A follow-up system and method for post-orthognathic surgery of maxillofacial deformity

Abstract

The application relates to the technical field of jaw and facial deformity treatment, and discloses a jaw and facial deformity orthognathic surgery postoperative follow-up system and method, which comprises an image analysis model and a follow-up evaluation model; the image analysis model carries out multi-dimensional feature fusion processing on initial postoperative data, and the data contains three-dimensional suture, occlusion function, soft tissue distribution and other feature data; the follow-up evaluation model inputs feature generation layers after dynamic interference suppression processing of real-time physiological data flow to simulate a recovery track, and then generates a follow-up evaluation instruction. The feature generation layers comprise an image preprocessing module and a dynamic feature generation module, the latter is obtained through joint training of multiple types of historical postoperative data and real-time physiological parameters, and comprises a feature analysis layer, a feature correction layer and an instruction triggering layer. The method applies the system to realize multi-dimensional and intelligent follow-up evaluation of postoperative recovery. The application improves the accuracy and real-time performance of follow-up, and provides effective support for postoperative recovery monitoring and intervention.

Description

Technical Field

[0001] This invention relates to the field of maxillofacial deformity treatment technology, specifically to a follow-up system and method for orthognathic surgery of maxillofacial deformities. Background Technology

[0002] Orthognathic surgery is an important treatment for maxillofacial deformities, restoring facial aesthetics and occlusal function by adjusting the position and shape of the jawbone. However, the postoperative recovery process is complex, involving the dynamic evolution of physiological indicators in multiple dimensions, such as suture healing, occlusal function reconstruction, and soft tissue morphological changes. Traditional follow-up methods have significant limitations.

[0003] From the perspective of data collection and analysis, traditional follow-up mainly relies on doctors' clinical examinations and patients' subjective feedback, lacking precise quantitative analysis of multi-dimensional data such as three-dimensional suture characteristics, occlusal function parameters, and soft tissue distribution. For example, the dynamic changes in bone displacement parameters and suture healing parameters are difficult to fully capture through traditional palpation or two-dimensional imaging, resulting in the inability to detect abnormal suture healing or deviations in occlusal function recovery in a timely manner. At the same time, the raw image data stream often contains a large amount of invalid signals, and traditional methods are unable to efficiently complete anatomical segmentation and noise filtering, affecting data accuracy.

[0004] Regarding the real-time and dynamic aspects of follow-up assessments, existing systems cannot effectively process real-time physiological data streams. During postoperative recovery, minute fluctuations in physiological parameters may indicate deviations in the recovery trajectory, but traditional methods lack dynamic interference suppression mechanisms and cannot eliminate interference from external factors in real time, leading to assessment delays. Furthermore, the system lacks sufficient capability for fusion analysis of multi-type feature data, and the dynamic mapping relationships between various feature vectors are not adequately modeled, making it difficult to simulate the true recovery trajectory and resulting in poor targeting and accuracy of follow-up assessment instructions.

[0005] From the perspective of model training and parameter optimization, traditional follow-up systems lack a joint training mechanism based on historical postoperative data and real-time physiological parameters. The absence of a dynamic feature generation module prevents personalized feature analysis and correction based on individual differences. For example, different patients exhibit variations in bone suture healing speed and soft tissue repair capacity, making it difficult for traditional methods to generate appropriate bias-compensating data. Furthermore, the lack of dynamic weight optimization and real-time abnormal state identification capabilities during command triggering makes it impossible to minimize evaluation errors through iterative optimization and accurately locate abnormal healing sites.

[0006] In clinical applications, traditional follow-up methods prevent doctors from fully grasping the subtle changes in patients' postoperative recovery, potentially missing the optimal intervention window. For example, if local misalignment during suture healing is not detected in time, it may lead to recurrence of maxillofacial deformities or worsening of occlusal dysfunction; abnormal soft tissue distribution, if not adjusted promptly, may affect the aesthetic recovery of the face. Furthermore, manual assessment is highly subjective and inefficient, making it difficult to meet the follow-up needs of a large number of postoperative patients. There is an urgent need for intelligent and precise follow-up systems to improve the quality of medical services. Summary of the Invention

[0007] The purpose of this invention is to provide a follow-up system and method for orthognathic surgery for maxillofacial deformities, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a follow-up system for orthognathic surgery for maxillofacial deformities, the system comprising:

[0009] An image analysis model is used to perform multi-dimensional feature fusion processing on postoperative initial data. The postoperative initial data includes a first spatial vector corresponding to three-dimensional suture feature data, a second correlation matrix corresponding to occlusal function feature data, and a third mapping sequence corresponding to soft tissue distribution feature data. The three-dimensional suture feature data includes a first distribution label generated by bone block displacement parameters and a second quantization label generated by suture healing parameters. The multi-dimensional feature fusion processing includes spatial coordinate normalization processing of the first spatial vector, time series alignment processing of the second correlation matrix, and anatomical region annotation processing of the third mapping sequence. The three types of processed data are input into a feature fusion network, and cross-modal fusion features are extracted through layer-by-layer feature cross-interaction to form a multi-dimensional fusion feature set.

[0010] A follow-up assessment model is used to perform dynamic interference suppression processing on real-time physiological data streams and input them into a feature generation layer for recovery trajectory simulation. Follow-up assessment instructions are generated based on the output of the feature generation layer. The dynamic interference suppression processing includes using an adaptive filtering algorithm to suppress high-frequency noise and removing abnormal fluctuation signals through joint time-domain and frequency-domain analysis. The recovery trajectory simulation includes matching the processed real-time physiological data stream with corresponding stage data of multiple types of historical postoperative data, extracting the characteristic change trends of each time node, and generating a postoperative recovery prediction trajectory through time-series prediction.

[0011] The feature generation layer includes an image preprocessing module and a dynamic feature generation module. The image preprocessing module is used to perform anatomical interval segmentation and invalid signal filtering on the original image data stream. The dynamic feature generation module is obtained by joint training based on multiple types of historical postoperative data and real-time physiological parameters. The dynamic feature generation module is used to construct a neural network architecture that includes a feature parsing layer, a feature correction layer and an instruction triggering layer.

[0012] Preferably, the feature parsing layer is used to perform anatomical feature association processing on different feature vectors in the original image data stream to generate feature fusion data; the feature correction layer is used to model the dynamic mapping relationship between the feature fusion data corresponding to each feature vector to generate deviation compensation data; and the instruction triggering layer is used to perform multi-dimensional feature fusion based on the deviation compensation data and the feature fusion data to generate follow-up evaluation instructions.

[0013] Preferably, the step of modeling the dynamic mapping relationship between the feature fusion data corresponding to each feature vector to generate bias compensation data includes:

[0014] The bone suture segmentation algorithm is used to identify key healing nodes in the feature fusion data, and the correction compensation mapping corresponding to each feature vector is determined based on the physiological parameter type corresponding to each key healing node.

[0015] Calculate the offset coefficient between healing nodes with the same parameters in the correction compensation map corresponding to any two feature vectors, and generate deviation compensation data between the two feature vectors based on the offset coefficient.

[0016] Preferably, calculating the offset coefficient between healing nodes with the same parameters in the corrected compensation mapping corresponding to any two feature vectors includes:

[0017] When the number of healing nodes in the correction compensation mapping corresponding to any two feature vectors is inconsistent, virtual healing point interpolation is performed based on the physiological parameters corresponding to the terminal healing node in the one with fewer healing nodes, and the offset coefficient between healing nodes with the same parameters is calculated based on the interpolated data.

[0018] Preferably, the image preprocessing module is specifically used for:

[0019] Based on the preset anatomical interval, the first spatial vector, the second correlation matrix, and the third mapping sequence are divided into equal gradients to generate standardized suture data, standardized occlusion data, and standardized soft tissue data.

[0020] The standardized suture data and standardized occlusion data are matched in real time using a dynamic parameter alignment method, and the standardized soft tissue data are steadily optimized using a fixed anatomical window mechanism, outputting a first calibration vector, a second calibration matrix and a third calibration sequence; wherein, the first calibration vector includes a calibrated first distribution label and a calibrated second quantization label.

[0021] Preferably, the image preprocessing module is further used for:

[0022] Calculate the healing offset coefficient between the calibrated first distribution label and the calibrated second quantization label during the historical follow-up period;

[0023] The expected distribution value of the calibrated second quantization label in the real-time follow-up phase is predicted based on the healing offset coefficient and the parameter value of the calibrated first distribution label in the real-time follow-up phase.

[0024] Target correction compensation data is generated based on the calibrated second quantization label and its expected distribution value, and the feature vector corresponding to the target correction compensation data is used as the first calibration vector.

[0025] Preferably, the feature correction layer specifically includes:

[0026] The feature tracing unit is used to perform healing feature tracing on each feature vector in the feature fusion data, so as to extract the corresponding healing conduction chain from each feature vector;

[0027] The parameter overlay unit is used to overlay the healing conduction chain extracted from each feature vector with the corresponding feature fusion data to generate bias compensation data.

[0028] Preferably, the feature correction layer further includes:

[0029] The noise suppression unit is used to perform phase offset effect elimination processing on the deviation compensation data.

[0030] Preferably, the instruction triggering layer specifically includes:

[0031] The multi-dimensional feature collaboration unit contains multiple instruction decision nodes, and each instruction decision node is connected to each feature vector in the deviation compensation data and feature fusion data through configuration parameters.

[0032] The dynamic weight optimization unit is used to iteratively optimize the configuration parameters through a dynamic weight adjustment algorithm to minimize the error between the follow-up evaluation instructions and the actual recovery trajectory.

[0033] An abnormal state identification unit is used to locate abnormal healing based on the deviation compensation data and feature fusion data, and generate follow-up evaluation instructions.

[0034] Preferably, the present invention further includes a follow-up method for orthognathic surgery for maxillofacial deformities, applied to the follow-up system for orthognathic surgery for maxillofacial deformities described in any one of the above-mentioned methods, comprising the following steps:

[0035] The initial postoperative data is processed by multi-dimensional feature fusion using an image analysis model. The initial postoperative data includes a first spatial vector corresponding to the three-dimensional suture feature data, a second correlation matrix corresponding to the occlusal function feature data, and a third mapping sequence corresponding to the soft tissue distribution feature data. The three-dimensional suture feature data includes a first distribution label generated by bone block displacement parameters and a second quantitative label generated by suture healing parameters.

[0036] The real-time physiological data stream is dynamically suppressed by a follow-up assessment model, and the processed data stream is input into the feature generation layer to simulate the recovery trajectory.

[0037] The image preprocessing module in the feature generation layer is used to perform anatomical region segmentation and invalid signal filtering on the raw image data stream;

[0038] A dynamic feature generation module, which is jointly trained from multiple types of historical postoperative data and real-time physiological parameters, is used to simulate the recovery trajectory. The dynamic feature generation module includes a feature parsing layer, a feature correction layer, and an instruction triggering layer connected in sequence.

[0039] Based on the output of the feature generation layer, the follow-up evaluation model generates follow-up evaluation instructions.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] In terms of data processing and feature analysis, the image analysis model can perform multi-dimensional feature fusion processing on initial postoperative data. By integrating the first spatial vector corresponding to the three-dimensional suture feature data, the second correlation matrix corresponding to the occlusal function feature data, and the third mapping sequence corresponding to the soft tissue distribution feature data, and combining the first distribution label generated by bone displacement parameters and the second quantitative label generated by suture healing parameters, refined modeling of key postoperative recovery indicators is achieved. The image preprocessing module divides the data into equal gradients through preset anatomical intervals, and combines dynamic parameter alignment and fixed anatomical window mechanisms to generate standardized and calibrated data streams, effectively improving data consistency and reliability, and providing an accurate data foundation for subsequent evaluation.

[0042] The dynamic interference suppression mechanism of the follow-up assessment model can eliminate noise and external interference in the real-time physiological data stream, ensuring that the data input to the feature generation layer truly reflects the patient's physiological state. The dynamic feature generation module of the feature generation layer is jointly trained based on multiple types of historical postoperative data and real-time physiological parameters, possessing strong generalization ability and personalized adaptability. The feature parsing layer generates feature fusion data through anatomical feature correlation processing; the feature correction layer generates bias compensation data by modeling the dynamic mapping relationship between feature vectors; and the instruction triggering layer generates follow-up assessment instructions through multi-dimensional feature fusion, forming a complete intelligent processing chain from data parsing to instruction generation.

[0043] In terms of feature correction and evaluation command optimization, the feature correction layer extracts the healing transmission chain through the feature tracing unit. Combined with parameter superposition and noise suppression processing, it can accurately capture the dynamic deviation between each feature vector and generate compensation data, effectively improving the accuracy of recovery trajectory simulation. The multi-dimensional feature collaboration unit of the command triggering layer connects each feature vector through configuration parameters, the dynamic weight optimization unit minimizes the evaluation error through iterative optimization, and the abnormal state identification unit accurately locates abnormal healing sites based on data, ensuring the scientific nature and pertinence of follow-up evaluation commands. For example, when the physiological parameters of key nodes in bone suture healing show abnormal deviations, the system can quickly identify and generate intervention commands to prevent further expansion of recovery deviations.

[0044] In terms of clinical application value, this system enables full-cycle, multi-dimensional dynamic monitoring of the postoperative recovery process, providing doctors with real-time and accurate assessment data. By predicting the healing trend of sutures, the recovery trajectory of occlusal function, and changes in soft tissue morphology in advance, doctors can adjust treatment plans in a timely manner, intervene in abnormal recovery processes, reduce the risk of recurrence of maxillofacial deformities, and improve the recovery of occlusal function and facial aesthetics. Simultaneously, the intelligent data processing and assessment mechanism significantly improves follow-up efficiency, reduces the subjectivity and error of manual assessment, and can be widely applied to postoperative patient follow-up, optimizing the allocation of medical resources and improving the overall level of medical services. Furthermore, the system's scalability provides a foundation for subsequent integration with more physiological parameters and new imaging technologies, contributing to the digital and intelligent development of the field of maxillofacial deformity treatment. Attached Figure Description

[0045] Figure 1 This is a schematic diagram illustrating the working principle of the orthognathic surgery follow-up system for maxillofacial deformities described in this invention.

[0046] Figure 2 This is a schematic diagram of the feature parsing-correction-triggering process in the feature generation layer.

[0047] Figure 3 This is a flowchart of the method for generating bias compensation data in the feature correction layer;

[0048] Figure 4 A flowchart for data calibration processing in the image preprocessing module;

[0049] Figure 5 This is a flowchart for predicting and compensating healing parameters based on historical data. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figures 1-5 This invention relates to a follow-up system and method for orthognathic surgery for maxillofacial deformities.

[0052] By combining image analysis and follow-up assessment models, accurate assessment and dynamic follow-up of postoperative recovery can be achieved. The specific implementation steps are as follows:

[0053] The image analysis model is used to perform multi-dimensional feature fusion processing on postoperative initial data. The postoperative initial data includes a first spatial vector corresponding to three-dimensional suture feature data, a second correlation matrix corresponding to occlusal function feature data, and a third mapping sequence corresponding to soft tissue distribution feature data. The three-dimensional suture feature data further includes a first distribution label generated from bone displacement parameters and a second quantization label generated from suture healing parameters. Through the fusion processing of the above multi-dimensional data, the image analysis model extracts key features reflecting the initial postoperative state.

[0054] In this embodiment, the multi-dimensional feature fusion processing of the image analysis model can be achieved in the following ways: In the initial postoperative data acquisition stage, three-dimensional bone suture feature data is acquired by cone-beam computed tomography (CBCT) scanning, and the raw data containing bone block displacement parameters (such as bone block translation and rotation angle) and bone suture healing parameters (such as bone suture density and callus coverage) is extracted by point cloud and converted into a first spatial vector for storage; occlusal function feature data is acquired by a dynamic occlusal recorder, which records parameters such as the distribution of occlusal contact points and occlusal force values ​​at different postoperative time points and converts them into a second correlation matrix arranged in time series; soft tissue distribution feature data is acquired by a three-dimensional facial scanner, which records parameters such as facial soft tissue thickness and volume distribution and converts them into a third mapping sequence containing anatomical location information.

[0055] For the spatial coordinate normalization processing of the first spatial vector, a standard coordinate system is first established with the patient's anterior nasal spine (ANS) and sella turcica (S) as reference points to establish the original three-dimensional coordinates of the CBCT scan. The original coordinate values ​​of each bone block displacement parameter are then converted into standardized coordinate values ​​relative to this reference coordinate system to eliminate spatial deviations caused by different scanning devices or scanning angles.

[0056] For the time series alignment processing of the second correlation matrix, the original timestamps of each time point in the occlusal function feature data are uniformly converted into postoperative days (such as 7 days, 14 days, 30 days, etc.) with the surgery date as the time origin. The occlusal data of missing time points are supplemented by linear interpolation to ensure that the time intervals of each time series in the second correlation matrix are consistent.

[0057] For the anatomical region annotation processing of the third mapping sequence, a soft tissue anatomical atlas template containing key anatomical landmarks such as the orbit, nasolabial fold, and mandibular border is pre-established. The soft tissue distribution data obtained by three-dimensional surface scanning is registered with the template, and the anatomical names corresponding to each soft tissue region (such as "cheek soft tissue" and "chin soft tissue") are automatically labeled to form a third mapping sequence with anatomical labels.

[0058] After the above preprocessing is completed, the normalized first spatial vector, the aligned second correlation matrix, and the labeled third mapping sequence are input into the feature fusion network. This network contains a three-layer feature cross structure: the first layer splices and fuses the local features of each modality data (such as the local displacement features of sutures, the local contact features of occlusion, and the local thickness features of soft tissue); the second layer performs weighted fusion of the key features of each modality (such as key areas of suture healing, abnormal points of occlusion function, and soft tissue depression areas) through an attention mechanism; the third layer performs global context association on the fused features output from the first two layers, ultimately forming a multi-dimensional fused feature set containing spatial location, temporal changes, and anatomical positioning information.

[0059] The follow-up assessment model is used to dynamically suppress interference in the real-time physiological data stream, removing noise or irrelevant interference during data acquisition to ensure the accuracy of the input data. The processed data stream is then input into the feature generation layer for trajectory simulation. The implementation process of the follow-up assessment model is as follows: Real-time physiological data streams are acquired in real time through wearable physiological monitoring devices (such as oral pressure sensors and facial motion sensors), including parameters such as occlusal force, mandibular movement trajectory, and soft tissue deformation. During dynamic interference suppression, high-frequency noise is first suppressed using an adaptive filtering algorithm: the filtering parameters are dynamically adjusted according to the noise level of the real-time data stream (e.g., when a high-frequency fluctuation higher than the baseline value is detected in the occlusal force data, the cutoff frequency of the low-pass filter is increased); subsequently, joint time-domain and frequency-domain analysis is performed to extract the time-domain statistical features (such as mean and variance) and frequency-domain power spectrum features (such as main frequency components) of the data stream. Signals that simultaneously meet the conditions of a time-domain variance exceeding three times the historical mean and the presence of non-physiological frequency components in the frequency domain (such as abnormal jitter >10Hz) are identified as abnormal fluctuations and removed.

[0060] In the recovery trajectory simulation phase, the processed real-time physiological data stream is matched with various types of historical postoperative data. Historical data is categorized and stored according to surgical type (e.g., bimaxillary surgery, unimaxillary surgery), patient age (e.g., 18-30 years, 31-50 years), and postoperative time stage (e.g., early healing period, bone remodeling period). During matching, the target historical dataset is first determined based on the current patient's surgical type and age, and then historical physiological data for the corresponding stage is extracted according to the postoperative time stage (e.g., if the current time is 2 weeks postoperative, the historical data for the period of 2 weeks ± 3 days postoperatively is matched). The characteristic change trends of each time node are extracted using the sliding window method (e.g., the weekly increase in occlusal force and the weekly expansion of mandibular range of motion). The change trend of the current real-time data is compared with the historical trend. If the difference exceeds a preset threshold (e.g., the increase in occlusal force is less than 50% of the historical average), the postoperative recovery prediction trajectory is generated through the time-series prediction module. This trajectory includes the expected change curves of occlusal force, mandibular range of motion, and soft tissue morphology within the next 4 weeks.

[0061] Finally, the follow-up assessment model generates specific follow-up assessment instructions based on the fusion features output by the feature generation layer and the recovery prediction trajectory. For example, when abnormal displacement parameters are detected in the key area of ​​bone suture healing and the occlusal force increase trend is lower than the historical average, the instruction "It is recommended to add imaging follow-up 4 weeks after surgery" is generated; when the deformation trend of the soft tissue depression area is consistent with the historical normal recovery trajectory, the instruction "The current recovery is good, maintain the existing follow-up plan" is generated.

[0062] The feature generation layer comprises an image preprocessing module and a dynamic feature generation module. The image preprocessing module is responsible for anatomical segmentation and invalid signal filtering of the raw image data stream to standardize the data format and eliminate interference. The dynamic feature generation module is jointly trained based on multiple types of historical postoperative data and real-time physiological parameters. The neural network architecture of the dynamic feature generation module is implemented as follows: a three-level cascaded deep learning network is constructed, including a feature parsing layer, a feature correction layer, and an instruction triggering layer. The layers form a collaborative optimization architecture through feature transfer and parameter sharing mechanisms.

[0063] The feature parsing layer employs a convolutional neural network structure with a multi-branch attention mechanism to process the preprocessed first calibration vector, second calibration matrix, and third calibration sequence in parallel. Specifically, a spatial attention module is designed for the first calibration vector, generating channel weights through 1×1 convolution and global pooling operations to enhance spatial attention to suture features; a temporal attention module is designed for the second calibration matrix, using gated recurrent units to capture temporal variations in occlusal function; and a hybrid attention module is designed for the third calibration sequence, fusing spatial and temporal attention mechanisms to process soft tissue distribution features. The outputs of each branch are concatenated using tensors to form multimodal feature fusion data.

[0064] The feature correction layer employs a residual connection encoding / decoding architecture, comprising a feature tracing unit and a parameter stacking unit. The feature tracing unit, implemented using a bidirectional long short-term memory network, performs temporal backtracking on the feature fusion data, extracting the healing conduction chain feature representation at each time step. The parameter stacking unit performs element-wise addition operations on the healing conduction chain features and the original feature fusion data to generate bias-compensated data containing temporal context information. During parameter stacking, learnable weight coefficients are introduced to dynamically adjust the healing conduction chain features; these weight coefficients are trained using an adaptive moment estimation optimization algorithm.

[0065] The instruction triggering layer employs a graph neural network structure, comprising a multi-dimensional feature coordination unit, a dynamic weight optimization unit, and an anomaly state recognition unit. The multi-dimensional feature coordination unit maps bias compensation data and feature fusion data into node features, learning the structural relationships between feature vectors through a graph convolutional network. The dynamic weight optimization unit, based on a reinforcement learning framework, dynamically adjusts the edge weights in the graph neural network using a policy gradient algorithm to minimize the error between the follow-up assessment instruction and the actual recovery trajectory. The anomaly state recognition unit uses the isolated forest algorithm to detect outliers in the output features of the graph neural network, triggering the corresponding follow-up assessment instruction when an abnormal healing feature is detected.

[0066] The entire neural network architecture employs an end-to-end training approach, optimizing parameters by minimizing the following objective function: using known recovery trajectories from historical postoperative data as supervision signals, a composite loss function is constructed comprising feature consistency loss, temporal continuity loss, and anomaly detection loss, and the network parameters are iteratively updated using stochastic gradient descent. During training, batch normalization is employed to accelerate convergence, and a dropout method is used to prevent overfitting, ultimately forming a dynamic feature generation module capable of accurately predicting the recovery trajectory after orthognathic surgery for maxillofacial deformities.

[0067] Example 1:

[0068] This embodiment relates to the specific working mechanism of the feature parsing layer. Its core function is to perform anatomical feature association processing on different feature vectors in the original image data stream to generate feature fusion data. The specific implementation method is as follows:

[0069] The raw image data stream originates from multimodal examination data of patients after orthognathic surgery for maxillofacial deformities. It includes three-dimensional suture feature data, occlusal function feature data, and soft tissue distribution feature data, presented in the form of a first spatial vector, a second correlation matrix, and a third mapping sequence, respectively. These data have different physical meanings and data structures: the first spatial vector represents the spatial distribution and dynamic changes of the three-dimensional sutures, including the first distribution label generated by bone block displacement parameters (such as discretized labels of displacement direction and displacement distance) and the second quantification label generated by suture healing parameters (such as continuous quantified values ​​of healing density and healing speed); the second correlation matrix reflects the mechanical transmission relationship of occlusal function, recording the force transmission intensity and synergy between different tooth positions and occlusal contact points; the third mapping sequence describes the spatial distribution morphology and time-varying trajectory of soft tissues (such as skin, muscles, and mucous membranes).

[0070] The feature analysis layer's processing flow begins with the anatomical correlation analysis of multi-source feature vectors. First, based on maxillofacial anatomy principles, the system establishes a spatial mapping relationship between bone and occlusal function for the three-dimensional suture feature data and occlusal function feature data. For example, the anterior displacement of the maxillary bone may lead to anterior shift of the anterior occlusal contact point. The feature analysis layer establishes a spatial positional correlation between these two by identifying the displacement direction of the maxillary bone (a key parameter in the first spatial vector) and the positional change of the anterior occlusal contact point (the corresponding parameter in the second correlation matrix). Specifically, the system pre-defines a computer-recognizable anatomical structure mapping table, dividing the maxilla into multiple sub-regions (such as anterior tooth bone segments and molar bone segments). Each sub-region corresponds to a specific occlusal contact area (such as the anterior occlusal surface and posterior occlusal surface). Through a coordinate mapping algorithm, the bone displacement parameters and occlusal contact point parameters are registered in three-dimensional space, forming a linked feature group of bone and occlusal function.

[0071] For three-dimensional suture feature data and soft tissue distribution feature data, the feature analysis layer analyzes the interaction between the suture healing area and the surrounding soft tissue based on anatomical hierarchical relationships. For example, during the healing process of the mandibular suture, changes in the tension of the surrounding masseter muscle may affect the stability of the bone fragment. The feature analysis layer establishes a correlation model between the soft tissue biomechanical state and the suture healing process by extracting the healing density parameter of the mandibular suture (second quantification label) and the masseter muscle thickness change parameter (soft tissue thickness parameter in the third mapping sequence). In specific operation, the system divides the masseter muscle region around the mandibular suture into multiple analysis units, each unit corresponding to a specific suture sub-region. Through time series analysis, the temporal correlation between changes in masseter muscle thickness and the increase in suture healing density is identified, generating a co-functional feature set of soft tissue-bone healing.

[0072] After completing the anatomical feature correlation analysis of different feature vectors, the feature parsing layer enters the feature fusion data generation stage. For the bone-occlusal function linkage feature group, the system adopts a tensor fusion algorithm to perform dimensional expansion and matrix multiplication operations on the bone block displacement direction and displacement distance parameters in the first spatial vector and the occlusal contact point position and occlusal force transmission intensity parameters in the second correlation matrix, generating a three-dimensional bone-occlusal function fusion tensor. Each element of this tensor represents the joint feature value of a specific bone segment displacement and the force on the corresponding occlusal contact point. For example, the element value can reflect the increase in occlusal force at the anterior tooth occlusal contact point when the maxillary anterior tooth bone segment moves forward by 0.5 mm.

[0073] For the soft tissue-bone healing synergistic feature group, the system employs a sequence alignment algorithm to align the soft tissue thickness change sequence in the third mapping sequence with the bone suture healing density change sequence in the first spatial vector along the time axis. Dynamic time warping (DTW) is then used to eliminate temporal deviations caused by differences in data acquisition time, thereby generating a two-dimensional soft tissue-bone healing fusion matrix. The row dimension of the matrix represents time nodes, and the column dimension represents anatomical location nodes. Each matrix element records the joint feature value of soft tissue thickness and bone suture healing density at the corresponding time point and anatomical location. For example, the element value can reflect the product of the change rate of masseter muscle thickness in the mandibular angle suture region and the increase rate of bone suture healing density in that region on postoperative day 7.

[0074] Furthermore, the feature analysis layer also needs to handle the correlation between occlusal function feature data and soft tissue distribution feature data. For example, abnormal occlusal function may lead to uneven force on the cheek soft tissue, causing changes in soft tissue texture. The feature analysis layer establishes a mechanical transmission feature set of occlusion-soft tissue by extracting the occlusal contact point distribution uniformity parameter from the second correlation matrix and the cheek soft tissue texture feature parameters (such as gray-level co-occurrence matrix eigenvalues) from the third mapping sequence. The system first divides the cheek soft tissue into meshes, with each mesh cell corresponding to a specific tooth position occlusal region. Through finite element analysis, it simulates the stress distribution of soft tissue under different occlusal contact modes, and then maps the occlusal contact point distribution uniformity parameter to the stress parameter of the soft tissue mesh cell, generating an occlusal-soft tissue stress fusion vector. Each component of this vector represents the joint feature value of the stress state and occlusal uniformity of the corresponding mesh cell.

[0075] During feature fusion, the system needs to normalize the feature parameters of different dimensions to eliminate the influence of dimensional differences on the fusion results. For the displacement distance parameter (unit: mm) in the first spatial vector and the bite force parameter (unit: N) in the second correlation matrix, the Z-score normalization method is used to convert each parameter into standard normal distribution data with a mean of 0 and a standard deviation of 1. For the soft tissue thickness parameter (unit: mm) and suture healing density parameter (dimensionless) in the third mapping sequence, the min-max normalization method is used to scale the parameter range to the [0,1] interval. The normalized data are combined into a unified feature vector space through feature concatenation, providing standardized input data for subsequent feature correction and instruction generation.

[0076] The feature parsing layer's processing is entirely based on anatomical principles and data-driven correlation analysis, without relying on any empirical assumptions or preset thresholds. Through the correlation and fusion of anatomical features from multi-dimensional feature vectors, the system can extract complex correlation features reflecting postoperative recovery after orthognathic surgery for maxillofacial deformities from the raw data. These features not only contain independent information from a single modality but also encompass information on the synergistic effects between multiple modalities, providing a more comprehensive feature expression for dynamic interference suppression and recovery trajectory simulation in the follow-up evaluation model. For example, the three-dimensional fusion features of bone-occlusion-soft tissue can comprehensively reflect the overall impact of changes in skeletal structure on function and appearance, enabling the system to comprehensively assess postoperative recovery from three levels: anatomical structure, functional status, and tissue morphology, avoiding the limitations of single-dimensional assessment.

[0077] Example 2:

[0078] This embodiment details the process of modeling dynamic mapping relationships and generating bias compensation data using the feature correction layer. Specific steps include:

[0079] First, a suture segmentation algorithm is used to identify key healing nodes in the feature fusion data. Based on the spatial distribution and grayscale differences of 3D suture feature data, the algorithm automatically identifies key locations in the suture healing process, such as bone fragment ends and the healing front. The algorithm first preprocesses the 3D suture feature data, including denoising, smoothing, and normalization operations, to improve feature clarity and distinguishability. Next, threshold segmentation and region growing techniques are used to separate the suture region from the surrounding tissue, generating a binary mask of the suture. Then, morphological operations and contour extraction are used to further refine the suture boundary, determining the precise location and shape of the suture.

[0080] After identifying the suture region, the algorithm further analyzes the internal structure and features of the suture to determine key healing nodes. These nodes typically correspond to specific anatomical locations within the suture, such as the junction of bone fragments, stress concentration areas, or the passage of neurovascular bundles. The algorithm identifies these key nodes by calculating the curvature, grayscale gradient, and texture features within the suture region, combined with pre-defined anatomical knowledge and rules. Each key healing node corresponds to a specific physiological parameter type, such as displacement velocity or healing density. Based on these parameter types, a corresponding correction and compensation mapping is determined for each feature vector, establishing a correspondence between each parameter in the feature vector and the key healing node.

[0081] Calculate the offset coefficient between healing nodes with the same parameters in the correction compensation maps corresponding to any two eigenvectors. When the number of healing nodes in the correction compensation maps of two eigenvectors is inconsistent, virtual healing point interpolation needs to be performed based on the physiological parameters corresponding to the terminal healing nodes of the eigenvector with fewer healing nodes. For example, if the correction compensation map of eigenvector A contains 5 healing nodes and the map of eigenvector B contains 3 healing nodes, then based on the terminal node parameters of eigenvector B, virtual parameter values ​​are generated at the positions of its missing nodes using an interpolation algorithm to align the number and position of nodes in both eigenvectors.

[0082] After interpolation, for healing nodes with the same parameters (such as displacement velocity), the percentage difference in parameter values ​​or the spatial distance difference is calculated as offset coefficients. The offset coefficients reflect the synergy or difference between different feature vectors along the same parameter dimension. Based on all offset coefficients, deviation compensation data between any two feature vectors is generated. This data is used to adjust the parameter weights or mapping relationships during feature fusion to compensate for evaluation bias caused by differences in feature vectors.

[0083] In generating deviation compensation data, the feature correction layer also considers the time dimension. Since the recovery after orthognathic surgery for maxillofacial deformities is a dynamic process, there may be temporal correlations between feature vectors at different time points. To capture these temporal correlations, the feature correction layer employs time series analysis to model and analyze feature vectors at different time points. Specifically, the algorithm first represents the feature vector at each time point as a point in a high-dimensional space, and then constructs a time series model of the feature vectors by calculating the distance and similarity between these points.

[0084] Based on this time series model, the algorithm can predict eigenvector values ​​at future time points and compare them with actual observations to detect any abnormal changes. If abnormal changes are detected, the algorithm automatically adjusts the bias compensation data to adapt to these changes. Furthermore, time series analysis can help identify lag effects and causal relationships between eigenvectors, further improving the accuracy and effectiveness of bias compensation.

[0085] In addition to considering the temporal dimension, the feature correction layer also considers the spatial dimension. Due to the complexity of the maxillofacial structure, there may be spatial correlations between feature vectors from different regions. To capture these spatial correlations, the feature correction layer employs spatial statistical analysis methods to model and analyze the feature vectors from different regions. Specifically, the algorithm first divides the maxillofacial region into multiple sub-regions, and then constructs a spatial correlation model of feature vectors by calculating the spatial distance and similarity between these sub-regions.

[0086] Based on this spatial correlation model, the algorithm can identify regions with similar feature patterns and integrate and analyze the feature vectors of these regions. Furthermore, spatial statistical analysis can help detect spatial heterogeneity and spatial autocorrelation between feature vectors, further improving the accuracy and effectiveness of bias compensation.

[0087] After generating the bias-compensated data, the feature correction layer further validates and optimizes it. Specifically, the algorithm first applies the bias-compensated data to the feature fusion process, and then verifies the effectiveness of the bias-compensated data by comparing the fusion results with the evaluation results of clinical experts. If a significant difference is found between the fusion results and the evaluation results of clinical experts, the algorithm automatically adjusts the parameters of the bias-compensated data to improve the accuracy and reliability of the fusion results.

[0088] Furthermore, the feature correction layer employs cross-validation and model selection techniques to compare and evaluate different bias compensation methods and parameters to select the optimal bias compensation scheme. In this way, the feature correction layer can continuously optimize the generation process of bias compensation data, improving its accuracy and effectiveness, thereby providing a more reliable basis for generating subsequent follow-up evaluation instructions.

[0089] Example 3:

[0090] This embodiment describes the specific processing flow of the image preprocessing module. This module, through operations such as anatomical region division, data standardization, dynamic matching, and steady-state optimization, achieves standardized processing and noise filtering of the raw image data stream, providing standardized input data for subsequent analysis. The specific implementation method is as follows:

[0091] The input data for the image preprocessing module consists of the first spatial vector, the second correlation matrix, and the third mapping sequence from the initial postoperative data, corresponding to three-dimensional suture feature data, occlusal function feature data, and soft tissue distribution feature data, respectively. The processing first involves spatially dividing the multi-source data based on preset anatomical regions. These preset anatomical regions are constructed according to maxillofacial anatomy standards. For example, the jawbone structure is divided into sub-regions such as the maxillary body, maxillary alveolar process, mandibular ramus, and mandibular body; the occlusal function region is divided into the anterior occlusal region, posterior occlusal region, left occlusal contact point group, and right occlusal contact point group; and the soft tissue distribution region is divided into buccal soft tissue, lip soft tissue, and soft tissue around the temporomandibular joint. Each anatomical region corresponds to a specific range of physiological parameter monitoring, such as bone displacement parameters of the maxillary body, occlusal force transmission parameters of the anterior occlusal region, and thickness change parameters of the buccal soft tissue.

[0092] Based on spatial partitioning, the module performs isogradient partitioning on each feature data. For the first spatial vector corresponding to the three-dimensional suture feature data, the system divides the sutures within each anatomical interval into several gradient layers according to the direction and distance of bone displacement, for example, using a gradient of 0.2 mm for each displacement distance, generating standardized suture data. For the second correlation matrix corresponding to the occlusal function feature data, gradient partitioning is performed according to the magnitude of the occlusal force transmission intensity (e.g., 0-50N, 50-100N, etc.) to generate standardized occlusal data. For the third mapping sequence corresponding to the soft tissue distribution feature data, gradient stratification is performed based on the variation amplitude of parameters such as soft tissue thickness and blood flow to generate standardized soft tissue data. The purpose of isogradient partitioning is to convert continuous physiological parameters into discrete feature levels, facilitating subsequent data matching and analysis.

[0093] After spatial partitioning and gradient stratification, the module performs dynamic parameter alignment on the standardized suture and occlusal data. Dynamic parameter alignment is based on time-series synchronous calibration logic. A two-dimensional mapping table of "timestamp-anatomical position" is established to temporally correlate suture displacement events with changes in occlusal function parameters. For example, when a 3D reconstructed image shows a 0.3mm anterior migration of the maxillary bone on postoperative day 3, the system automatically retrieves the pressure change values ​​at the anterior occlusal contact points from the concurrent occlusal function monitoring data. An interpolation algorithm fills in missing values ​​in the data acquisition interval, ensuring a one-to-one correspondence between the two on the time axis. This process employs a dynamic time warping (DTW) algorithm to eliminate temporal deviations caused by differences in sampling frequencies from different monitoring devices, creating a comparable synchronous sequence of suture displacement and dynamic changes in occlusal function.

[0094] For standardized soft tissue data, the module employs a fixed anatomical window mechanism for steady-state optimization. The fixed anatomical window is a 5mm cube-shaped analysis area centered on the surgical incision, encompassing the surrounding soft tissue region (the window size can be adjusted according to anatomical structures). Soft tissue parameters within the window (such as grayscale values ​​in ultrasound images and blood flow parameters from laser Doppler flowmeters) are continuously monitored. Steady-state optimization is achieved through a low-pass filtering algorithm, setting a cutoff frequency of 0.1Hz to filter high-frequency noise and retain low-frequency signals reflecting the long-term healing trend of soft tissue. For example, in the early postoperative period, soft tissue may experience a short-term surge in blood flow due to inflammation; filtering can extract the baseline trend of blood flow changes, avoiding interference with the assessment of suture healing status. The processed soft tissue data generates a third calibration sequence, which includes the mean, variance, and trend characteristic values ​​of the parameters for each anatomical window.

[0095] In addition, the image preprocessing module also needs to perform cross-parameter correlation analysis on the calibrated feature data. Taking the first distribution label (bone block displacement label) and the second quantification label (suture healing density) in the first calibration vector as an example, the system calculates the healing offset coefficient of the two within the historical follow-up period. The healing offset coefficient is defined as the Pearson correlation coefficient between bone block displacement distance and suture healing density within the same anatomical interval, reflecting the synergy between changes in bone structure and the healing process. When the correlation coefficient of a certain area is lower than a preset threshold (such as 0.5), it indicates that there may be delayed healing or abnormal displacement, which needs to be monitored closely during the real-time follow-up stage.

[0096] Based on historical offset coefficients and real-time displacement parameters, the module predicts the expected distribution value of suture healing density. Specifically, a linear regression model is used to fit the functional relationship between displacement distance and healing density in historical data (e.g., healing density = α × displacement distance + β). The real-time displacement parameters are then substituted into the model to calculate the expected value. If the difference between the real-time monitored healing density value and the expected value exceeds 15%, target correction compensation data is generated to adjust the weight parameters of the feature vector for that region. For example, when a mandibular segment shifts by 0.8 mm, the model predicts a healing density of 0.6 g / cm³. If the measured value is 0.4 g / cm³, a compensation factor is generated to weight and correct the feature vector for that region, highlighting any abnormal signals in the healing density parameter.

[0097] In the data output stage, the module converts the processed standardized data into a unified data format: the first calibration vector integrates the spatial location, displacement gradient, and healing density characteristics of the sutures, represented as a three-dimensional coordinate vector; the second calibration matrix records the spatial distribution and force transmission synergy of the occlusal contact points, using an N×N matrix (N being the number of occlusal contact points) to store the correlation strength between each point; the third calibration sequence arranges the steady-state characteristic values ​​of soft tissue parameters in chronological order, forming a one-dimensional time-series array. These three types of data are correlated through timestamp indexes to ensure the spatiotemporal consistency of the multimodal data.

[0098] Throughout the processing, the module strictly adheres to anatomical zoning principles to ensure accurate spatial localization. Gradient partitioning and parameter alignment achieve dimensional uniformity and temporal synchronization of multi-source data. Fixed-window filtering and cross-parameter prediction enhance data stability and clinical interpretability. For example, when analyzing patients post-operatively fractured zygomatic bone complexes, the module dynamically matches suture displacement data in the zygomatic alveolar ridge region with force transmission data at the ipsilateral posterior occlusal contact point. Simultaneously, it assesses the impact of local blood supply on healing through blood flow parameters in the cheek soft tissue window. Ultimately, it outputs multidimensional calibration data encompassing spatial location, functional status, and tissue physiology, providing high-quality input for subsequent feature generation layers.

[0099] Example 4:

[0100] This embodiment focuses on the feature tracing and parameter superposition process of the feature correction layer. The feature correction layer includes a feature tracing unit, a parameter superposition unit, and a noise suppression unit. Through deep analysis and dynamic correction of the feature fusion data, it generates deviation compensation data that reflects the dynamic mapping relationship of the healing process. The specific implementation method is as follows:

[0101] The core function of the feature tracing unit is to trace the healing features of each feature vector in the feature fusion data to extract the corresponding healing pathway. The feature fusion data contains multi-dimensional information from three-dimensional sutures, occlusal function, and soft tissue distribution, such as the bone-occlusal function fusion tensor formed by fusing the first spatial vector and the second correlation matrix, and the soft tissue-bone healing fusion matrix formed by fusing the third mapping sequence and the first spatial vector. For each feature vector (such as the bone displacement feature vector in the fusion tensor and the soft tissue thickness feature vector in the fusion matrix), the feature tracing unit identifies the key feature chains directly related to the healing process through time series analysis and spatial path tracing.

[0102] Taking the three-dimensional suture feature vector as an example, the feature tracing unit first performs time-series expansion on the bone block displacement parameters in the first spatial vector, extracting the displacement trajectory of each anatomical sub-region (such as the maxillary anterior tooth region bone segment) at different follow-up time points (e.g., 0.2 mm displacement on postoperative day 1 and 0.8 mm cumulative displacement on postoperative day 7). Combined with the suture healing parameters (second quantitative label), the system constructs a "displacement-healing" correlation map, tracking the healing density change nodes on the displacement path (e.g., the healing density at point A in the displacement path reaches 0.4 g / cm³ on postoperative day 3 and 0.7 g / cm³ at point B on postoperative day 10), forming a healing conduction chain from the initial position of the bone block to the current position. This conduction chain contains a series of ordered (displacement position, healing density, time) triples.

[0103] For occlusal function feature vectors (such as the occlusal force transmission vector in the second correlation matrix), the feature tracing unit analyzes the transfer trajectory of occlusal contact points over time. For example, in the early postoperative period after orthognathic surgery, due to the unstable bone displacement, occlusal contact points may be concentrated in the molar region; as the sutures heal, the contact points gradually shift to the anterior region. The system generates a healing transmission chain of occlusal function by identifying the spatial migration path of occlusal contact points (such as the transfer from the mandibular first molar to the mandibular central incisor) and the corresponding occlusal force values ​​at time points (such as the occlusal force in the molar region decreasing from 80N to 50N, and the occlusal force in the anterior region increasing from 20N to 60N). This transmission chain reflects the recovery process of occlusal function from a compensatory state to a normal state.

[0104] The healing feature tracing of soft tissue distribution feature vectors is based on the spatiotemporal changes of the third mapping sequence. For example, the soft tissue thickness parameter around the surgical incision thickens due to edema in the early postoperative period and then gradually subsides. The feature tracing unit extracts the soft tissue thickness change curve within a 5mm range around the incision (e.g., the thickness increases by 1.5mm on postoperative day 1 and returns to the preoperative level on day 14), and combines it with the blood flow parameters in this area (e.g., the peak blood flow on postoperative day 3 is 2.3 times the baseline value, and it decreases to 1.2 times the baseline value on day 7) to construct a soft tissue healing conduction chain. This chain records the temporal characteristics of edema subsidence and blood revascularization.

[0105] The function of the parameter overlay unit is to superimpose the healing conduction chain extracted from each feature vector with the corresponding feature fusion data using anatomical parameters to generate bias compensation data. Anatomical parameter overlay is not a simple data splicing, but rather a cross-integration of dynamic features in the conduction chain and static features in the fusion data based on the correlation of anatomical structures. For example, in the bone-occlusal function fusion tensor, the bone displacement nodes in the healing conduction chain (e.g., anterior tooth bone segment displacement of 0.6 mm) are superimposed with the corresponding occlusal contact point parameters in the fusion tensor (anterior tooth occlusal force of 55 N) to generate a new composite parameter—the "displacement-occlusal force synergistic value," which is equal to the product of the displacement distance and the occlusal force (0.6 mm × 55 N = 33 N). (mm), reflecting the load state of the engagement function under this displacement.

[0106] For the soft tissue-bone healing fusion matrix, the parameter overlay unit overlays the peak blood flow time point (postoperative day 3) in the soft tissue healing conduction chain with the healing density parameter (0.3 g / cm³) of the corresponding suture region in the fusion matrix to generate a "blood supply-healing correlation value." This value, after normalization, reflects the degree of influence of blood supply status on healing speed. Furthermore, for the occlusal-soft tissue stress fusion vector, the system overlays the contact point transfer node in the occlusal conduction chain (e.g., the contact point transfers from the premolar region to the molar region on postoperative day 7) with the soft tissue stress parameter (soft tissue stress in the molar region increases by 15 kPa) to generate an "occlusal stress transfer amount," used to assess the mechanical impact of changes in occlusal function on soft tissue.

[0107] The noise suppression unit performs phase shift effect elimination processing on the superimposed bias compensation data. Phase shift effect originates from time delays or spatial registration errors during multimodal data acquisition. For example, a 5-second time difference between CT image acquisition and occlusal force sensor data acquisition causes a temporal misalignment between suture displacement parameters and occlusal data. The noise suppression unit uses a cross-correlation algorithm to calculate the time delay of different modal data and performs time shifting on the lagging data through linear interpolation to achieve phase alignment. For spatial registration errors (such as coordinate deviations between the 3D suture model and soft tissue ultrasound images), the system uses the Iterative Closest Point (ICP) algorithm to perform spatial transformation, aligning the coordinate systems of different modalities to the maxillofacial anatomical reference coordinate system, thus eliminating spatial phase deviations.

[0108] Throughout the feature correction layer's processing flow, the system follows a logical framework of "anatomical structural constraints - temporal sequence correlation - dynamic parameter overlay." For example, when analyzing patients after sagittal split mandibular surgery, the feature tracing unit first tracks the posterior displacement trajectory of the mandibular ramus segment and the corresponding changes in suture healing density, generating a suture healing conduction chain. Simultaneously, it analyzes the thickness changes of the soft tissues around the ipsilateral temporomandibular joint and the tension changes of the masseter muscle, generating a soft tissue conduction chain. The parameter overlay unit overlays the posterior displacement of the bone segment (1.2 mm) with the masseter muscle tension value (45 N) to form a deviation compensation parameter reflecting the skeletal-muscle biomechanical balance. The noise suppression unit eliminates phase errors caused by differences in device sampling frequencies by calibrating the timestamps of CT and electromyographic signals, ensuring the spatiotemporal consistency of the compensation parameters.

[0109] The output data of the feature correction layer—bias compensation data—is stored in the form of multidimensional tensors. Each tensor element corresponds to a feature correction value at a specific anatomical location and time point. For example, a three-dimensional tensor element (x, y, z, t) represents the bias compensation value at anatomical coordinates (x, y, z) and time t. This value is generated after superposition through conduction chains and noise removal, and is used to adjust the weight coefficients of each parameter in the feature fusion data. For example, when the bone suture healing density in a certain area is lower than the expected value, the bias compensation data increases the weight of this parameter in feature fusion, causing the system to prioritize this abnormal area in subsequent evaluations.

[0110] Example 5:

[0111] This embodiment describes the specific structure and working principle of the instruction triggering layer. The instruction triggering layer includes a multi-dimensional feature coordination unit, a dynamic weight optimization unit, and an abnormal state recognition unit. Through the collaboration of multiple units, the process of generating follow-up evaluation instructions from feature data is realized. The specific implementation method is as follows:

[0112] The multidimensional feature collaboration unit consists of multiple instruction decision nodes, each corresponding to a clinical assessment logic. These nodes are connected to feature vectors in the bias compensation data and feature fusion data through configuration parameters. For example, a node might correlate the healing density of three-dimensional sutures (from the first calibration vector) with the uniformity of occlusal force transmission (from the second calibration matrix), with its configuration parameters being preset density thresholds (e.g., 0.6 g / cm³) and uniformity thresholds (e.g., 85%). When the healing density of sutures in a certain anatomical region is below the threshold and the uniformity of occlusal force is also below the threshold, the node triggers an assessment instruction of "delayed healing and functional abnormality." The node's assessment logic is based on clinical guidelines, such as the correlation criteria between bone healing and functional recovery in the "Postoperative Follow-up Standards for Orthognathic Surgery."

[0113] The dynamic weight optimization unit iteratively optimizes the configuration parameters using a dynamic weight adjustment algorithm to minimize the error between the follow-up evaluation instructions and the actual recovery trajectory. This algorithm employs gradient descent, and the objective function is defined as:

[0114]

[0115] in, This is the error value. For the sample size, The actual recovery status vector labeled by clinical experts. This is the evaluation instruction vector generated by the system. The algorithm updates the configuration parameters (such as thresholds and weight coefficients) of each instruction decision node through backpropagation, so that the objective function... Gradually reduce the density. For example, for the "suture healing density-occlusal force uniformity" node, the algorithm may adjust the density threshold from 0.6 g / cm³ to 0.55 g / cm³ to adapt to the healing characteristics of a specific patient group.

[0116] The abnormal state identification unit locates abnormal healing based on deviation compensation data and feature fusion data. The system first establishes a feature template for normal healing patterns, generated by analyzing the mean and standard deviation of a large amount of historical postoperative data. For real-time input feature data, the unit calculates the Mahalanobis distance between the feature vector of each anatomical sub-region and the template.

[0117]

[0118] in, For real-time feature vectors, The template mean vector, This is the template covariance matrix. When the Mahalanobis distance of a certain region exceeds a preset critical value (e.g., ...), ... When abnormal healing is detected in the area, the system accurately locates the specific anatomical position (such as the anterior wall suture zone of the maxilla or the right posterior tooth occlusal contact point group) and generates a follow-up assessment instruction containing information on the type of abnormality (such as delayed healing or stress concentration) and its location.

[0119] In the instruction generation process, the node outputs of the multi-dimensional feature collaboration unit are combined through logic gates (such as "AND" and "OR") to form preliminary assessment conclusions. For example, the node "bone suture healing density is below the threshold" and the node "soft tissue edema lasts for more than 2 weeks" are combined through "AND" logic to generate the instruction "there may be infection or blood supply disorder". The dynamic weight optimization unit adjusts the triggering conditions of the logic gates based on historical assessment data, for example, increasing the triggering probability of "OR" logic from 40% to 60% to accommodate high-risk patient groups.

[0120] The localization accuracy of the abnormal state identification unit depends on the spatial resolution of the feature data. For example, in three-dimensional suture data, the system divides the jawbone into voxel units of 0.5mm × 0.5mm × 0.5mm, with each unit corresponding to a feature vector. By calculating the Mahalanobis distance on a voxel-by-voxel basis, millimeter-level localization of abnormal areas is achieved. For occlusal function data, each occlusal contact point is used as the smallest analysis unit, and occlusal trauma risk areas are identified by abnormal force transmission patterns at the contact points (such as a sudden 50% increase in unilateral occlusal force).

[0121] The output commands of the command trigger layer adopt a structured format, containing the following fields: anatomical location (e.g., "posterior border of the mandibular ramus"), abnormality type (e.g., "delayed suture healing"), recommended measures (e.g., "increase the frequency of CT follow-up examinations"), and timestamp (e.g., "14 days postoperatively"). For example, when the system detects that the suture healing density of the posterior border of the mandibular ramus is lower than 80% of the template mean for two consecutive weeks, and the blood flow of the masseter muscle in the corresponding area is higher than 1.5 times the mean, the command is generated as follows: "Anatomical location: posterior border of the mandibular ramus; Abnormality type: delayed suture healing with soft tissue inflammation; Recommended measures: perform local ultrasound examination and adjust antibiotic medication; Timestamp: May 22, 2025, 14:30".

[0122] The entire instruction triggering layer integrates data-driven algorithm optimization with clinical knowledge-based rule constraints. The dynamic weight optimization unit enhances the system's adaptability to individual differences through machine learning, while the multi-dimensional feature collaboration unit and abnormal state recognition unit ensure that the evaluation logic conforms to clinical diagnostic and treatment guidelines. For example, when dealing with patients with temporomandibular joint disorder, the system optimizes the weights of the "joint area suture displacement - condylar movement trajectory" node, prioritizing characteristic data of the joint region and avoiding potential missed diagnoses in traditional assessments.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A follow-up system for orthognathic surgery for maxillofacial deformities, characterized in that, include: An image analysis model is used to perform multi-dimensional feature fusion processing on postoperative initial data. The postoperative initial data includes a first spatial vector corresponding to three-dimensional suture feature data, a second correlation matrix corresponding to occlusal function feature data, and a third mapping sequence corresponding to soft tissue distribution feature data. The three-dimensional suture feature data includes a first distribution label generated by bone block displacement parameters and a second quantization label generated by suture healing parameters. The multi-dimensional feature fusion processing includes spatial coordinate normalization processing of the first spatial vector, time series alignment processing of the second correlation matrix, and anatomical region annotation processing of the third mapping sequence. The three types of processed data are input into a feature fusion network, and cross-modal fusion features are extracted through layer-by-layer feature cross-interaction to form a multi-dimensional fusion feature set. A follow-up assessment model is used to perform dynamic interference suppression processing on real-time physiological data streams and input them into a feature generation layer for recovery trajectory simulation. Follow-up assessment instructions are generated based on the output of the feature generation layer. The dynamic interference suppression processing includes using an adaptive filtering algorithm to suppress high-frequency noise and removing abnormal fluctuation signals through joint time-domain and frequency-domain analysis. The recovery trajectory simulation includes matching the processed real-time physiological data stream with corresponding stage data of multiple types of historical postoperative data, extracting the characteristic change trends of each time node, and generating a postoperative recovery prediction trajectory through time-series prediction. The feature generation layer includes an image preprocessing module and a dynamic feature generation module. The image preprocessing module is used to perform anatomical interval segmentation and invalid signal filtering on the original image data stream. The dynamic feature generation module is obtained by joint training based on multiple types of historical postoperative data and real-time physiological parameters. The dynamic feature generation module is used to construct a neural network architecture that includes a feature parsing layer, a feature correction layer and an instruction triggering layer. The feature parsing layer is used to perform anatomical feature association processing on different feature vectors in the original image data stream to generate feature fusion data; the feature correction layer is used to model the dynamic mapping relationship between the feature fusion data corresponding to each feature vector to generate deviation compensation data; the instruction triggering layer is used to perform multi-dimensional feature fusion based on deviation compensation data and feature fusion data to generate follow-up evaluation instructions. The process of modeling the dynamic mapping relationship between the feature fusion data corresponding to each feature vector and generating deviation compensation data includes: The bone suture segmentation algorithm is used to identify key healing nodes in the feature fusion data, and the correction compensation mapping corresponding to each feature vector is determined based on the physiological parameter type corresponding to each key healing node. Calculate the offset coefficient between healing nodes with the same parameters in the correction compensation map corresponding to any two feature vectors, and generate deviation compensation data between the two feature vectors based on the offset coefficient; The calculation of the offset coefficient between healing nodes with the same parameters in the corrected compensation mapping corresponding to any two feature vectors includes: When the number of healing nodes in the correction compensation mapping corresponding to any two feature vectors is inconsistent, virtual healing point interpolation is performed based on the physiological parameters corresponding to the terminal healing node in the one with fewer healing nodes, and the offset coefficient between healing nodes with the same parameters is calculated based on the interpolated data. The feature correction layer specifically includes: The feature tracing unit is used to perform healing feature tracing on each feature vector in the feature fusion data, so as to extract the corresponding healing conduction chain from each feature vector; The parameter overlay unit is used to overlay the healing conduction chain extracted from each feature vector with the corresponding feature fusion data to generate bias compensation data.

2. The follow-up system for orthognathic surgery for maxillofacial deformities as described in claim 1, characterized in that, The image preprocessing module is specifically used for: Based on the preset anatomical interval, the first spatial vector, the second correlation matrix, and the third mapping sequence are divided into equal gradients to generate standardized suture data, standardized occlusion data, and standardized soft tissue data. The standardized suture data and standardized occlusion data are matched in real time using a dynamic parameter alignment method, and the standardized soft tissue data are steadily optimized using a fixed anatomical window mechanism, outputting a first calibration vector, a second calibration matrix and a third calibration sequence; wherein, the first calibration vector includes a calibrated first distribution label and a calibrated second quantization label.

3. The follow-up system for orthognathic surgery for maxillofacial deformities as described in claim 2, characterized in that, The image preprocessing module is also used for: Calculate the healing offset coefficient between the calibrated first distribution label and the calibrated second quantization label during the historical follow-up period; The expected distribution value of the calibrated second quantization label in the real-time follow-up phase is predicted based on the healing offset coefficient and the parameter value of the calibrated first distribution label in the real-time follow-up phase. Target correction compensation data is generated based on the calibrated second quantization label and its expected distribution value, and the feature vector corresponding to the target correction compensation data is used as the first calibration vector.

4. The follow-up system for orthognathic surgery for maxillofacial deformities as described in claim 1, characterized in that, The feature correction layer further includes: The noise suppression unit is used to perform phase offset effect elimination processing on the deviation compensation data.

5. The follow-up system for orthognathic surgery for maxillofacial deformities as described in claim 1, characterized in that, The instruction triggering layer specifically includes: The multi-dimensional feature collaboration unit contains multiple instruction decision nodes, and each instruction decision node is connected to each feature vector in the deviation compensation data and feature fusion data through configuration parameters. The dynamic weight optimization unit is used to iteratively optimize the configuration parameters through a dynamic weight adjustment algorithm to minimize the error between the follow-up evaluation instructions and the actual recovery trajectory. An abnormal state identification unit is used to locate abnormal healing based on the deviation compensation data and feature fusion data, and generate follow-up evaluation instructions.

6. A method for follow-up after orthognathic surgery for maxillofacial deformities, applied to the follow-up system for orthognathic surgery for maxillofacial deformities as described in any one of claims 1 to 5, characterized in that, Includes the following steps: The initial postoperative data is processed by multi-dimensional feature fusion using an image analysis model. The initial postoperative data includes a first spatial vector corresponding to the three-dimensional suture feature data, a second correlation matrix corresponding to the occlusal function feature data, and a third mapping sequence corresponding to the soft tissue distribution feature data. The three-dimensional suture feature data includes a first distribution label generated by bone block displacement parameters and a second quantitative label generated by suture healing parameters. The real-time physiological data stream is dynamically suppressed by a follow-up assessment model, and the processed data stream is input into the feature generation layer to simulate the recovery trajectory. The image preprocessing module in the feature generation layer is used to perform anatomical region segmentation and invalid signal filtering on the raw image data stream; A dynamic feature generation module, which is jointly trained from multiple types of historical postoperative data and real-time physiological parameters, is used to simulate the recovery trajectory. The dynamic feature generation module includes a feature parsing layer, a feature correction layer, and an instruction triggering layer connected in sequence. Based on the output of the feature generation layer, the follow-up evaluation model generates follow-up evaluation instructions.

Images