Method for judging jaw malformation based on facial soft tissue feature points

Abstract

The invention discloses a method for judging jaw malformation based on facial soft tissue feature points, which comprises the following steps of: automatically positioning key feature points in a side face optical image by using a pre-trained neural network model, establishing a coordinate system, correcting the deviation of a head posture in the image, and pre-diagnosing the jaw malformation condition; in the model training stage, a face optical image and an X-ray image which are paired are used, common feature points are extracted to register soft tissue and a bony structure, a deep neural network is used for learning the space mapping relation of the face optical image and the X-ray image, a coordinate system error introduced by an optical photo posture is calculated, and a head position normalization neural network model is formed; in the application, only a single face optical image is input to the model, the pre-training model can automatically identify the feature points, posture calibration and space standardization are completed, and automatic classification and auxiliary diagnosis results of jaw malformation are output according to the posture calibration and the space standardization.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, specifically to a method for determining jaw deformities based on facial soft tissue feature points. Background Technology

[0002] Jaw deformities are a common oral and maxillofacial disease that not only affects a patient's appearance but can also cause functional impairments in chewing, swallowing, and breathing, and even impact growth, development, and mental health. Therefore, timely and accurate diagnosis of jaw deformities is crucial for treatment and prevention. Currently, the clinical diagnosis of jaw deformities primarily utilizes cephalometric techniques (lateral cephalometric radiographs). This method involves taking images of the patient using specialized radiographic equipment, with medical professionals locating and measuring key points on the cephalometric film. However, this traditional method has the following significant drawbacks:

[0003] (1) Radiation issues: Cephalometric technique uses X-rays for imaging. Long-term or frequent use may cause radiation damage to patients, especially for adolescents who are in the growth and development stage, the radiation risk is particularly noteworthy.

[0004] (2) Reliance on professional equipment and personnel: Traditional cephalometric methods require a high level of expertise from both the equipment and the operator, which limits the application of the technology to specialized medical institutions and makes it difficult to popularize in primary hospitals and remote areas, thus hindering large-scale and efficient disease screening.

[0005] In recent years, with the development of artificial intelligence and computer vision technologies, some studies have proposed using optical images (such as photographs) for preliminary identification and diagnosis of jaw deformities. While this method avoids radiation exposure issues, traditional optical facial recognition methods are easily affected by patient posture, lighting conditions, and differences in imaging equipment, resulting in low diagnostic accuracy and failing to meet clinical application needs. Therefore, how to improve diagnostic accuracy and reduce the influence of posture and equipment in optical image diagnostic methods to achieve rapid and accurate screening of jaw deformities has become a key problem urgently needing to be solved by those skilled in the art.

[0006] In view of this, the present invention proposes a method for judging jaw deformities based on facial soft tissue feature points. Summary of the Invention

[0007] The purpose of this invention is to provide a method for judging jaw deformities based on facial soft tissue feature points. By accurately locating facial feature points and establishing a standardized facial coordinate system, the diagnostic errors caused by differences in shooting posture and equipment are significantly reduced, effectively improving the accuracy and universality of jaw deformity diagnosis.

[0008] In a first aspect, the present invention provides a method for determining jawbone deformities based on facial soft tissue feature points, comprising the following steps:

[0009] S101. Obtain an optical image of the region to be analyzed, automatically detect soft tissue feature points in the optical image based on a pre-trained neural network model, and extract the initial coordinates of the soft tissue feature points.

[0010] S102. Based on the initial coordinates, establish a standardized soft tissue coordinate system and determine the spatial measurement data between multiple facial soft tissue feature points. The spatial measurement data includes feature point distances and spatial angle vectors.

[0011] S103: Input the spatial measurement data into the pre-trained deep learning neural network, and automatically calculate the coordinate correction values ​​of each soft tissue feature point based on the head position normalization and bony reference calibration module of the deep learning neural network, so as to realize the pose normalization and bony reference alignment of the soft tissue feature points in the input optical image.

[0012] S104. Update the initial coordinates according to the coordinate correction value to obtain the calibrated facial soft tissue feature point coordinates and spatial measurement data.

[0013] S105: Input the calibrated spatial measurement data into the classification module of the pre-trained deep learning neural network model to achieve automatic classification and assisted diagnosis of the jaw deformity category of the target object.

[0014] As a preferred embodiment of the first aspect of the present invention, obtaining the initial coordinates of the soft tissue feature points in the region to be analyzed includes:

[0015] The facial region of the optical image of the area to be analyzed is standardized.

[0016] The standardized image is trained by a machine learning model and decoded from an integer heatmap using feature maps. The integer coordinates with the highest probability values ​​are extracted for preliminary localization.

[0017] In the region surrounding the initially located integer coordinates, the initial coordinates of the soft tissue feature points in the region to be analyzed are determined through local classification regression calculation.

[0018] As a preferred embodiment of the first aspect of the present invention, the construction of the standardized soft tissue coordinate system specifically includes:

[0019] The initial coordinates of the soft tissue feature points in the region to be analyzed are organized into a feature point set; using the custom soft tissue feature points as the reference feature points, a horizontal baseline is selected to construct a standardized two-dimensional coordinate system for facial soft tissue.

[0020] A horizontal baseline is defined based on the origin of the coordinate system. The horizontal baseline is rotated to the horizontal axis of the coordinate system, and all feature points are rotated accordingly to obtain the coordinates of the rotated feature points. The facial length is defined as a standardized scale, and the initial coordinates of all soft tissue feature points are processed by standardization to obtain a coordinate system with uniform size.

[0021] After establishing a standardized coordinate system, spatial measurement data of facial soft tissue feature points are calculated. The spatial measurement data includes feature point distances and spatial angle vectors. The distance between facial soft tissue feature points is defined and marked as the feature point distance, and the spatial angle data formed by any three facial soft tissue feature points is marked as the spatial angle vector.

[0022] As a preferred technical solution of the first aspect of the present invention, based on the spatial measurement data of facial soft tissue feature points, all soft tissue feature point combinations are calculated one by one to form a spatial measurement dataset.

[0023] Based on multiple soft tissue feature points, and according to clinical diagnostic needs, several facial feature angles that are recognized as having diagnostic value in the field of orthodontics are selected from a predefined set of angles for calculation to constitute spatial measurement data of facial soft tissue feature points.

[0024] As a preferred embodiment of the first aspect of the present invention, the construction logic of the deep learning neural network is as follows:

[0025] During the model training phase, X-ray image data of the target object is acquired, and bony soft tissue feature points are marked in the X-ray image data. Key bony feature points include, but are not limited to, saddle point, nasal root point, maxillary anterior point, mandibular anterior point, infraorbital margin point, and ear canal point.

[0026] Based on the aforementioned bony feature points, an X-ray cephalometric coordinate system is constructed, and the corresponding clinical standard bony measurement angles are calculated.

[0027] Spatial alignment and mapping are performed between spatial measurement data obtained from optical images and corresponding bony measurement angles in X-ray image data.

[0028] Using paired facial optical images and X-ray image data, a spatial registration relationship is established based on shared feature points, and the coordinate system error caused by changes in head posture in the optical images is learned, including:

[0029] Overall rotational error caused by attitude change;

[0030] Overall translational error caused by changes in head position;

[0031] Scale deviation caused by different image acquisition conditions;

[0032] Local drift deviation of soft tissue feature points in a local area;

[0033] An error compensation mechanism is established based on coordinate system error to optimize the fitting accuracy of the standardized soft tissue coordinate system relative to the bone coordinate system, so as to achieve spatial uniformity and measurement consistency between different image modalities.

[0034] The aforementioned spatial registration relationship and coordinate system error compensation mechanism are embedded into the neural network structure to complete the pre-training process of the deep learning neural network based on multimodal data alignment.

[0035] As a preferred embodiment of the first aspect of the present invention, the calculation logic for the coordinate correction value is as follows:

[0036] Based on the input initial soft tissue feature point coordinates and the constructed standardized soft tissue coordinate system, spatial measurement data between each feature point is extracted;

[0037] By utilizing a pre-trained head position normalization and bony reference calibration module in a deep learning neural network, systematic errors between input spatial measurement data and the standard bony coordinate system are identified.

[0038] Based on the identified systematic and local drift errors, the network generates coordinate correction values ​​for each soft tissue feature point, wherein the correction values ​​are:

[0039] The coordinate correction values ​​of each feature point in the X and Y dimensions; based on the coordinate correction values, the feature point coordinates are mapped from the soft tissue reference frame of the optical image to a pose space consistent with the bony reference frame.

[0040] As a preferred embodiment of the first aspect of the present invention, the pre-training logic of the deep learning neural network is as follows:

[0041] The deep learning neural network includes a training dataset and a test dataset, which include a large number of paired optical images and X-ray image data.

[0042] During the training phase, optical images with different head poses and soft tissue feature point distributions are used as input factors. The paired optical images and X-ray image data are optimized and trained in the deep learning neural network to automatically output coordinate correction values ​​consistent with the bony reference.

[0043] During the generation of coordinate correction values, the historical attitude error distribution of the input data is learned simultaneously, and the generated coordinate correction values ​​are dynamically optimized by combining the current input spatial measurement data.

[0044] As a preferred embodiment of the first aspect of the present invention, the logic for obtaining the calibrated facial soft tissue feature point coordinates and spatial measurement data is as follows:

[0045] Using the initial coordinates of facial soft tissue feature points as a reference, the coordinate positions of facial soft tissue feature points are automatically calibrated based on the coordinate correction values ​​generated by dynamic optimization.

[0046] The standardized coordinate system is updated based on the calibrated feature point coordinates, and the spatial measurement data between facial soft tissue feature points is recalculated to obtain calibrated spatial measurement data.

[0047] As a preferred embodiment of the first aspect of the present invention, the training process of the deep learning neural network classification module includes:

[0048] Using a dataset of historical jaw deformity patients, a feature vector with several feature angle values ​​is constructed as the input layer. The hidden layer uses the ReLU activation function, which includes the spatial angles of facial soft tissue feature points and known jaw classification information. The output layer uses the Softmax function to obtain the probability of each category.

[0049] The model is gradually optimized through iterative training, and the backpropagation algorithm is used to train the model until it stabilizes and converges, until the confidence probability value of the model's output classification reaches the preset classification accuracy.

[0050] As a preferred embodiment of the first aspect of the present invention, the iterative optimization of training the deep learning neural network specifically includes:

[0051] The error gradient is calculated using the initial classification results and known classification information;

[0052] The network weights are dynamically adjusted based on the error gradient, and the optimization is repeated multiple times until the classification error is less than the error threshold.

[0053] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0054] This invention automatically identifies facial feature points in optical images using a neural network model and utilizes X-ray cephalometric data as the gold standard for diagnosis, significantly improving the accuracy and reliability of jaw deformity diagnosis. By introducing artificial intelligence optimization fitting algorithms and deep learning neural networks, it achieves automatic identification of abnormal areas and automatic calibration of feature points, reducing the influence of subjective human factors, greatly improving diagnostic efficiency, adapting to individual differences among patients, and exhibiting good adaptability and stability. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0056] Figure 1 This is a flowchart of a method for judging jaw deformities based on facial soft tissue feature points according to the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0058] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0059] Example 1

[0060] Please see Figure 1 As shown, this embodiment provides a method for determining jawbone deformities based on facial soft tissue feature points, including the following steps:

[0061] S101. Obtain an optical image of the region to be analyzed, automatically detect soft tissue feature points in the optical image based on a pre-trained neural network model, and extract the initial coordinates of the soft tissue feature points.

[0062] It should be noted that the region to be analyzed is the target side facial region. Optical images of the target side facial region are extracted, and a neural network model (CNN) is used to automatically locate specific anatomical landmarks (such as the soft tissue nasal root point, subnasal point, and soft tissue anterior chin point) in the optical image to obtain an integer heatmap of facial soft tissue feature points. Based on the integer heatmap, the feature point coordinates of each location are calculated to obtain the initial coordinates of the soft tissue feature points in the region to be analyzed. In other words, a trained image processing algorithm is used as input to the optical image to extract the positions of soft tissue feature points, thereby obtaining a non-linear mapping relationship between the optical image and the soft tissue feature points. After network pre-training, inputting an optical image of any region to be analyzed will automatically output the initial coordinates of each soft tissue feature point.

[0063] The pre-training logic of the neural network model is as follows:

[0064] Acquire facial optical images and corresponding skull X-ray images of the region to be analyzed, mark identifiable anatomical feature points in both, and construct a common coordinate reference between soft tissue and bony structures;

[0065] A spatial registration model between optical images and X-ray images is established based on shared anatomical feature points, and the nonlinear mapping relationship between the two is learned.

[0066] By iteratively training a deep neural network, a head-position normalization model is constructed to correct coordinate system errors caused by shooting posture in optical images, and then embedded in the pre-trained neural network model.

[0067] Specifically, the training logic for a neural network model is as follows:

[0068] A neural network model (such as the Stacked Hourglass Network, HRNet, or the CHC Network proposed in this paper) is used to train the soft tissue feature point detection model. X-ray images are used as the gold standard to establish a bony reference coordinate system. The neural network is trained using a large number of paired facial optical images and X-ray images. During pre-training, it learns the spatial mapping and fitting rules between optical and bony images, completing the spatial alignment and calibration of soft tissue and bony images. Once trained, this knowledge is embedded in the neural network model parameters. In practical applications, the target lateral facial optical image can be automatically reconstructed using the neural network model to obtain the spatial structure of soft tissue feature points conforming to the standard head posture, achieving automatic calibration and efficient diagnosis without the need for X-ray images.

[0069] Input: A large number of side-view facial photos with labeled soft tissue feature points.

[0070] Output: A probability heatmap for each soft tissue feature point.

[0071] Training method: Based on the initial coordinates of manually labeled facial soft tissue feature points, a Gaussian distribution heatmap of each soft tissue feature point is constructed as a supervision signal; mean squared error (MSE) or cross-entropy loss function is used to optimize network parameters until the model converges stably.

[0072] The logic for obtaining the initial coordinates of the facial soft tissue feature points is as follows:

[0073] The optical image of the area to be analyzed is standardized into a standardized photograph with a uniform size and resolution (e.g., 256×256 pixels);

[0074] Input a standardized photo into a pre-trained neural network model. The model outputs an integer heatmap of multiple soft tissue feature points. Each integer heatmap has the same size as the input image. The value of each position in the integer heatmap represents the probability that the corresponding position is a soft tissue feature point.

[0075] The point with the highest probability value in the integer heatmap is represented by its integer coordinates. Integer coordinates are used to represent the approximate location, but using only integer coordinates cannot meet the accuracy requirements. Therefore, it is necessary to calculate the precise sub-pixel coordinates from the integer heatmap.

[0076] Using the integer coordinates of the feature points as the center, a local area of ​​3×3 or 5×5 pixels is extracted from the surrounding area; for the heat map of the extracted local area, Gaussian fitting or quadratic surface fitting is used to determine the coordinates of the sub-pixel level feature points.

[0077] The integer coordinates of the feature points are corrected to sub-pixel level feature point coordinates, and the sub-pixel level feature point coordinates are marked as the initial coordinates of the facial soft tissue feature points.

[0078] S102. Based on the initial coordinates of facial soft tissue feature points, establish a standardized soft tissue coordinate system and determine the spatial measurement data between multiple soft tissue feature points. The spatial measurement data includes feature point distances and spatial angle vectors.

[0079] Specifically, the construction logic of the standardized soft tissue coordinate system is as follows:

[0080] Organize the initial coordinates of the soft tissue feature points in the region to be analyzed into a feature point set; use the custom soft tissue feature points as the reference feature points and select the horizontal baseline;

[0081] It should be noted that: the custom soft tissue feature points are soft tissue feature points that the operator considers to have anatomical significance and strong stability. A horizontal baseline is defined based on the origin of the coordinate system. The horizontal baseline is rotated to the horizontal axis of the coordinate system, and all feature points are rotated accordingly to obtain the coordinates of the rotated feature points. The facial length is defined as a standardized scale, and the initial coordinates of all soft tissue feature points are processed to standardize the scale to obtain a coordinate system with uniform size.

[0082] After establishing a standardized coordinate system, spatial measurement data of facial soft tissue feature points are calculated. The spatial measurement data includes feature point distance and spatial angle vector. That is, any two soft tissue feature points are selected and the feature point distance between the two points is calculated, which represents the degree of proximity of the two soft tissue feature points in the standardized space and has clinical significance.

[0083] Three soft tissue feature points on the face are selected, with one of the soft tissue feature points as the vertex of the angle and the other two soft tissue feature points as the two endpoints of the angle. The three points together form a spatial angle vector, which reflects the spatial layout relationship between the three soft tissue feature points. This method is widely used in the diagnosis and analysis of jaw deformities.

[0084] Based on spatial measurement data of soft tissue feature points, all combinations of soft tissue feature points are calculated one by one to form a spatial measurement dataset. For example, based on multiple soft tissue feature points, several representative facial feature angles are selected for calculation according to clinical diagnostic needs to form spatial measurement data of facial soft tissue feature points. In this embodiment, 33 soft tissue feature points are preferably used as a basis, and 54 feature angles with clinical diagnostic significance among facial feature points may be selected for calculation. The calculated 54 feature angles constitute the final spatial measurement data of facial soft tissue feature points to meet the accuracy and reliability requirements of diagnosis. The specific number of feature points and feature angles does not limit the scope of protection of this invention. The spatial measurement dataset is used as input for MLP classification models or other diagnostic models for oral and maxillofacial deformity diagnosis.

[0085] S103: Input the spatial measurement data into the pre-trained deep learning neural network, and automatically calculate the coordinate correction values ​​of each soft tissue feature point based on the head position normalization and bony reference calibration module of the deep learning neural network, so as to realize the pose normalization and bony reference alignment of the soft tissue feature points in the input optical image.

[0086] It should be noted that X-ray imaging data, through X-rays penetrating the human head tissues, clearly reveals the position and spatial relationships of key bony structures such as the skull, maxilla, mandible, and teeth. This information directly reflects abnormalities in the shape, size, and position of the jawbone itself. However, in reality, there are significant individual differences in the jawbone and soft tissue structures of each patient. Therefore, it is impossible to define the normal range of soft tissue spatial relationships using preset standard values. In clinical applications, doctors typically use X-ray imaging data (such as ANB angle, SNB angle, and SNA angle) as the standard for diagnosing jawbone deformities. By accurately locating skeletal landmarks using X-ray imaging data, an X-ray coordinate system is constructed, and clinically standard measurement angles between bones are directly calculated. These angles have definitive diagnostic significance in clinical practice and can serve as reliable standard references. Artificial intelligence optimization fitting algorithms (such as least squares fitting, weighted least squares fitting, or nonlinear optimization algorithms) are used to fit the optical measurement data to determine its optimal spatial mapping relationship with the X-ray measurement data. Based on the residual value after fitting (i.e., the difference between the actual measured value of the optical feature point and the X-ray fitted prediction value), abnormal areas are automatically identified and determined. For example: X-ray measurement shows an ANB angle of 7° (deterministic ossification Class II); preliminary optical measurement yields a corresponding soft tissue angle of 3°; after fitting and mapping, a soft tissue angle deviation of 4° (significant abnormality) is found. AI automatically marks this area as an abnormal feature point region and guides subsequent correction or diagnosis.

[0087] Specifically, the construction logic of the deep learning neural network is as follows:

[0088] During the model training phase, X-ray image data of the target object is acquired, and bony soft tissue feature points are marked in the X-ray image data. Key bony feature points include, but are not limited to, saddle point, nasal root point, maxillary anterior point, mandibular anterior point, infraorbital margin point, and ear canal point.

[0089] It should be noted that: clear common anatomical landmarks are selected in optical images and X-ray image data, and these anatomical landmarks are used as the reference for spatial alignment;

[0090] It should be noted that anatomical landmarks include the nasal root point (Nasion, N point), the ear canal point (Porion, Po point), the infraorbital margin point (Orbitale, Or point), and the upper and lower lip points (references for the correspondence between soft tissue and dental points). These anatomical landmarks can be identified in X-ray images but cannot be located in optical images. Therefore, they need to be predicted and simulated using algorithms. A two-dimensional standard coordinate system for optical images is established, using the soft tissue nasal root point (N' point) as the origin, the eye-ear plane (FH plane, Porion-Orbitale connection line) predicted by the algorithm in the optical image as the X-axis, and the direction perpendicular to the X-axis as the Y-axis.

[0091] A two-dimensional standard coordinate system for X-ray image data is constructed with the nasal root point (N point) as the origin, the FH plane (Porion-Orbitale) marked by the X-ray image data as the X-axis, and the direction perpendicular to the FH plane as the Y-axis.

[0092] Extract the coordinates of common anatomical landmarks in two sets of coordinate systems, map the optical coordinate system onto the X-ray coordinate system through the spatial relationship between the common anatomical landmarks, and then align the feature points to their new coordinates in the X-ray coordinate system.

[0093] Algorithms (such as AI fitting algorithms or least squares optimization algorithms) are used to align the standardized coordinate systems of optical images and X-ray image data. This involves superimposing and matching them to make the positions of common feature points on the two images match as closely as possible. This can be understood as fitting the "outer surface" of facial soft tissue onto a standard template of the internal skeleton.

[0094] Based on the aforementioned bony feature points, an X-ray cephalometric coordinate system is constructed, and the corresponding clinical standard bony measurement angles are calculated.

[0095] Spatial alignment and mapping are performed between spatial measurement data obtained from optical images and corresponding bony measurement angles in X-ray image data.

[0096] Using paired facial optical images and X-ray image data, a spatial registration relationship is established based on shared feature points, and the coordinate system error caused by changes in head posture in the optical images is learned, including:

[0097] Overall rotational error caused by attitude change;

[0098] Overall translational error caused by changes in head position;

[0099] Scale deviation caused by different image acquisition conditions;

[0100] Local drift deviation of soft tissue feature points in a local area;

[0101] An error compensation mechanism is established based on coordinate system error to optimize the fitting accuracy of the standardized soft tissue coordinate system relative to the bone coordinate system, so as to achieve spatial uniformity and measurement consistency between different image modalities.

[0102] The aforementioned spatial registration relationship and coordinate system error compensation mechanism are embedded into the neural network structure to complete the pre-training process of the deep learning neural network based on multimodal data alignment.

[0103] Further explanation of the calculation logic for the coordinate correction value:

[0104] Based on the input initial soft tissue feature point coordinates and the constructed standardized soft tissue coordinate system, spatial measurement data between each feature point is extracted;

[0105] By utilizing a pre-trained head position normalization and bony reference calibration module in a deep learning neural network, systematic errors between input spatial measurement data and the standard bony coordinate system are identified.

[0106] Based on the identified systematic and local drift errors, the network generates coordinate correction values ​​for each soft tissue feature point, wherein the correction values ​​are:

[0107] The coordinate correction values ​​of each feature point in the X and Y dimensions; based on the coordinate correction values, the feature point coordinates are mapped from the soft tissue reference frame of the optical image to a pose space consistent with the bony reference frame.

[0108] After registration, the system automatically compares the differences between the same feature point in the two types of images (e.g., the difference between the lip position in optical data and the lip bone position in X-ray image data). Using an AI-optimized fitting algorithm, the coordinate systems of the optical image and X-ray image data are mapped and matched. By calculating the deviation between the optical and X-ray image data, the coordinate positions of the optical feature points are automatically calibrated. Using the X-ray image data as the "standard," the coordinates of the optical image feature points are automatically corrected and calibrated. For example, if the position of a feature point in the optical image significantly deviates from the standard position in the X-ray data, the position of that point is automatically corrected to the standard X-ray position, obtaining the initial coordinate correction value for the soft tissue feature point.

[0109] To further explain, the pre-training logic of the deep learning neural network is as follows:

[0110] The deep learning neural network includes a training dataset and a test dataset, which include a large number of paired optical images and X-ray image data.

[0111] During the training phase, optical images with different head poses and soft tissue feature point distributions are used as input factors. The paired optical images and X-ray image data are optimized and trained in the deep learning neural network to automatically output coordinate correction values ​​consistent with the bony reference.

[0112] During the generation of coordinate correction values, the historical attitude error distribution of the input data is learned simultaneously, and the generated coordinate correction values ​​are dynamically optimized by combining the current input spatial measurement data.

[0113] S104. Update the initial coordinates according to the coordinate correction value to obtain the calibrated facial soft tissue feature point coordinates and spatial measurement data.

[0114] Specifically, the automatically calibrated data meets the requirements for classification and diagnosis, and there is no need to be overly concerned about whether it fully meets the ideal deviation convergence standard. The image registration model calls on the pose alignment parameters and distortion compensation experience learned in the pre-training stage to automatically calibrate the feature point data of the input image, completing the spatial standardization equivalent to the X-ray reference coordinate system.

[0115] Specifically, the logic for obtaining the calibrated facial soft tissue feature point coordinates and spatial measurement data is as follows:

[0116] Using the initial coordinates of facial soft tissue feature points as a reference, the coordinate positions of facial soft tissue feature points are automatically calibrated based on the coordinate correction values ​​generated by dynamic optimization.

[0117] The standardized coordinate system is updated based on the calibrated feature point coordinates, and the spatial measurement data between facial soft tissue feature points is recalculated to obtain calibrated spatial measurement data.

[0118] It should be noted that: after visualization through the spatial mapping map, the locations of abnormal feature points (such as feature points that are seriously deviated from their normal positions) are clearly identified: based on the mapping map, areas where the density of feature points is too high, too low, or significantly deviates from the expected position are identified; the identified areas are the abnormal areas that need to be adjusted, and the coordinate positions of the identified abnormal feature points are adjusted. Using the adjacent normal areas (feature points that are not marked as abnormal) as reference areas, the average position of these normal feature points is calculated, and the position of the abnormal feature points is moved towards this average position. The adjustment range is determined according to the degree of deviation of the feature points from the normal range.

[0119] Repeatedly update the standardized soft tissue coordinate system and reselect the originally set reference feature points (such as the nasal root point and key points on the eye-ear plane); recalculate the rotation angle of the baseline based on the adjusted feature point coordinates; rotate and scale normalize the coordinates of all feature points to obtain the updated standardized soft tissue coordinate system. The specific rotation and scale normalization formulas are the same as those used when initially constructing the standardized coordinate system.

[0120] By using the updated standardized soft tissue coordinate system, the spatial measurement data between feature points is recalculated: the distance between feature points and the spatial angle vector are recalculated. The new spatial measurement data will be more accurate and stable, effectively improving the accuracy of subsequent diagnosis.

[0121] After the above adjustment and update process, it can be executed repeatedly (e.g., 2-3 times), with each iteration further refining the position coordinates of the feature points; ultimately reaching a stable state of spatial measurement data and coordinates (the adjustment amplitude tends to 0 or is below a preset threshold). This yields calibrated facial soft tissue feature point coordinates and spatial measurement data; multiple iterations ensure improved accuracy of feature point coordinates and reduce measurement errors.

[0122] S105: Input the calibrated spatial measurement data into the classification module of the pre-trained deep learning neural network model to achieve automatic classification and assisted diagnosis of the jaw deformity category of the target object.

[0123] The spatial angles between multiple facial soft tissue feature points are determined using a standardized soft tissue coordinate system. These spatial angles are then input into a pre-trained deep learning neural network to calculate spatial angle vectors for selected spatial combination feature points, forming a spatial angle dataset. Assuming a total of 54 diagnostically significant spatial angles are selected, these angles collectively constitute the input data feature vector:

[0124] The training process of the deep learning neural network classification module includes:

[0125] Using a dataset of historical jaw deformity patients, a feature vector with several feature angle values ​​is constructed as the input layer. The hidden layer uses the ReLU activation function, which includes the spatial angles of facial soft tissue feature points and known jaw classification information. The output layer uses the Softmax function to obtain the probability of each category.

[0126] The model is gradually optimized through iterative training, and the backpropagation algorithm is used to train the model until it stabilizes and converges, until the confidence probability value of the model's output classification reaches the preset classification accuracy.

[0127] Collect a large number of case data with known diagnostic categories (Class I, Class II, Class III). For each case, obtain 54 spatial angular feature values ​​through the above method to form a training dataset. The input feature vector is a 54-dimensional vector, and the label is the diagnostic category (Class I, Class II, or Class III).

[0128] An MLP network is constructed with 54 neurons in the input layer, each corresponding to a spatial angle feature value; one or more hidden layers (e.g., 2 layers) are used, each with a certain number of neurons (e.g., 64, 128); the ReLU function is used as the activation function to improve the model's expressive power; the output layer outputs the number of classes (3 classes in this experiment: class I, class II, and class III), and the Softmax function is used to obtain the probability of each class.

[0129] The case data with known labels are input into the MLP for forward propagation. The loss between the predicted output and the true class is calculated. The network parameters are adjusted using the backpropagation algorithm. The process is iterated repeatedly until convergence (the loss function reaches the predetermined accuracy or the number of training iterations reaches the set threshold). After training, a pre-trained MLP model is obtained (the weight parameters are fixed).

[0130] For new cases, obtain their spatial angle vectors (54 feature angles); input the spatial angle vectors into the pre-trained MLP model; calculate the probability of each category (I, II, III); select the category with the highest probability as the preliminary prediction result of jaw deformity.

[0131] For example, if the model outputs the class probabilities as follows: Class I: 0.85, Class II: 0.10, Class III: 0.05, then the preliminary prediction result of jaw deformity is Class I.

[0132] The angle is recalculated based on the latest coordinates of the feature points in the standardized soft tissue coordinate system. For any three feature points, the angle is calculated one by one for all combinations of feature points. Finally, the updated angle feature vector is obtained (for example, still 54 angles): the output is the probability value of the three diagnostic categories (for example, the probability of Bone Category I is 0.92, Category II is 0.06, and Category III is 0.02).

[0133] The category with the highest probability is ultimately selected as the final classification result, providing a predicted category (Class I, Class II, or Class III); the confidence score (probability value) for each category further assists clinical decision-making. For example: classification result = bone type I probability 0.92, type II probability 0.06, type III probability 0.02.

[0134] This embodiment introduces a real bone reference standard based on X-ray data, solving the practical problem of difficulty in presetting the normal reference range of soft tissue in actual applications; it clarifies the AI ​​optimization fitting algorithm to realize the mapping and registration of optical data and X-ray data; and it emphasizes the automatic calibration of feature point positions and the data iterative optimization mechanism, realizing a closed-loop structure.

[0135] In summary, during the deep learning neural network training phase, this embodiment utilizes simultaneously acquired lateral facial optical images and corresponding X-ray images to extract common anatomical landmarks and construct soft tissue and bony coordinate systems. The spatial mapping between the two is then learned through a deep learning neural network. By using artificial intelligence to optimize the fitting algorithm, the neural network can automatically model the nonlinear transformation relationship between optical feature points and bony benchmarks. During the training process, the network learns from a large number of samples and solidifies the learning into the network parameters.

[0136] The iterative optimization of training the deep learning neural network specifically includes:

[0137] The error gradient is calculated using the initial classification results and known classification information;

[0138] The network weights are dynamically adjusted based on the error gradient, and the optimization is repeated multiple times until the classification error is less than the error threshold.

[0139] In practical applications, it can operate independently of X-ray images, relying solely on a single input facial optical image. A pre-trained neural network automatically identifies feature points, establishes a soft tissue coordinate system, and performs posture calibration. Based on X-ray mapping experience accumulated during training, it achieves rapid spatial standardization of the input image. The standardized feature angle vector is then input into a pre-trained classification network to automatically classify and diagnose jaw deformities. It automatically identifies and classifies the jaw deformity type of the target object, directly outputting the jaw deformity prediction result.

[0140] Example 2

[0141] This embodiment, based on Embodiment 1, conducts multi-faceted experimental verification of the diagnosis of oral and maxillofacial deformities, specifically including:

[0142] A multi-sensor method for diagnosing oral and maxillofacial deformities includes the following steps:

[0143] First, lateral facial photographs and corresponding lateral cephalometric X-ray data were collected from 1105 patients, including 580 patients with skeletal type I (ANB angle ranging from 0.7° to 4.7°, representing normal maxillary-mandibular relationship), 275 patients with skeletal type II (ANB angle greater than 4.7°, representing relative mandibular retrusion), and 250 patients with skeletal type III (ANB angle less than 0.7°, representing relative mandibular protrusion).

[0144] Subsequently, using a neural network model combining heatmap regression and coordinate regression (CHC network), 33 key anatomical landmarks were automatically located on the facial soft tissue in each lateral facial photograph. Based on these key points, 54 clinically significant facial soft tissue measurement angles were generated by calculating the lines formed between pairs of key points and the angles formed by combinations of three points.

[0145] In the ablation experiment on the influence of different angles on the diagnostic results, as shown in Table 1, in order to verify the influence of different angles on the diagnostic accuracy in the diagnosis of dentofacial deformities, based on 54 original angles, the imaging degree of different angles on the diagnostic results was judged. Only one angle was input at a time, and different angles were input in turn to judge the importance of different angles.

[0146] Table 1: Ablation experimental data on the impact of different angles on diagnostic results:

[0147]

[0148]

[0149] Analysis: Among the 54 angles, the diagnostic accuracy changed when any angle was entered, and the overall accuracy fluctuated significantly, proving that different angles have different levels of importance and different levels of help to the diagnostic results. As can be seen from Table 1, the diagnostic accuracy was higher when angles 1-10 were entered, proving that these angles have a greater impact on the diagnostic results.

[0150] In the ablation experiment investigating the impact of the number of angles on diagnostic accuracy, as shown in Table 2, to verify the influence and contribution of different angle numbers on the diagnostic results, we set up a comparative experiment. We progressively reduced the number of angles input into the diagnostic model to observe the impact of different final input angle numbers on the diagnostic results. First, we tested the accuracy using 54 angles. Then, we randomly selected 53 angles, reducing both the training and validation sets to 53 angles. We used this 53-angle training set to train the diagnostic model and simultaneously tested the validation set accuracy. Next, we selected 52 angles from the 53 angles, again reducing both the training and validation sets to 52 angles, trained the model, and validated the accuracy. Subsequently, we progressively reduced the number of angles by one from each previous experiment until only one angle was used. Each time the number of angles was reduced, the model needed to be trained, and the accuracy validated. We repeated this experiment six times to reduce the impact of randomly selected angles on the overall results and comprehensively evaluate the impact of angle reduction on the overall diagnostic accuracy.

[0151] Table 2: Ablation experimental data on the impact of the number of angles on the accuracy of diagnostic results:

[0152]

[0153]

[0154] Analysis: The number of angles in Table 2 was randomly selected. The data in the table shows that when the number of angles is greater than 20, the overall accuracy does not change much. As the number of angles decreases, the diagnostic accuracy gradually decreases, proving that the number of angles affects the diagnostic results. When the number of angles decreases to a certain extent, there are fewer important angles, resulting in a lower accuracy of the results.

[0155] In the integer heatmap resolution ablation experiment, as shown in Table 3, the influence of integer heatmap resolution on quantization error was verified. The fractional coordinate regression resolution was kept at 16×2, and only the integer heatmap resolution was adjusted to compare the performance of different downsampling ratios.

[0156] Table 3: Ablation experimental data on the effect of integer heatmap resolution on quantization error:

[0157]

[0158] Analysis: A 64×64 resolution strikes the best balance between quantization error and computational complexity. Increasing the resolution to 128×128 does not significantly improve performance, indicating that excessively high resolution introduces redundant computation. When the resolution is reduced to 32×32, the NME (Network Mean Estimation) increases significantly, verifying that the downsampling scale directly affects coordinate accuracy, impacting the side profile coordinate localization and diagnostic accuracy.

[0159] The ablation experimental data in Tables 1-3 were used to verify the importance of key features and the rationality of model parameter settings, thereby addressing the following issues: angle parameters contribute most to the diagnosis of jaw deformities; parameters are sufficient to support high-precision diagnosis; and the feature extraction network structure (such as heatmap resolution) was optimized to ensure diagnostic effectiveness and computational efficiency.

[0160] The MLP model was trained using three training schemes, as shown in Table 4, which illustrates the accuracy achieved under different batch sizes.

[0161] Table 4: Accuracy achieved under three training schemes and different batch sizes

[0162]

[0163] The experimental results are as follows:

[0164] Training scheme A: Three-class labeling was performed on 200 patients with osteopathic class I, 200 patients with osteopathic class II, and 200 patients with osteopathic class III, and then random sampling was used to form a balanced training dataset.

[0165] Training Plan B: 200 patients with osteopathic class I disease were labeled as negative, and 200 patients with osteopathic class II disease and 200 patients with osteopathic class III disease were labeled as positive, forming a binary classification training dataset.

[0166] Training scheme C: 400 patients with osteopathic class I were labeled as negative, and 200 patients with osteopathic class II and 200 patients with osteopathic class III were each labeled as positive, forming another binary classification training dataset.

[0167] During model training, tests were conducted for different batch sizes. Batch sizes were set to 8, 16, 32, and 64, and the model was tested for each of the three test schemes. Test scheme A included a balanced distribution of 50 cases of type I, 50 cases of type II, and 50 cases of type III; test scheme B included 100 cases of type I and a total of 100 cases of type II and type III; test scheme C simulated a natural distribution, including 180 cases of type I, 75 cases of type II, and 45 cases of type III.

[0168] The accuracy results of the model testing showed that training scheme A (balanced three-class classification training set) performed best under the conditions of batch size 64 and testing scheme A (balanced test set), achieving an accuracy of 94.7%, which proves that the method has significant clinical application potential and diagnostic advantages in the optical diagnosis of maxillofacial deformities.

[0169] Therefore, this embodiment achieves efficient and accurate diagnosis of oral and maxillofacial deformities by accurately locating soft tissue feature points and calculating facial spatial relationships, combined with an optimized MLP classification model. This overcomes the shortcomings of traditional diagnostic methods, especially avoiding the risk of X-ray radiation and reducing the reliance on professional medical personnel.

[0170] During training, the exact same multilayer perceptron (MLP) model was used for the classification and diagnosis of jaw deformities, but the input data differed. This input data came from different facial landmark detection networks, including HRnet, PIPnet, SLPT, HIH, and the CHC network proposed in this paper. After feature point detection on facial photographs using these networks, angle data was calculated and used as the model input.

[0171] The accuracy rates of different facial landmark detection networks in diagnostic tasks are shown in Table 5.

[0172] Table 5: Accuracy of different facial feature point detection network models in diagnostic tasks

[0173]

[0174] As can be seen from the table, the CHC network proposed in this paper, after feature point detection, achieves a diagnostic accuracy of 94.7% on the input data, which is the best among all methods and significantly outperforms other comparative methods. This indicates that the CHC network designed in this study can extract facial feature points more effectively and accurately, thereby improving the overall performance of the multilayer perceptron (MLP) model in diagnosing jaw deformities.

[0175] This embodiment also uses a test set of 150 patients with maxillofacial deformities, who were divided into three groups—skeletal Class I, skeletal Class II, and skeletal Class III—based on imaging examinations and cephalometric measurements before treatment. All cases were diagnosed and classified based on photographs using CHC+MLP, three junior physicians (graduate residents specializing in orthodontics), and three experienced specialists. Finally, these results were compared with the gold standard to examine the difference in diagnostic efficiency between the algorithm and clinical diagnosis by physicians.

[0176] The results are compared with the gold standard, as shown in Table 6:

[0177] Table 6: Comparison of the accuracy of CHC+MLP model and physician diagnosis

[0178]

[0179] This table visually verifies the diagnostic performance of the CHC+MLP model in actual clinical applications, clearly showing the difference in diagnostic ability between the algorithm (CHC network combined with MLP model) and doctors with different experience levels (young doctors and experts); by comparing the accuracy and AUC values ​​of the algorithm and doctors, it highlights the significant advantages of the algorithm in actual diagnostic scenarios.

[0180] Clearly, the diagnostic accuracy of the algorithm used in this embodiment (94.7%) is significantly higher than that of each doctor group (56.0%-70.0% for young doctors and 69.3%-73.3% for experts), and the AUC value is also much higher than that of doctors' diagnostic level. This indicates that the model's diagnostic efficacy is more stable and reliable. The significant statistical difference (P value < 0.001) confirms that the model has obvious and stable advantages in actual medical scenarios and has good application prospects. It provides clear evidence to promote the further promotion and application of model technology in primary healthcare, telemedicine, and initial screening diagnosis.

[0181] Example 3

[0182] This embodiment also provides a smart maxillofacial deformity examination instrument, including:

[0183] The system includes a detector with two independent cameras, a monitor, a computer, a support frame, and wheelsets.

[0184] The intelligent maxillofacial deformity examination instrument can collect clinical information, such as frontal and lateral facial photographs, anterior teeth occlusion photographs, dynamic videos of mouth opening, and patient symptoms, and then transmit this information to the data processing center. At the data processing center, multimodal feature extraction, recognition, fusion, and alignment operations are performed for preliminary diagnosis of maxillofacial deformities and prediction of temporomandibular joint disorder risk.

[0185] Guided by a smart maxillofacial deformity screening device, a person without a medical background used the equipment to collect data from students for preliminary diagnosis. No medical professionals were involved in this process. Afterwards, the students were taken to a hospital for further examination by dentists, including radiological evaluations.

[0186] A total of 440 students were examined by the intelligent maxillofacial deformity examination system, dentists, and radiologists. Based on cephalometric measurements, 50 students with skeletal Class I, 50 with skeletal Class II, and 50 with skeletal Class III deformities were randomly selected from these students. The preliminary jaw deformity predictions provided by the intelligent maxillofacial deformity examination system were compared with the cephalometric results. It was found that integrating this algorithm into the intelligent maxillofacial deformity examination system achieved an accuracy rate of 90.67% compared to the gold standard. Detailed results are shown in Table 7: Accuracy of Preliminary Diagnosis by the Intelligent Maxillofacial Deformity Examination System (IMDEU).

[0187]

[0188] Diagnostic accuracy: The table data shows that the overall diagnostic accuracy of the intelligent examination instrument reached 90.67%. Specifically: Class I accuracy: 82% (41 / 50), Class II accuracy: 96% (48 / 50), Class III accuracy: 94% (47 / 50). The experimental results fully demonstrate that the Intelligent Maxillofacial Deformity Examination Instrument (IMDEU) can reliably perform preliminary diagnosis of jaw deformities, and its performance is close to the gold standard of radiological examination (cephalometrics). It also verifies that even non-medical professionals can accurately complete data collection under the guidance of the device, proving the practicality and promotional value of IMDEU, especially in grassroots or school environments lacking professional medical personnel, where it can exert significant social and medical benefits.

[0189] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for determining jawbone deformities based on facial soft tissue feature points, characterized in that, Includes the following steps: S101. Obtain an optical image of the region to be analyzed, automatically detect soft tissue feature points in the optical image based on a pre-trained neural network model, and extract the initial coordinates of the soft tissue feature points. S102. Based on the initial coordinates, establish a standardized soft tissue coordinate system and determine the spatial measurement data between multiple facial soft tissue feature points. The spatial measurement data includes feature point distances and spatial angle vectors. S103: Input the spatial measurement data into the pre-trained deep learning neural network, and automatically calculate the coordinate correction values ​​of each soft tissue feature point based on the head position normalization and bony reference calibration module of the deep learning neural network, so as to realize the pose normalization and bony reference alignment of the soft tissue feature points in the input optical image. S104. Update the initial coordinates according to the coordinate correction value to obtain the calibrated facial soft tissue feature point coordinates and spatial measurement data. S105: Input the calibrated spatial measurement data into the classification module of the pre-trained deep learning neural network model to achieve automatic classification and assisted diagnosis of the jaw deformity category of the target object.

2. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 1, characterized in that, The process of obtaining the initial coordinates of the soft tissue feature points in the region to be analyzed includes: The facial region of the optical image of the area to be analyzed is standardized. The standardized image is trained by a machine learning model and decoded from an integer heatmap using feature maps. The integer coordinates with the highest probability values ​​are extracted for preliminary localization. In the region surrounding the initially located integer coordinates, the initial coordinates of the soft tissue feature points in the region to be analyzed are determined through local classification regression calculation.

3. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 2, characterized in that, The construction of the standardized soft tissue coordinate system specifically includes: The initial coordinates of the soft tissue feature points in the region to be analyzed are organized into a feature point set; using the custom soft tissue feature points as the reference feature points, a horizontal baseline is selected to construct a standardized two-dimensional coordinate system for facial soft tissue. A horizontal baseline is defined based on the origin of the coordinate system. The horizontal baseline is rotated to the horizontal axis of the coordinate system, and all feature points are rotated accordingly to obtain the coordinates of the rotated feature points. The facial length is defined as a standardized scale, and the initial coordinates of all soft tissue feature points are processed by standardization to obtain a coordinate system with uniform size. After establishing a standardized coordinate system, spatial measurement data of facial soft tissue feature points are calculated. The spatial measurement data includes feature point distances and spatial angle vectors. The distance between facial soft tissue feature points is defined and marked as the feature point distance, and the spatial angle data formed by any three facial soft tissue feature points is marked as the spatial angle vector.

4. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 3, characterized in that, Based on the spatial measurement data of facial soft tissue feature points, all soft tissue feature point combinations are calculated one by one to form a spatial measurement dataset. Based on multiple soft tissue feature points, facial feature angles with diagnostic value are selected from a predefined set of angles according to clinical diagnostic needs, and spatial measurement data of facial soft tissue feature points are constructed.

5. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 1, characterized in that, The construction logic of the deep learning neural network: During the model training phase, X-ray image data of the target object is acquired, and bony soft tissue feature points are marked in the X-ray image data. Key bony feature points include, but are not limited to, saddle point, nasal root point, maxillary anterior point, mandibular anterior point, infraorbital margin point, and ear canal point. Based on the aforementioned bony feature points, an X-ray cephalometric coordinate system is constructed, and the corresponding clinical standard bony measurement angles are calculated. Spatial alignment and mapping are performed between spatial measurement data obtained from optical images and corresponding bony measurement angles in X-ray image data. Using paired facial optical images and X-ray image data, a spatial registration relationship is established based on shared feature points, and the coordinate system error caused by changes in head posture in the optical images is learned, including: Overall rotational error caused by attitude change; Overall translational error caused by changes in head position; Scale deviation caused by different image acquisition conditions; Local drift deviation of soft tissue feature points in a local area; An error compensation mechanism is established based on coordinate system error to optimize the fitting accuracy of the standardized soft tissue coordinate system relative to the bone coordinate system, so as to achieve spatial uniformity and measurement consistency between different image modalities. The aforementioned spatial registration relationship and coordinate system error compensation mechanism are embedded into the neural network structure to complete the pre-training process of the deep learning neural network based on multimodal data alignment.

6. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 5, characterized in that, The calculation logic for the coordinate correction value is as follows: Based on the input initial soft tissue feature point coordinates and the constructed standardized soft tissue coordinate system, spatial measurement data between each feature point is extracted; By utilizing a pre-trained head position normalization and bony reference calibration module in a deep learning neural network, systematic errors between input spatial measurement data and the standard bony coordinate system are identified. Based on the identified systematic and local drift errors, the network generates coordinate correction values ​​for each soft tissue feature point, wherein the correction values ​​are: The coordinate correction values ​​of each feature point in the X and Y dimensions; based on the coordinate correction values, the feature point coordinates are mapped from the soft tissue reference frame of the optical image to a pose space consistent with the bony reference frame.

7. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 6, characterized in that, The pre-training logic of the deep learning neural network: The deep learning neural network includes a training dataset and a test dataset, which include a large number of paired optical images and X-ray image data. During the training phase, optical images with different head poses and soft tissue feature point distributions are used as input factors. The paired optical images and X-ray image data are optimized and trained in the deep learning neural network to automatically output coordinate correction values ​​consistent with the bony reference. During the generation of coordinate correction values, the historical attitude error distribution of the input data is learned simultaneously, and the generated coordinate correction values ​​are dynamically optimized by combining the current input spatial measurement data.

8. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 7, characterized in that, The logic for obtaining the calibrated facial soft tissue feature point coordinates and spatial measurement data is as follows: Using the initial coordinates of facial soft tissue feature points as a reference, the coordinate positions of facial soft tissue feature points are automatically calibrated based on the coordinate correction values ​​generated by dynamic optimization. The standardized coordinate system is updated based on the calibrated feature point coordinates, and the spatial measurement data between facial soft tissue feature points is recalculated to obtain calibrated spatial measurement data.

9. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 1, characterized in that, The training process of the deep learning neural network classification module includes: Using a dataset of historical jaw deformity patients, a feature vector with several feature angle values ​​is constructed as the input layer. The hidden layer uses the ReLU activation function, which includes the spatial angles of facial soft tissue feature points and known jaw classification information. The output layer uses the Softmax function to obtain the probability of each category. The model is gradually optimized through iterative training, and the backpropagation algorithm is used to train the model until it stabilizes and converges, until the confidence probability value of the model's output classification reaches the preset classification accuracy.

10. The method for determining jawbone deformity based on facial soft tissue feature points according to claim 9, characterized in that, The iterative optimization of training the deep learning neural network specifically includes: The error gradient is calculated using the initial classification results and known classification information; The network weights are dynamically adjusted based on the error gradient, and the optimization is repeated multiple times until the classification error is less than the error threshold.

Images