Fracture treatment scheme and operation scheme simulation system based on big data analysis

Abstract

The invention relates to the technical field of big data processing, in particular to a fracture treatment scheme and operation scheme simulation system based on big data analysis, which comprises a data preprocessing module, an adversarial network module and a model training and optimizing module. And carrying out transfer learning on the fracture treatment decision knowledge learned by the plurality of heterogeneous teacher models to a student model. According to the application, through a strategy of fusing confrontation generative network (GAN) sample expansion and heterogeneous teacher model knowledge distillation transfer learning, the contradiction between single hospital comminuted fracture case quantity scarcity (sample sparsity) and medical resource and capability high personalization (resource difference) is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data processing, in particular to a fracture treatment scheme and surgical scheme simulation system based on big data analysis. BACKGROUND

[0002] The content of this part only provides background information related to the present application, which may not constitute prior art.

[0003] In fracture treatment, comminuted fractures are the most complex and uncertain type of injury due to their severe bone fragmentation, unstable fracture ends, and difficulty in reduction and fixation. The selection of treatment options is extremely challenging and requires consideration of various personalized factors such as fracture location, comminution degree, displacement direction, soft tissue injury, patient age and physical condition, and underlying diseases. In clinical practice, the best treatment option with better prognosis (e.g., better functional recovery and fewer complications) needs to be selected by weighing the risks and benefits among various complex combinations of surgical approaches (e.g., open reduction and internal fixation, minimally invasive percutaneous fixation), internal fixation devices (e.g., bone plates, intramedullary nails, external fixation frames), and auxiliary technologies (e.g., bone grafting).

[0004] To assist doctors in decision-making, surgical scheme simulation systems based on big data analysis are currently being used. These systems typically establish complex statistical models or machine learning models (such as deep learning, decision trees, random forests, etc.) to deeply mine and analyze vast amounts of historical case data. Based on the input of new patient characteristics (such as fracture type and patient indicators), the system retrieves or learns from the database to generate the treatment scheme with the best prognosis and can simulate and simulate the final treatment effect (such as bone healing progress, functional score prediction, and complication probability estimation) that the scheme may bring. Finally, the simulation results are presented to doctors and patients in an intuitive way (such as three-dimensional visualization simulation, data charts, and risk-benefit analysis reports) as an important reference for decision-making.

[0005] Although existing big data driven surgical plan simulation systems can mine the general rules of historical data and recommend treatment plans with good prognosis in a "macro" sense, the optimal plan recommended by the system is usually "population-based optimal" based on the overall statistical results of its training data set, rather than "localized optimal" for the specific hospital's own medical resource allocation and ability characteristics. Different hospitals have significant differences in the accuracy of surgical equipment models, the types of inventory instruments, the availability of consumables, and most importantly, the proficiency and preference of the surgeon team for specific technologies (such as minimally invasive technology and specific instrument operation). A theoretically optimal plan may not be effectively implemented due to the lack of corresponding resources or the lack of relevant technical experience of the doctors in the target hospital, and even increase the risk (resource and ability mismatch risk). However, due to the extremely limited number of comminuted fracture case data samples that meet the high-quality modeling requirements that a single hospital can accumulate (sample sparsity), the size of its own data is usually insufficient to support the training of a special big data model that can accurately capture its unique resource and ability constraints and recommend a localized treatment plan that is most suitable for its conditions. The contradiction between "sample sparsity" and "resource difference" leads to the fact that the actual implementation feasibility and final effect prediction accuracy of the simulation results of the recommended plan by the existing simulation system are very low when applied to different hospitals. SUMMARY

[0006] The summary part of the present application is used to introduce the concepts in a brief form, which will be described in detail in the specific embodiment part. The summary part of the present application is not intended to identify key features or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of the present application propose a big data analysis based fracture treatment plan surgical plan simulation system to solve the technical problems mentioned in the background part.

[0008] As a first aspect of the present application, some embodiments of the present application provide a big data analysis based fracture treatment plan surgical plan simulation system, comprising: a data preprocessing module configured to collect a hospital local case data set including pre-examination information, intermediate treatment information and prognosis information; an adversarial network module receiving the hospital local case data set and training and sample expanding the hospital local case data set using at least one adversarial generative neural network model to generate a simulated case data set; The simulated case data set matches the local medical resource and ability conditions represented by the hospital local case data set in statistical distribution; The model training and optimization module is configured to: Pre-training a plurality of heterogeneous teacher models on a general large-scale fracture treatment dataset, and transferring the fracture treatment decision knowledge learned by the plurality of heterogeneous teacher models to a student model for transfer learning, The student model is retrained using the simulated case dataset generated by the adversarial network module during transfer learning; The simulation module integrates the student model as a final localized simulation model, which is configured to: Input the pre-examination information of the target patient; and based on the input pre-examination information, simulate and output the intermediate treatment information of the potential fracture surgery plan for the target patient and the corresponding prognosis information prediction results.

[0009] The present application overcomes the contradiction between the scarcity of comminuted fracture cases (sample sparsity) and the high individualization of medical resources and capabilities (resource difference) by fusing the strategy of generative adversarial network (GAN) sample expansion and heterogeneous teacher model knowledge distillation transfer learning. The core effect is that the simulated case dataset generated by the system not only significantly expands the sample size of localized modeling, but more importantly, accurately matches the unique constraints of the target hospital (such as equipment models, instrument inventory, and doctor-specific technical proficiency); combined with the rich fracture treatment knowledge and pattern recognition ability transferred from the general large-scale data, the final localized simulation student model not only recommends treatment plans that are feasible under the existing resources and capabilities of the target hospital, but also highly reliably simulates and predicts the execution process (intermediate treatment information) of these plans and their actual prognosis effects (such as bone healing progress, functional score, and complication probability), significantly improving the adaptability, execution feasibility, and prognosis prediction accuracy of the recommended plans in the actual clinical environment, thereby bridging the gap between macro-optimal plans and localized optimal execution effects.

[0010] Further, the data preprocessing module is configured to: Using a structural feature extraction network, the non-numerical or unstructured data in the collected pre-examination information, intermediate treatment information, and prognosis information are converted into structured numerical feature vectors according to a pre-set standard coding table; Using an image feature extraction network, the medical image data in the collected pre-examination information, intermediate treatment information, and prognosis information are processed to extract structured visual feature vectors related to fracture morphology, comminution degree, and surrounding soft tissue condition; Using a feature organization network, the pre-examination information, intermediate treatment information, and prognosis information are respectively converted into structured numerical feature vectors and structured visual feature vectors to obtain digital examination information, digital treatment information, and digital prognosis information.

[0011] The data preprocessing module realizes the standardization, deep structuring conversion and efficient fusion of multi-source heterogeneous medical information (unstructured text / numerical, multi-modal image) through deep feature engineering, the synergistic effect of structure feature extraction network, image feature extraction network and feature organization network. The original chaotic and difficult to directly model non-numerical / unstructured data (such as diagnosis description, operation selection) is accurately converted into structured numerical feature vector according to the standard code, and at the same time, the image feature extraction network specially designed for fracture scene accurately analyzes the quantitative visual feature vector of fracture morphology (such as fragment number, displacement angle, soft tissue injury grade, etc.) from the original medical image. The feature organization network organizes these key information of different dimensions into structured, high-quality digital examination information, treatment information and prognosis information.

[0012] Further, the adversarial network module comprises: a noise source generation module, which determines a noise source generation range based on the medical conditions of the hospital, and generates noise signals within the noise source generation range; a generator, which generates simulation samples based on the noise signals, the simulation samples including simulated digital examination information, digital treatment information and digital prognosis information; a discriminator, which forms an adversarial network with the generator, and discriminates the authenticity of the input simulation samples and real samples; a simulation data generator, which internally stores the trained generator, generates new simulation samples according to new noise sequences, and generates simulation case data sets using the new simulation samples.

[0013] The adversarial network module effectively generates high-quality simulation case data sets that are highly consistent with the unique medical resource capability boundaries (such as specific device types, limited instrument inventory, and physician technical proficiency levels) of the target hospital in statistical distribution through its unique "constraint noise source generation" and bidirectional adversarial training mechanism. The noise source generation module determines the generation range of the noise according to the specific conditions of the hospital (such as the list of equipment models, the range of operable operation, and the experience level of the physician team), and constrains the generation space of the simulation cases from the source, ensuring that the new samples are aligned with the local reality in the "resource supply capability" dimension. Under this constraint, the generator and the discriminator dynamically optimize through adversarial game, and finally drive the generator to output "pseudo-true" case samples that contain reasonable medical logic (such as the matching of fracture type and treatment plan) and naturally integrate local resource limitation features (such as preferential use of inventory instruments and avoidance of high-risk minimally invasive operations). This mechanism significantly overcomes the sample scarcity problem and accurately captures key local factors that affect the feasibility and execution effect of the scheme, providing a large amount of training data for the subsequent training of the local decision-making model (student model) that is sufficient in quantity, reliable in quality and truly reflects "what can be executed in this hospital and what effect has been achieved in this hospital".

[0014] Further, the noise signal is generated in the following manner: S1: load the local hospital's medical resources, generate an original encoding table of the local medical resources, map the original encoding table to a standard encoding table, and determine the longitudinal dimension of the noise signal according to the dimension of the original encoding table; Load the maximum dimension of the real case feature vector of the local hospital as the transverse dimension of the noise signal; Wherein, the longitudinal dimension is used to constrain the numerical fluctuation range of the noise signal, and the transverse dimension is used to constrain the length of the noise signal; S2: Predefine an entropy source seed signal source and collect random signals generated by the entropy source seed signal source; S3: Map each element in the random signal to the corresponding position element of the noise signal based on the proportion information, and generate the noise signal.

[0015] The noise signal generation mechanism limits the numerical range of the noise based on the "longitudinal dimension" of the actual medical resources (mapped to the standard encoding) of the local hospital, and defines the structural length of the noise based on the "transverse dimension" of the maximum feature complexity of the real cases, which fundamentally ensures that the generated noise signal strictly obeys the ability boundary of the target hospital (such as available instrument types, equipment parameter limits, and operation range) and can carry all the key features of the local cases; At the same time, using a high-entropy true random source seed and through a proportional mapping rule to generate an initial random signal, under the premise of meeting the strict two-dimensional constraints, the diversity and unpredictability of the noise signal are effectively guaranteed, thereby providing a high-quality "seed" for the adversarial generation network, which can accurately reflect the local resource constraints and drive the generation of rich and reasonable medical logic simulation cases. The cornerstone of the high-fidelity simulation case data set with statistical distribution highly consistent with the unique conditions of the target hospital.

[0016] Further, the noise signal is , ; represents the jth element in , j represents the index of the element in the noise signal; ; Wherein, represents the value of the jth element of the random signal, represents the maximum value of the random signal, represents the maximum value of the noise signal at the jth position, represents the floor function.

[0017] The accurate quantization calculation method of the noise signal element realizes the deterministic and discrete conversion of the high-entropy random source signal to the noise element strictly conforming to the local resource and capability boundary conditions of the target hospital by combining the normalized proportion control with the local resource upper limit value. The core effect is that it effectively applies the mathematical constraint "maximum value" of hospital resources and capabilities (such as the upper limit value of a specific instrument model, the concurrent operation complexity threshold allowed in the operating room, or the upper limit of the proficiency level of a certain technical skill of a doctor team) as a hard ceiling for noise element generation, and ensures that each noise element value generated is an integer and falls within the actual executable range of the target hospital in that feature dimension, thereby forcibly embedding the fundamental constraints of local resources at the most basic unit level of the noise space, and laying a mathematical foundation for the subsequent generator to output simulation cases that absolutely conform to the hospital's real operation conditions (such as only using the equipment models that exist in stock, and only handling surgical complexity within the doctor's experience threshold).

[0018] Further, the generator comprises: a generator input layer for receiving a noise vector; a feature conversion layer for extracting hidden features from the noise vector based on a fully connected layer and a residual network; a generator output layer for outputting a simulation sample according to the input hidden features.

[0019] The generator network accurately receives the noise signal under the local constraint through the input layer, and deeply analyzes the local resource capability boundary information and potential medical logic association contained in the noise through the feature conversion layer composed of the fully connected and residual networks. The residual structure in the feature conversion layer effectively overcomes the gradient vanishing problem of deep network, stably captures the complex and nonlinear deep structure and association in the fracture diagnosis and treatment process (such as the implicit relationship between specific image features and complication risk, and the differential influence of different surgical options on the prognosis model of patients with specific constitution), and ensures that the generated samples have high medical logic authenticity and pattern diversity while complying with strict resource constraints.

[0020] Further, the discriminator comprises: a feature pyramid backbone network for inputting a simulation sample or a real sample to generate a case feature; a true-false discrimination head based on a Sigmoid function to regress the case feature to the probability that the sample is true.

[0021] The discriminator resolves simulated and real case samples (digitalized examination, treatment, prognosis information) through a feature pyramid backbone network, uses its multi-level and multi-scale perception ability to accurately capture the complex and cross-dimensional medical logic associations (such as the matching degree of fracture comminution morphology and internal fixation device selection, the implicit mapping of intraoperative specific operation details and postoperative complication risk) and the statistical patterns of key local resource constraints (such as the usage frequency of preferred device models, the traces of specific team operation habits) contained in the case data, and the authenticity judgment head of the discriminator efficiently focuses these high-dimensional and complex fusion features into the final sample authenticity probability judgment.

[0022] Further, the loss function of the generator and the discriminator is: ; ; ; represents the loss of the discriminator, E represents the mathematical expectation operator, represents a real sample x sampled from a real data distribution, represents a penalty sample generated by interpolation, represents a mixed random number, x represents a real sample, represents a generated sample generated by the generator; represents a noise vector z sampled from a noise distribution, represents the score of the discriminator on the real sample, represents the score of the discriminator on the generated sample ; represents the weight coefficient of the gradient penalty term, represents a sample sampled from a penalty distribution, represents the gradient output by the discriminator with respect to the input ; represents the feature matching loss, l represents the level index of the feature pyramid, represents the feature map output by the discriminator on the real sample x at the lth layer, represents the feature map output by the discriminator on the generated sample at the lth layer, represents the square of the L2 norm.

[0023] Further, the plurality of heterogeneous teacher models includes an intermediate treatment information generation model and a prognosis information generation model; The intermediate treatment information generation model is trained based on a first sample set prepared in advance; The prognosis information generation model is trained based on a second sample set prepared in advance; The intermediate treatment information generation model generates optimal intermediate treatment information based on the input preliminary examination information. The prognostic information generation model generates prognostic information based on the input preliminary examination information and intermediate treatment information.

[0024] This heterogeneous teacher model design decomposes treatment plan decisions into a model focused on generating optimal "intermediate treatment information" and a model focused on predicting "prognostic information." Both models are deeply pre-trained using large-scale professional sample sets, achieving a strong cohesion and abstraction of highly specialized knowledge inherent in the core decision-making chain of fracture treatment ("determining the treatment plan based on the injury" and "evaluating the effectiveness of the plan"). This division of labor enables the two teacher models to achieve expert-level accuracy and depth of understanding in optimal matching reasoning of treatment plans (inputting preliminary examinations and outputting treatment information) and accurate simulation of treatment plan effects (inputting preliminary examinations and treatment information and outputting prognostic information). This lays a solid foundation for the subsequent complete, independent, and faithful injection of these universal, big data-validated deep decision-making knowledge—especially the logic of generating high-value treatment plans and complex prognostic causal mapping relationships—into the local student model through transfer learning. This effectively improves the student model's ability to formulate feasible treatment plans with good prognostic prospects under local constraints and enhances the accuracy of its prognostic predictions.

[0025] Furthermore, the student model includes: The feature input layer is used to input digital inspection information and extract feature information from the digital inspection information; The treatment information generation layer outputs corresponding digital treatment information based on the input feature information. The prognostic information generation layer generates digital prognostic information based on the digital treatment information and digital examination information output by the treatment information generation layer.

[0026] Furthermore, the intermediate treatment information generation model is signal-connected to the treatment information generation layer, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is used as the first loss term. The prognostic information generation model is connected to the prognostic information generation layer signal, and the output of the prognostic information generation model and the error between the prognostic information generation layer are used as the second loss term. The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer; The second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer.

[0027] The student model accurately inherits the chain decision-making ability (treatment scheme generation → effect prediction) of the heterogeneous teacher model through modular hierarchical design (treatment information generation layer, prognosis information generation layer), and innovatively adopts a double-channel differentiated knowledge transfer mechanism: the output of the treatment information teacher model is used as the benchmark to calculate the first loss term to guide the treatment layer parameter optimization, and the output of the prognosis information teacher model is used as the benchmark to calculate the second loss term to guide the prognosis layer parameter update, to strengthen the localization learning (the total loss covers the local case generated by the generative adversarial network) while realizing two core values: 1) the treatment decision layer accurately absorbs the generation logic of the universal optimal scheme (such as the deep mapping of fracture characteristics and instrument selection), 2) the prognosis prediction layer fine internalizes the universal law of the scheme-effect causal chain (such as the relevance of a specific instrument and bone healing rate), and finally forms a decision-making closed loop with a global big data perspective (teacher knowledge) and local resource adaptability (simulation data training), so that the system can highly reliably simulate and predict the actual effect (prognosis information) of the treatment scheme under the conditions of the target hospital, and completely break through the clinical decision-making chain from “feasible scheme under resource constraints” to “accurate prognosis evaluation”. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A flowchart of a fracture treatment scheme surgery scheme simulation system based on big data analysis.

[0029] Figure 2 A training diagram of a heterogeneous teacher model and a student model. DETAILED DESCRIPTION

[0030] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely in conjunction with specific implementation manners. The same reference signs in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0031] Compared with the embodiments shown in the drawings, the feasible implementation solutions within the protection scope of the present application can have fewer components, other components not shown in the drawings, different components, differently arranged components or differently connected components, etc. In addition, two or more components in the drawings can be implemented in a single component, or a single component shown in the drawings can be implemented as a plurality of separate components.

[0032] Unless otherwise defined, technical terms or scientific terms used herein shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "first", "second", and similar terms do not imply any order, quantity, or importance, but are used to identify different components. Similarly, the terms "one", "another", and similar terms do not necessarily denote a quantity of only one. The terms "upper", "lower", and similar terms are used to describe relative positions, and when the absolute positions of the described objects are changed, the relative positions may also be changed accordingly.

[0033] Reference Figure 1 , Embodiment 1: A fracture treatment plan surgery plan simulation system based on big data analysis includes a data preprocessing module, an adversarial network module, a model training and optimization module, and a simulation module. The data preprocessing module, the adversarial network module, the model training and optimization module, and the simulation module are connected in sequence, and the data preprocessing module is signal connected with the simulation module.

[0034] The function of the data preprocessing module is to convert the original case data containing text, records, and images (such as CT) in the hospital into structured and numerical feature vectors, providing uniform and processable digital input for subsequent modules.

[0035] The data preprocessing module converts non-numerical examination information, treatment plans, and prognosis records into structured numerical feature vectors according to standard medical coding tables through a structural feature extraction network (such as an encoding converter); then uses a special image feature extraction network (such as a convolutional neural network) to analyze medical images and extract structured vectors describing key visual features of fracture morphology and fragmentation details; then, through a feature organization network, the two parts of information are integrated to form digital examination information, digital treatment information, and digital prognosis information that represent the patient's condition, achieving standardized and structured expression of multi-source heterogeneous medical data.

[0036] Specifically, the data preprocessing module is configured to collect a hospital local case data set including pre-examination information, intermediate treatment information, and prognosis information; The data preprocessing module is configured to: Use a structural feature extraction network to convert non-numerical or unstructured data in the collected pre-examination information, intermediate treatment information, and prognosis information into structured numerical feature vectors according to a pre-set standard coding table; Use an image feature extraction network to process medical image data in the collected pre-examination information, intermediate treatment information, and prognosis information, and extract structured visual feature vectors related to fracture morphology, fragmentation degree, and surrounding soft tissue condition; The pre-examination information, intermediate treatment information and prognosis information are respectively converted into structured numerical feature vectors and structured visual feature vectors by using a feature organization network, to obtain digital examination information, digital treatment information and digital prognosis information.

[0037] Specifically, the pre-examination information, intermediate treatment information and prognosis information belong to common information in a case. If there is incomplete information, data filling needs to be performed.

[0038] The pre-examination information is related information collected before the surgical treatment of the patient, such as name, age, gender, medical history, CT image, fracture position picture and the like.

[0039] The intermediate treatment information is the surgical plan of the patient, such as steel plate screw fixation, intramedullary nail, closed reduction external fixation, screw / tension screw fixation, combined application.

[0040] The prognosis information is the prognosis condition of the patient, which is generally the recovery condition from postoperative to discharge, usually including wound healing recorded by nurses, whether there is complication or not and the like.

[0041] The pre-examination information, intermediate treatment information and prognosis information above all belong to text information + video, image information. These information cannot be directly used for subsequent processing. Therefore, the text information is converted into code by using a structure feature extraction network. The video, image information is converted into structured visual feature vectors by using an image feature extraction network. Then the pre-examination information, intermediate treatment information and prognosis information can be converted into digital examination information, digital treatment information and digital prognosis information.

[0042] Specifically, the digital examination information, digital treatment information and digital prognosis information are essentially digital processing of information.

[0043] Specifically, the digital examination information is essentially the text case of the patient + image examination data. The text case will be converted into a corresponding code table according to a pre-set coding rule. For example, 0011 represents that the patient is male, age 22 and the like. Only need to arrange the data according to the rule, and then the text information can be converted into code according to the coding dictionary.

[0044] The image examination data is various ultrasonic examination, CT examination and X-ray examination data. This part of examination data is converted into specific examination results by using an image feature extraction network. For example, the ultrasonic data shows that a certain position is a first degree comminuted fracture (comminuted intensity) and the like. After being converted into specific examination results, the code data is converted, so as to obtain the digital examination information.

[0045] The digital treatment information is to encode the treatment process. Generally, the fracture surgery plan can be roughly divided into: Operation time (distance from fracture time); Anesthesia method (general anesthesia, local anesthesia); Operation method (open surgery or external fixation); Bone fragment processing method (bone fragment removal or bone fragment retention); Fracture fixation method (intramedullary nail fixation, steel plate fixation, steel nail fixation); Postoperative fixation range (whether it needs to cover the entire limb).

[0046] The above parts can be flexibly adjusted according to actual needs. For each part, the corresponding data items can be added or reduced, and converted into corresponding codes to obtain digital prognosis information.

[0047] The digital prognosis information is to describe the prognosis situation in stages, for example: Postoperative day 1: wake up, no fever; Postoperative day 5: wound is not inflamed; Postoperative day 10: wound is not inflamed, etc.

[0048] For each prognosis situation, appropriate labels can be set according to actual needs, such as good, poor, and poor. Further converting these labels into codes can obtain digital prognosis information.

[0049] The adversarial network module receives the hospital local case data set, and uses at least one adversarial generative neural network model to train and sample expand the hospital local case data set to generate a simulated case data set; The simulated case data set is matched with the local medical resource and ability condition represented by the hospital local case data set in statistical distribution.

[0050] The number of cases in each hospital is relatively limited. Using limited cases to train neural networks generally has poor results. Therefore, the adversarial network module is used to expand the data. Specifically: The adversarial network module includes: A noise source generation module determines a noise source generation range based on the medical conditions of the hospital, and generates a noise signal within the noise source generation range; A generator generates simulated samples based on the noise signal, the simulated samples including simulated digital examination information, digital treatment information, and digital prognosis information; A discriminator forms an adversarial network with the generator to determine the authenticity of the input simulated samples and real samples; The simulation data generator is internally provided with a trained generator, and generates new simulation samples according to new noise sequences, and generates a simulation case data set by using the new simulation samples.

[0051] The noise signal is generated in the following manner: S1: loading medical resources of a local hospital, generating an original coding table of the local medical resources, mapping the original coding table to a standard coding table, and determining a longitudinal dimension of the noise signal according to a dimension of the original coding table; loading a maximum dimension of a real case feature vector of the local hospital as a transverse dimension of the noise signal; The longitudinal dimension is used to constrain the numerical fluctuation range of the noise signal, and the transverse dimension is used to constrain the length of the noise signal. S2: defining an entropy source seed signal source in advance, and collecting a random signal generated by the entropy source seed signal source; S3: mapping each element in the random signal to an element at a corresponding position of the noise signal based on scale information, to generate the noise signal.

[0052] The noise signal is , ; denotes the jth element in , and j represents an index of an element in the noise signal; ; wherein denotes a value of the jth element of the random signal, denotes a maximum value of the random signal, denotes a maximum value of the noise signal at the jth position, denotes a floor function.

[0053] The generator comprises: a generator input layer configured to receive a noise vector; a feature conversion layer configured to extract hidden features from the noise vector based on a fully connected layer and a residual network; a generator output layer configured to output a simulation sample according to the input hidden features.

[0054] The discriminator comprises: a feature pyramid backbone network configured to input a simulation sample or a real sample to generate a case feature; a true-false discrimination head configured to regress the case feature to a probability that the sample is true based on a Sigmoid function.

[0055] The loss function of the generator and the discriminator is: ; ; ; denotes the discriminator loss, E denotes the mathematical expectation operator, denotes a real sample x sampled from a real data distribution, denotes a penalty sample generated by interpolation, denotes a mixed random number, x denotes a real sample, denotes a generated sample generated by the generator; denotes a noise vector z sampled from a noise distribution, denotes the score of the discriminator on the real sample, denotes the score of the discriminator on the generated sample , denotes a weight coefficient of the gradient penalty term, denotes a sample sampled from a penalty distribution, denotes the gradient output by the discriminator with respect to the input ; denotes a feature matching loss, l denotes a hierarchical index of a feature pyramid, denotes a feature map output by the discriminator on the real sample x at the lth layer, denotes a feature map output by the discriminator on the generated sample at the lth layer, denotes the square of the L2 norm.

[0056] In this way, after jointly training the generator and the discriminator, an accurate generator can be obtained. By inputting random noise into the generator, a required number of simulated samples can be obtained, and after obtaining a sufficient number of simulated samples, a student model can be trained.

[0057] The model training and optimization module is configured to: pre-train a plurality of heterogeneous teacher models on a general large-scale fracture treatment dataset, and perform transfer learning of fracture treatment decision knowledge learned by the plurality of heterogeneous teacher models to a student model, using the simulated case data set generated by the adversarial network module to perform secondary training on the student model during the transfer learning.

[0058] The plurality of heterogeneous teacher models includes an intermediate treatment information generation model and a prognosis information generation model; The intermediate treatment information generation model is trained based on a first sample set prepared in advance; The prognosis information generation model is trained based on a second sample set prepared in advance; The intermediate treatment information generation model generates optimal intermediate treatment information according to input early examination information; The prognosis information generation model generates prognosis information according to input pre-examination information and intermediate treatment information.

[0059] The pre-examination information, the intermediate treatment information, and the prognosis information are digital examination information, digital treatment information, and digital prognosis information described above. The intermediate treatment information generation model is a pre-trained decision model based on a deep residual neural network architecture, which inputs standardized and structured digital examination information of a patient and outputs digital treatment information adapted to the input injury condition. The model learns high-dimensional mapping logic between fracture types, patient physiological parameters, and optimal treatment plans by supervised training on a first sample set composed of global multi-center high-quality treatment cases, forming a treatment strategy generator with general medical decision-making ability.

[0060] The prognosis information generation model is a spatio-temporal prediction neural network using a multi-modal feature fusion mechanism, which inputs digital examination information (initial injury condition of the patient) and digital treatment information (executed plan), and outputs structured digital prognosis information. The model training relies on a second sample set with complete follow-up records (covering 6-24 month prognosis trajectories under different treatment plans), and realizes fine simulation prediction of the execution effect of a specific treatment plan by modeling the dynamic causal chain of "injury condition-treatment-outcome".

[0061] The intermediate treatment information generation model and the prognosis information generation model are large neural network models, and their specific structures are prior art. Given the known model input and output, a deep residual neural network and a spatio-temporal prediction neural network can be used to respectively construct the required intermediate treatment information generation model and prognosis information generation model. In practice, it is necessary to construct an intermediate treatment information generation model and a prognosis information generation model with complex hierarchical structure and high depth to increase the data processing and learning ability of the two models.

[0062] The key of the present application is to distill and learn a local student model by using a general and large intermediate treatment information generation model and a prognosis information generation model, thereby increasing the applicability of the student model locally.

[0063] Reference Figure 2 The student model includes: A feature input layer is used to input digital examination information and extract feature information from the digital examination information.

[0064] The feature input layer is a fully connected feature fusion network used to extract feature information from the digital examination information and splice the feature information.

[0065] A treatment information generation layer outputs corresponding digital treatment information according to the input feature information.

[0066] The treatment information generation layer is a Transformer-based sequence decision network, which simulates the progressive decision-making process of the clinical treatment path. It regards the treatment plan generation as a multi-step sequence decision task: decomposed into specific operation steps in logical order (such as surgical plan → rehabilitation plan). Each step dynamically selects the current optimal operation from the preset feasible operation set through a micro-classifier.

[0067] The prognosis information generation layer generates digital prognosis information according to the digital treatment information and digital examination information output by the treatment information generation layer.

[0068] The prognosis information generation layer is a gated feature interaction network that builds a mapping relationship between "patient basic state + treatment plan" and "prognosis result". The treatment plan vector is decoded into a computable embedding representation (such as different surgical methods corresponding to different numerical impacts), and dynamically weighted with patient examination features. Through the gating unit, the contribution weights of "patient basic condition" and "treatment plan" to the prognosis are automatically adjusted, and finally the corresponding digital prognosis information is output through a multi-level nonlinear transformation of the residual network.

[0069] The specific functions and structures of the feature input layer, the treatment information generation layer, and the prognosis information generation layer have been described, and the specific network structure and input data format of the feature input layer, the treatment information generation layer, and the prognosis information generation layer will not be further described. The key of the present application is to set different loss terms for each functional part of the student model and perform corresponding training design.

[0070] The intermediate treatment information generation model is signal connected with the treatment information generation layer, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is taken as the first loss term; The prognosis information generation model is signal connected with the prognosis information generation layer, and the error between the output of the prognosis information generation model and the output of the prognosis information generation layer is taken as the second loss term; The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer; The second loss term and the total loss of the student model are used to update the model parameters of the prognosis information generation layer.

[0071] The total loss of the student model is the loss between the final output obtained when the student model inputs the training sample and the label in the training sample. The total loss is used to describe the gap between the predicted data of the student model and the labeled data. According to the gap between the predicted data and the labeled data, the network parameters of the student model are corrected in reverse.

[0072] The training of the student model is realized by two-stage knowledge transfer and local data fine-tuning. After obtaining a training sample, the training sample is input into the student model and the teacher model respectively, the output of the intermediate treatment information generation model in the teacher model is calculated with the output of the treatment information generation layer, and the first loss term is calculated; Then, the output of the prognosis information generation model in the teacher model is calculated with the output of the updated prognosis information generation layer, and the second loss term is calculated; Finally, the total loss of the student model is calculated by comparing the output of the student model with the label of the training sample. Then, the total loss and the first loss term are fused into a first loss function value according to the weighting parameter, and the model parameters of the treatment information generation layer are corrected by the first loss function value.

[0073] The total loss and the second loss term are fused into a second loss function value according to the weighting parameter, and the model parameters of the treatment information generation layer are corrected by the second loss function value.

[0074] Therefore, compared with the traditional neural network model training scheme, the original total loss used to guide the update of the model parameters in the whole model is replaced by two different loss information to replace the model parameters of two different network layers. In this way, the training accuracy of the student model is ensured.

[0075] How to update the model parameters according to the loss function value is prior art, for example, using Adam optimizer to update the model parameters.

[0076] The prognosis information generation layer integrates the inspection information and the generated treatment information, outputs the prognosis prediction result, and compares it with the prediction effect of the prognosis information generation model to generate the second loss term, which drives the parameter update in the prognosis stage. In this process, both kinds of teacher loss terms are combined with the supervision loss (total loss) of the local real cases for back propagation, and the global knowledge inheritance and local feature adaptation are coordinated through the weighted fusion mechanism, so that the model can simultaneously improve the localization feasibility of the treatment scheme and the clinical reliability of the prognosis evaluation in the continuous iteration.

[0077] The above gives the specific structure of the student model, the construction method of the teacher model, and the training principle. The specific training process is not described here. The general logic is that during training, each training sample is input into the student model and the teacher model to obtain the first loss term, the second loss term, and the total loss term. The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer, and the second loss term and the total loss of the student model are used to update the model parameters of the prognosis information generation layer, until the training result of the student model reaches the expectation. Then, the student model that reaches the expectation is put into the simulation module.

[0078] An emulation module integrates a student model as a final localized emulation model configured to: input examination information of a target patient; and based on the input examination information, emulate output intermediate treatment information of a potential fracture surgery plan made for the target patient and a corresponding prognosis information prediction result.

[0079] The above only is the preferred embodiment of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A surgical simulation system for fracture treatment based on big data analysis, characterized in that, include: The data preprocessing module is configured to collect local hospital case datasets that include preliminary examination information, intermediate treatment information, and prognostic information. The adversarial network module receives a local hospital case dataset and uses at least one adversarial generative neural network model to train and augment the local hospital case dataset to generate a simulated case dataset. The simulated case dataset matches the local medical resources and capabilities represented by the hospital's local case dataset in terms of statistical distribution; The model training and optimization module is configured as follows: We pre-trained a general large-scale fracture treatment dataset using multiple heterogeneous teacher models, and then transferred the fracture treatment decision-making knowledge learned by the multiple heterogeneous teacher models to a single student model. During transfer learning, the student model is trained a second time using a simulated case dataset generated by the adversarial network module. The simulation module integrates a student model as the final localized simulation model, which is configured as follows: The process involves inputting preliminary examination information of the target patient, and simulating the output of intermediate treatment information and corresponding prognostic prediction results for the potential fracture surgical plan developed for the target patient based on the input preliminary examination information.

2. The fracture treatment surgical simulation system based on big data analysis according to claim 1, characterized in that, The data preprocessing module is configured as follows: Using a structural feature extraction network, non-numerical or unstructured data in the collected preliminary examination information, intermediate treatment information, and prognostic information are converted into structured numerical feature vectors according to a preset standard encoding table. Using an image feature extraction network, medical image data from the collected preliminary examination information, intermediate treatment information, and prognostic information are processed to extract structured visual feature vectors related to fracture morphology, degree of comminutedness, and surrounding soft tissue condition. By utilizing feature organization networks, preliminary examination information, intermediate treatment information, and prognostic information are transformed into structured numerical feature vectors and structured visual feature vectors, respectively, to obtain digital examination information, digital treatment information, and digital prognostic information.

3. The fracture treatment surgical simulation system based on big data analysis according to claim 2, characterized in that, The adversarial network module includes: The noise source generation module determines the noise source generation range based on the local medical conditions of the hospital, and generates noise signals within the noise source generation range; The generator generates simulated samples based on noise signals. The simulated samples include simulated digital examination information, digital treatment information, and digital prognostic information. The discriminator, forming an adversarial network with the generator, distinguishes the authenticity of input simulated samples from real samples; The simulation data generator has a built-in pre-trained generator that generates new simulation samples based on new noise sequences, and uses these new simulation samples to generate a simulated case dataset.

4. The fracture treatment surgical simulation system based on big data analysis according to claim 3, characterized in that, The noise signal is generated as follows: S1: Load the medical resources of the local hospital, generate the original encoding table from the local medical resources, map the original encoding table to the standard encoding table, and determine the vertical dimension of the noise signal based on the dimension of the original encoding table. Load the maximum dimension of the feature vector of real cases from the local hospital and use it as the horizontal dimension of the noise signal; The vertical dimension is used to constrain the numerical fluctuation range of the noise signal, while the horizontal dimension is used to constrain the length of the noise signal. S2: Predefine an entropy source seed signal source and collect random signals generated by the entropy source seed signal source; S3: Based on the proportional information, each element in the random signal is sequentially mapped to the element at the corresponding position in the noise signal to generate the noise signal.

5. The fracture treatment surgical simulation system based on big data analysis according to claim 4, characterized in that, The noise signal is , ; express The j-th element in the noise signal, where j represents the index of the element in the noise signal; ； in, This represents the value of the j-th element of the random signal. This represents the maximum value of the random signal. This represents the maximum value of the noise signal at the j-th bit. This indicates rounding down to the nearest integer.

6. The fracture treatment surgical simulation system based on big data analysis according to claim 5, characterized in that, The generator includes: The generator input layer is used to receive the noise vector; The feature transformation layer, based on fully connected layers and residual networks, extracts hidden features from noise vectors; The generator output layer outputs simulated samples based on the hidden features of the input.

7. The fracture treatment surgical simulation system based on big data analysis according to claim 6, characterized in that, The discriminator includes: The feature pyramid backbone network is used to take simulated or real samples as input to generate case features; The true / false discrimination head, based on the Sigmoid function, regresses case features to the probability that the sample is true.

8. The fracture treatment surgical simulation system based on big data analysis according to claim 7, characterized in that, The loss functions for the generator and discriminator are: ； ； ； Let E represent the discriminator loss, and let E represent the mathematical expectation operator. Let x represent a real sample drawn from the real data distribution. This represents the penalty sample generated through interpolation. Let x represent a mixed random number, and let x represent a real sample. This represents the generated samples produced by the generator; Let z represent the noise vector sampled from the noise distribution. This represents the discriminator's score for the real sample. This indicates that the discriminator evaluates the generated samples. The rating, The weights of the gradient penalty term are represented by the coefficients. This represents a sample drawn from a penalty distribution. This indicates that the discriminator output is related to the input. The gradient; Let l represent the feature matching loss, and l represent the hierarchical index of the feature pyramid. This represents the feature map output by the discriminator for the real sample x at the l-th layer. This indicates that the discriminator generates samples for each pair. The feature map output at layer l. This represents the square of the L2 norm.

9. The fracture treatment surgical simulation system based on big data analysis according to claim 2, characterized in that, The student model includes: The feature input layer is used to input digital inspection information and extract feature information from the digital inspection information; The treatment information generation layer outputs corresponding digital treatment information based on the input feature information. The prognostic information generation layer generates digital prognostic information based on the digital treatment information and digital examination information output by the treatment information generation layer.

10. The fracture treatment surgical simulation system based on big data analysis according to claim 9, characterized in that, The intermediate treatment information generation model is connected to the treatment information generation layer signal, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is used as the first loss term. The prognostic information generation model is connected to the prognostic information generation layer signal, and the output of the prognostic information generation model and the error between the prognostic information generation layer are used as the second loss term. The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer; The second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer.

Images