for: By combining the growth rate mapping model with the online symptom information of the target patient, the growth rate is predicted by traversing each spatial grid region of the stage benchmark callus model to obtain the growth rate of multiple units. The growth rates of multiple units are traversed, and time integration is performed along a preset growth direction to obtain the newly added bone tissue information of each spatial grid region. The newly added bone tissue information is aggregated, and the stage baseline callus model is updated using a geometry engine method to obtain the stage predicted callus model.

[0117] In one embodiment, the discrimination module 15 is used to: Each time the stage prediction callus model is obtained in the iteration, the comparative image data of the target patient is updated based on the preset time step. Based on the image acquisition parameters of the predicted callus model and the comparative image data, virtual images are generated to obtain reference image data. The reference image data is compared with the comparison image data and the residual is calculated. If the residual is greater than or equal to a preset residual constraint, a healing anomaly is reported, and the distribution of the residual is reported simultaneously.

[0118] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the data acquisition and processing module 11 first constructs a joint training dataset by integrating simulation case data and clinical imaging data. A lightweight growth rate mapping model is then trained using the knowledge distillation concept to achieve synergistic effects. By fusing the deep biophysical laws revealed by high-fidelity biomechanical simulation with real clinical imaging data, the actual healing situation is realistically reflected, addressing the problem of weak generalization ability of single-data-source models. Furthermore, by encapsulating the growth rate mapping knowledge in the computationally intensive prior simulation model into a regression model, accurate prediction from the patient's current state to the local callus growth rate is achieved, providing an efficient quantitative prediction tool for clinical practice.

[0119] Secondly, through the regression model construction module 12, a weighted training loss function integrating distillation loss, soft target loss, and hard target loss is constructed to balance the granularity of model training, ensuring that the lightweight regression model monitors the true growth rate simultaneously during learning. This prevents bias or overfitting caused by excessive reliance on a single target during model training, improves the prediction accuracy of the final growth rate mapping model, and enhances the reliability of the model output.

[0120] Furthermore, through the baseline model construction module 13, a physical model of the fracture site is established and gridded in the digital space based on the patient's specific online symptom information. This enables the personalized and precise initialization of the stage baseline callus model, providing a digital model that strictly matches the patient's real anatomical structure, fracture type, and initial condition for subsequent growth simulation. This ensures that the entire monitoring process has highly individualized characteristics and improves the accuracy of prediction.

[0121] Simultaneously, through the prediction model construction module 14, the growth rate is predicted by traversing the grid cells, time integration is performed, and the geometric model is updated, achieving efficient numerical simulation of the callus growth process based on dynamic rate. The prediction of callus growth rate is transformed into an intuitive prediction of the evolution of the macroscopic morphology of the callus over time. The complete dynamic process of callus growth from initial growth to later remodeling is simulated, enabling the prediction of callus morphology at different time points and providing sufficient decision-making basis for more accurate rehabilitation decisions.

[0122] Finally, by running the discrimination module 15, the virtual images generated by the stage prediction callus model are iteratively compared with the patient's actual follow-up images. The residuals are calculated and analyzed to construct a closed-loop online dynamic monitoring system, objectively evaluating the actual healing trajectory. If healing is abnormal, timely warnings can be issued and the spatial area of ​​the abnormality can be located, providing alarm information to doctors and realizing intelligent management of the patient's healing status.

[0123] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0124] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0125] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for monitoring fracture healing process based on callus growth simulation, characterized in that, include: Using big data methods, we collect case data on callus growth for fracture healing and construct a joint training dataset based on the case data. The joint training dataset includes a first training dataset and a second training dataset. Initialize the regression model, and train the regression model based on the joint training dataset and the knowledge distillation concept to obtain the growth rate mapping model; Obtain online symptom information of the target patient, and initialize the configuration stage baseline callus model based on the online symptom information; Based on the stage baseline callus model and the growth rate mapping model, and combined with a preset time step, callus growth simulation based on dynamic growth rate is performed to obtain the stage prediction callus model. Based on the stage-predicted callus model, the fracture healing process of the target patient is monitored iteratively.

2. The method for monitoring fracture healing process based on callus growth simulation as described in claim 1, characterized in that, The data on callus growth examples includes simulation case data and clinical imaging data.

3. The method for monitoring fracture healing process based on callus growth simulation as described in claim 2, characterized in that, A joint training dataset was constructed based on the aforementioned callus growth paradigm data, including: Based on the simulation case data, a first training dataset is constructed by combining the corresponding prior simulation model and clinical validation image data, wherein the first training dataset includes soft targets; Based on clinical imaging data, bimodal analysis is performed using the prior simulation model to construct a second training dataset, which includes soft targets and hard targets. The first training dataset and the second training dataset are merged to output the joint training dataset.

4. The method for monitoring fracture healing process based on callus growth simulation as described in claim 3, characterized in that, Based on the simulation case data, a first training dataset is constructed by combining the corresponding prior simulation model and clinical validation image data, including: Based on the simulation case data, extract the first sample state features corresponding to each simulation case; By combining the simulation case data with the prior simulation model corresponding to the simulation case data, the virtual biofield data before the output layer is matched and extracted, and the output is the first soft target. The first sample state features are merged with the first soft target to obtain the first training dataset.

5. The method for monitoring fracture healing process based on callus growth simulation as described in claim 3, characterized in that, Based on clinical imaging data, a bimodal analysis is performed using the prior simulation model to construct a second training dataset, which includes soft and hard targets, including: The clinical image data is segmented and analyzed using machine vision methods to obtain the first callus state characteristics. Based on the first callus state characteristics, a neighborhood temporal change analysis is performed to calculate and obtain the true generation rate as the second hard target; The second sample state features corresponding to the clinical imaging data are obtained, and the second sample state features are input into the prior simulation model to extract the virtual biofield data before the output layer as the second soft target. The second training dataset is constructed by combining the second sample state features, the second hard target, and the second soft target.

6. The method for monitoring fracture healing process based on callus growth simulation as described in claim 1, characterized in that, Initialize the regression model, and train the regression model using the joint training dataset and the concept of knowledge distillation to obtain the growth rate mapping model. Prior to this, the process includes: Construct a training loss function, which includes: Distillation loss term, wherein the distillation loss term is the loss of the regression model on the soft target of the first training dataset; The soft target loss term is the loss of the regression model on soft targets in the second training dataset. Hard target loss term, wherein the hard target loss term is the loss of the regression model on hard targets in the second training dataset; The distillation loss term, the soft target loss term, and the hard target loss term are weighted and fused together by coefficients that sum to 1 to form the training loss function.

7. The method for monitoring fracture healing process based on callus growth simulation as described in claim 1, characterized in that, Acquire online symptom information of the target patient, and initialize a baseline callus model for the configuration phase based on the online symptom information, including: Based on the online medical information, a physical model of the fracture site is established in the digital space; Based on the shape characteristics of the physical model, the physical model is divided into spatial grids to obtain the stage reference callus model.

8. The method for monitoring fracture healing process based on callus growth simulation as described in claim 1, characterized in that, Based on the stage-based baseline callus model and the growth rate mapping model, and combined with a preset time step, callus growth simulation based on dynamic growth rate is performed to obtain a stage-predicted callus model, including: By combining the growth rate mapping model with the online symptom information of the target patient, the growth rate is predicted by traversing each spatial grid region of the stage benchmark callus model to obtain the growth rate of multiple units. The growth rates of multiple units are traversed, and time integration is performed along a preset growth direction to obtain the newly added bone tissue information of each spatial grid region. The newly added bone tissue information is aggregated, and the stage baseline callus model is updated using a geometry engine method to obtain the stage predicted callus model.

9. The method for monitoring fracture healing process based on callus growth simulation as described in claim 1, characterized in that, Based on the stage-based callus prediction model, the fracture healing process of the target patient is iteratively monitored, including: Each time the stage prediction callus model is obtained in the iteration, the comparative image data of the target patient is updated based on the preset time step. Based on the image acquisition parameters of the predicted callus model and the comparative image data, virtual images are generated to obtain reference image data. The reference image data is compared with the comparison image data and the residual is calculated. If the residual is greater than or equal to a preset residual constraint, a healing anomaly is reported, and the distribution of the residual is reported simultaneously.

10. A fracture healing process monitoring system based on callus growth simulation, characterized in that, For implementing the fracture healing process monitoring method based on callus growth simulation according to any one of claims 1-9, the system comprises: The data acquisition and processing module is used to collect callus growth example data for fracture healing through big data methods, and to construct a joint training dataset based on the callus growth example data, wherein the joint training dataset includes a first training dataset and a second training dataset. The regression model building module is used to initialize the regression model and train the regression model based on the joint training dataset, combined with the knowledge distillation idea, to obtain the growth rate mapping model. The baseline model construction module is used to acquire online symptom information of the target patient and initialize the configuration stage baseline callus model based on the online symptom information. The prediction model construction module is used to simulate callus growth based on dynamic growth rate by combining the stage baseline callus model and the growth rate mapping model with a preset time step, and obtain the stage prediction callus model. The discrimination module is used to predict the callus model based on the stage and iteratively monitor the fracture healing process of the target patient.

Images