Mandibular symmetry detection method and system

Abstract

The present invention discloses a method and system for detecting mandibular symmetry. The method includes the following steps: first, a feature point position prediction model is trained using a mandibular feature point training dataset to obtain a trained mandibular feature point positioning model; second, a mandibular segmentation model is trained using a mandibular segmentation training dataset to obtain a trained mandibular segmentation model; third, a three-dimensional CBCT maxillofacial image to be processed is input into the trained mandibular feature point positioning model and the trained mandibular segmentation model, respectively, to obtain a mandibular feature point set and a mandibular fine segmentation result; fourth, a surface reconstruction is performed based on the mandibular fine segmentation result to obtain a triangular mesh mandibular model; and finally, a mandibular symmetry analysis is performed to obtain a symmetry analysis result. The present invention can identify mandibular shape and locate feature points in CBCT maxillofacial images, thereby achieving quantitative analysis of mandibular symmetry.

Description

Technical Field

[0001] The invention belongs to the field of medical graphics and image processing, and in particular relates to a mandibular symmetry detection method and system. Background Art

[0002] Jaw morphology and growth play a crucial role in orthodontics. Jaw symmetry influences masticatory function and aesthetics. In cases of malocclusion, such as asymmetrical molar lockjaw, or those requiring mandibular ramus osteotomy, jaw symmetry analysis helps measure the degree of jaw asymmetry and dynamically track the patient's orthodontic treatment and surgical outcomes.

[0003] Commonly used two-dimensional radiographs in clinical practice, such as lateral, posteroanterior, and panoramic radiographs, have limitations in their anterior-posterior overlap, severely impacting jaw symmetry analysis. Consequently, two-dimensional measurements are limited in the analysis and treatment planning of asymmetric deformities.

[0004] Current examination technology already supports the acquisition and data analysis of three-dimensional information of teeth, soft tissues, and hard tissues, and allows for evaluation. For hard tissues, unlike 2D X-rays, CBCT's reconstructed models allow for observation of multiple 3D views in space, rather than a single sagittal view. Furthermore, while 2D X-rays can only measure point positions or lengths, CBCT can provide measurements of surfaces, volumes, and other content, enriching the dimensions of jaw symmetry analysis. Furthermore, CBCT segmentation of the jaw can be constructed into a personalized three-dimensional digital jaw model for further evaluation.

[0005] At present, some technologies have proposed to perform symmetry analysis on the left and right temporomandibular joint discs of patients with asymmetric molar lockjaw, or to perform surface distance analysis related to jaw symmetry after manual segmentation and jaw partitioning. However, these studies rely on some manual recognition operations and are not yet fully automated. The accuracy of shape segmentation also needs to be improved. Summary of the Invention

[0006] In order to solve the problems and needs in the background technology, the purpose of the present invention is to provide a method and system for detecting mandibular symmetry.

[0007] The technical solutions of the present invention are as follows:

[0008] 1. A method for detecting mandibular symmetry

[0009] Step 1: After training the feature point position prediction model using the mandibular feature point training dataset, a trained mandibular feature point positioning model is obtained;

[0010] Step 2: After training the mandibular segmentation model using the jaw segmentation training dataset, a trained mandibular segmentation model is obtained;

[0011] Step 3: Input the 3D CBCT maxillofacial image to be processed into the trained jaw feature point positioning model, which outputs a set of jaw feature points. Furthermore, input the 3D CBCT maxillofacial image to be processed into the trained mandibular segmentation model, which outputs a precise mandibular segmentation result. Surface reconstruction is then performed based on the precise mandibular segmentation result to obtain a triangular mesh mandibular model.

[0012] Step 4: Perform symmetry analysis on the mandible based on the mandibular feature point set and the triangular mesh mandibular model to obtain the symmetry analysis results.

[0013] In step 1, the mandibular feature point training data set includes a low-resolution three-dimensional CBCT maxillofacial image and a corresponding high-resolution three-dimensional CBCT maxillofacial image, and a landmark point set, wherein the low-resolution and high-resolution three-dimensional CBCT maxillofacial images include complete upper and lower mandibles, the most anterior end of the nasal bone, left and right infraorbital points, and a complete C4 vertebral structure, and the landmark point set includes feature point coordinates for determining the midsagittal plane, feature point coordinates for cephalometric measurement, and feature point coordinates for mandibular morphological zoning and symmetry analysis.

[0014] In step 1, the feature point position prediction model includes a first position prediction model and a second position prediction model. The low-resolution three-dimensional CBCT maxillofacial image is used as input to the first position prediction model, and the first position prediction model is used to predict the rough position of the jaw feature points; the high-resolution three-dimensional CBCT maxillofacial image and the rough position set of the jaw feature points predicted by the first position prediction model are used together as input to the second position prediction model, and the second position prediction model is used to predict the precise position of the jaw feature points.

[0015] The first position prediction model includes a first image encoder, a multi-layer perceptron and a vector transformation module connected in sequence.

[0016] In the vector transformation module, first, the feature vector output by the multilayer perceptron is split into an x-direction coordinate vector △x^, a y-direction coordinate vector △y^, a z-direction coordinate vector △z^, a linear transformation vector △R^, and a displacement vector t^; then, the linear transformation vector △R^ is deformed to obtain a linear transformation matrix R^; finally, according to the x-direction coordinate vector △x^, the y-direction coordinate vector △y^, the z-direction coordinate vector △z^, the displacement vector t^ and the transformation matrix R^ and the learnable template coordinate parameter C p Generate the predicted rough position of the jaw feature points using the following formula:

[0017] C pred =C p ·R^ T +1 35 ·t^ T+[△x^ T △y^ T △z^ T ]

[0018] Among them, C pred is the rough position of the jaw feature point predicted by the first position prediction model, 1 35 is a column vector of size 35*1, and T represents the matrix transpose operation.

[0019] The second position prediction model includes a position preprocessing module, a relative position relationship generation module, a second image encoder, a feature extraction network and a self-attention module; the rough position set of the jaw feature points predicted by the first position prediction model is input into the position preprocessing module, the position preprocessing module outputs the transformed rough position set of the jaw feature points and the image subspace corresponding to all feature points, and the relative position relationship generation module generates a relative position matrix M according to the transformed rough position set of the jaw feature points. rel The second image encoder is connected to the feature extraction network. The image subspace corresponding to all feature points is used as the input of the second image encoder. The feature extraction network outputs the intermediate features corresponding to all feature points. The intermediate features corresponding to all feature points output by the feature extraction network form a feature matrix F. The feature matrix F and the relative position matrix M are combined. rel After splicing, the data are input into the self-attention module, which outputs the offset matrix N. Finally, the rough positions of each feature point in the rough position set of the transformed mandibular feature points are adjusted according to the offset matrix N to obtain the precise position of each feature point, thereby outputting the precise position set of the mandibular feature points.

[0020] In the position preprocessing module, the rough position set of the jaw feature points predicted by the first position prediction model is first spatially transformed to obtain the transformed rough position set of the jaw feature points. The formula is as follows:

[0021] C pred [i]=(C pred [i,1],C pred [i,2],C pred [i,3])

[0022] C L pred [i]= (C pred [i,1]×S x / 128,C pred [i,2]×S y / 128,C pred [i,3]×S z / 128)

[0023] Among them, C pred[i] is the coordinate of the i-th feature point predicted by the first position prediction model, C pred [i,1], C pred [i,2], C pred [i,3] are the rough positions C of the jaw feature points predicted by the first position prediction model pred The values ​​of row i, column 1, row i, column 2, and row i, column 3, C L pred [i] is the initial coordinate of the i-th feature point in the high-resolution space after spatial transformation, S x 、S y 、S z is the resolution size of high-resolution three-dimensional CBCT maxillofacial images in the x, y, and z dimensions;

[0024] Then, for the coarse position of each feature point in the transformed rough position set of mandibular feature points, the corresponding image subspace is intercepted in the high-resolution three-dimensional CBCT maxillofacial image with the coarse position of the feature point as the center. After traversing the coarse positions of all feature points in the transformed rough position set of mandibular feature points, the image subspace corresponding to all feature points is obtained.

[0025] In the relative position relationship generation module, a relative position matrix M is established based on the relative position relationship of each feature point in the rough position set of the transformed jaw feature points. rel , relative position matrix M rel The values ​​of each element in satisfy the following formula:

[0026] M rel [i, j]=C L pred [i,0]-C L pred [j,0]

[0027] M rel [i, j+35]=C L pred [i,1]-C L pred [j,1]

[0028] M rel [i, j+70]=C L pred [i,2]-C L pred [j,2]

[0029] Among them, M rel [i,y] is the relative position matrix M rel The value of the i-th row and y-th column in the matrix, y=j, j+35 or j+70, CL pred [z,w] is the rough position set C of the transformed jaw feature points L pred The value in the zth row and wth column, where z=i or j, and w=0, 1, or 2.

[0030] 2. A Mandibular Symmetry Detection System

[0031] A mandibular feature point positioning unit, used to store a feature point position prediction model and train the feature point position prediction model using a mandibular feature point training data set;

[0032] a mandibular segmentation unit, configured to store a mandibular segmentation model and train the mandibular segmentation model using a mandibular segmentation training dataset;

[0033] The mandibular model reconstruction unit is used to input the 3D CBCT maxillofacial image to be processed into the trained mandibular segmentation model. The model outputs the mandibular segmentation result, and then performs surface reconstruction based on the mandibular segmentation result to obtain a triangular mesh mandibular model.

[0034] The mandibular symmetry analysis unit is used to perform symmetry analysis on the mandibular according to the mandibular feature point set and the triangular mesh mandibular model to obtain a symmetry analysis result.

[0035] 3. A computer device

[0036] The device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the mandibular symmetry detection method when executing the computer program.

[0037] The present invention can achieve the following beneficial effects:

[0038] The present invention proposes a mandibular feature point positioning model and a mandibular segmentation model, which can locate the mandibular feature points and recognize the shape of the mandibular face in new CBCT maxillofacial images.

[0039] The present invention can quantitatively analyze the symmetry of the mandible in new CBCT maxillofacial images according to the positions and shapes of the mandibular feature points.

[0040] In clinical practice, the present invention can automatically obtain the symmetry status of the patient's mandible that the doctor is concerned about and present it visually, thereby reducing the doctor's workload, providing a more accurate basis for treatment plan design, reducing the possibility of misdiagnosis, and optimizing the overall diagnosis and treatment experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flow chart of the method of the present invention.

[0042] Figure 2 Schematic diagram of the original CBCT maxillofacial image in Example 1.

[0043] Figure 3 This is a schematic diagram of the mandibular contour zoning in Example 1.

[0044] Figure 4 Schematic diagram of jaw symmetry analysis in Example 1. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0046] Example 1

[0047] like Figure 1 As shown, a mandibular symmetry detection method proposed by the present invention includes the following steps:

[0048] Step 1: After training the feature point position prediction model using the mandibular feature point training dataset, a trained mandibular feature point positioning model is obtained;

[0049] The mandibular landmark training dataset consists of low-resolution 3D CBCT maxillofacial images (at least 150 mm x 150 mm x 150 mm), corresponding high-resolution 3D CBCT maxillofacial images, and a landmark set. Both the low-resolution and high-resolution 3D CBCT maxillofacial images contain the complete maxillary and mandibular structures, the anterior end of the nasal bones, the left and right infraorbital points, and the complete C4 vertebral structure. 3D CBCT maxillofacial images (in NIfTI or DICOM format) were read in and landmark measurements were performed using 3DSlicer software using a manual marking method. The landmark set includes the coordinates of landmark points (Type I points) used to determine the midsagittal plane (point A, ANS, and N), 16 feature point coordinates used for cephalometric measurements (Type II points in Table 1), and 16 feature point coordinates used for jaw morphological zoning and symmetry analysis (Type III points in Table 1). Details are shown in Table 1.

[0050] Table 1 Introduction to the 35 feature points of the landmark point set

[0051]

[0052] The feature point position prediction model includes a first position prediction model and a second position prediction model. The low-resolution three-dimensional CBCT maxillofacial image is used as the input of the first position prediction model, and the first position prediction model is used to predict the rough position of the jaw feature points; the high-resolution three-dimensional CBCT maxillofacial image and the rough position set of the jaw feature points predicted by the first position prediction model are used as the input of the second position prediction model, and the second position prediction model is used to predict the precise position of the jaw feature points.

[0053] The first position prediction model includes a first image encoder, a multi-layer perceptron, and a vector transformation module connected in sequence. Specifically, the first image encoder is a Swin autoencoder (Swin Transformer) model, which constructs a multi-level feature pyramid through a sliding window self-attention mechanism (Shifted Window Self-Attention), downsamples the input image step by step, and extracts multi-scale semantic features. The window division strategy reduces the computational complexity while retaining spatial locality; the feature maps of different levels of the encoder are cross-scale fused through jump connections and deformable convolutions to enhance the collaborative representation of fine-grained boundaries and global context information. The first image encoder can also be a multi-layer 3D convolutional neural network or a multi-layer self-attention network. The multi-layer perceptron converts the output of the first image encoder into a one-dimensional vector, that is, the multi-layer perceptron outputs a 117-dimensional feature vector. The multi-layer perceptron includes multiple fully connected layers connected in sequence. The input of the vector transformation module also includes the template coordinate parameter C p , which is of size 35×3 and is a learnable variable.

[0054] In the vector transformation module, first, the 117-dimensional feature vector output by the multilayer perceptron is split into a 35-dimensional x-direction coordinate vector △x^, a 35-dimensional y-direction coordinate vector △y^, a 35-dimensional z-direction coordinate vector △z^, a 9-dimensional linear transformation vector △R^, and a 3-dimensional displacement vector t^; then, after the 9-dimensional linear transformation vector △R^ is deformed, a 3×3 linear transformation matrix R^ is obtained, satisfying R^=E+△R^, where E is a 3×3 unit matrix; finally, according to the 35-dimensional x-direction coordinate vector △x^, the 35-dimensional y-direction coordinate vector △y^, the 35-dimensional z-direction coordinate vector △z^, the 3-dimensional displacement vector t^, the 3×3 transformation matrix R^, and the learnable template coordinate parameter C p Generate the predicted rough position of the jaw feature points using the following formula:

[0055] C pred =C p ·R^ T +1 35 ·t^ T +[△x^ T △y^ T △z^T ]

[0056] Among them, C pred The rough position of the jaw feature points predicted by the first position prediction model, with a size of 35×3, 1 35 is a column vector of size 35*1, 1 35 ·t^ T The size of is 35×3, and T represents the matrix transpose operation.

[0057] During the experiment, the original three-dimensional CBCT maxillofacial image (in the format specified by NIfTI or DICOM) is first read in and converted into a tensor form that can be read by the first position prediction model used in the present invention. The three-dimensional CBCT maxillofacial image is then spatially normalized, and the image is scaled to a low-resolution three-dimensional CBCT maxillofacial image of uniform size (128*128*128 pixels) using cubic interpolation. Then, the low-resolution three-dimensional CBCT maxillofacial image is grayscale normalized. The normalization method is: first calculate the average value of the image grayscale as v' and the variance as v d ; Then perform the following calculation for all pixels v in the image: v←v / v d - v', obtain the final low-resolution three-dimensional CBCT maxillofacial image I1, that is, the low-resolution three-dimensional CBCT maxillofacial image in the jaw feature point training data set.

[0058] During model training, the average of the differences between the predicted coordinates of all feature points and the coordinates of their corresponding manually marked feature points is used as the loss function. The AdamW gradient descent optimizer is used to perform backpropagation and update parameters to complete the training. During prediction, the forward propagation calculation method is the same as described above.

[0059] The first position prediction model and the second position prediction model are trained separately. After the first position prediction model is trained, the second position prediction model is trained.

[0060] The second position prediction model includes a position preprocessing module, a relative position relationship generation module, a second image encoder, a feature extraction network and a self-attention module. The rough position set of the jaw feature points predicted by the first position prediction model is input into the position preprocessing module. The position preprocessing module outputs the transformed rough position set of the jaw feature points and the image subspace corresponding to all feature points. The relative position relationship generation module generates a relative position matrix M based on the transformed rough position set of the jaw feature points. rel , the second image encoder is connected to the feature extraction network, the image subspace corresponding to all feature points is used as the input of the second image encoder, and the feature extraction network outputs the intermediate features corresponding to all feature points.

[0061] Specifically, the second image encoder is a Swin autoencoder (Swin Transformer) model. This model uses a shifted window self-attention mechanism to construct a multi-level feature pyramid, progressively downsampling the input image and extracting multi-scale semantic features. This window partitioning strategy reduces computational complexity while preserving spatial locality. Feature maps from different encoder layers are fused across scales using skip connections and deformable convolutions to enhance the collaborative representation of fine-grained boundaries and global contextual information. The feature extraction network is a multi-layer perceptron. This feature extraction network takes the output of the second image encoder as input and outputs a one-dimensional vector with a fixed length of 512. The self-attention module positionally encodes the n input features, then uses the self-attention mechanism to mine feature correlations and output 35 three-dimensional numerical values. The self-attention module consists of multiple sequentially connected self-attention blocks.

[0062] In the position preprocessing module, the rough position set of the jaw feature points predicted by the first position prediction model is first spatially transformed to obtain the transformed rough position set of the jaw feature points. The formula is as follows:

[0063] C pred [i]=(C pred [i,1],C pred [i,2],C pred [i,3])

[0064] C L pred [i]= (C pred [i,1]×S x / 128, C pred [i,2]×S y / 128, C pred [i,3]×S z / 128)

[0065] Among them, C pred [i] is the coordinate of the i-th feature point predicted by the first position prediction model, with a size of 35×3, C pred [i,1], C pred [i,2], C pred [i,3] are the rough position set C of the jaw feature points predicted by the first position prediction model pred (which is in matrix form) the values ​​of row i, column 1, row i, column 2, and row i, column 3, C L pred [i] is the initial coordinate of the i-th feature point in the high-resolution space after spatial transformation, S x 、S y 、Sz The resolution of high-resolution three-dimensional CBCT maxillofacial images in the x, y, and z dimensions is the same;

[0066] Next, for each feature point in the transformed rough jaw feature point location set, a corresponding image subspace is extracted from the high-resolution 3D CBCT maxillofacial image, centered at that feature point's rough location. This subspace is denoted as subimg[i], where i = 1, 2, …, 35 and the length, width, and height of the image subspace are 64 pixels. After traversing the coarse locations of all feature points in the transformed rough jaw feature point location set, the image subspace corresponding to each feature point is obtained. Each image subspace serves as the input to the second image encoder, which outputs the corresponding intermediate feature, feature[i], from the feature extraction network.

[0067] In the relative position relationship generation module, the relative position matrix M is established based on the relative position relationship of each feature point in the rough position set of the transformed jaw feature points. rel , the size is 35*105. Relative position matrix M rel The values ​​of each element in satisfy the following formula:

[0068] M rel [i, j]=C L pred [i,0]-C L pred [j,0]

[0069] M rel [i, j+35]=C L pred [i,1]-C L pred [j,1]

[0070] M rel [i, j+70]=C L pred [i,2]-C L pred [j,2]

[0071] Among them, M rel [i,y] is the relative position matrix M rel The value of the i-th row and y-th column in the matrix, y=j, j+35, or j+70, C L pred [z,w] is the rough position set C of the transformed jaw feature points L pred The value in the zth row and wth column, where z=i or j, and w=0, 1, or 2.

[0072] The intermediate features corresponding to all feature points output by the feature extraction network form the feature matrix F, and the feature matrix F and the relative position matrix M rel After splicing, the data are input into the self-attention module, which outputs a 35*3 offset matrix N, which contains the position offsets corresponding to the 35 feature points. Finally, the rough positions of each feature point in the rough position set of the transformed mandibular feature points are adjusted according to the offset matrix N to obtain the precise position of each feature point, thereby outputting the precise position set of the mandibular feature points.

[0073] During the experiment, the original three-dimensional CBCT maxillofacial image (in the format specified by NIfTI or DICOM) is first read in and converted into a tensor form that can be read by the second position prediction model used in the present invention. The three-dimensional CBCT maxillofacial image is then spatially normalized, and the image is scaled to a uniform size and isotropic high-definition (0.3mm / px*0.3mm / px*0.3mm / px) image using cubic interpolation. Then, the low-resolution three-dimensional CBCT maxillofacial image is grayscale normalized. The normalization method is: first calculate the average value of the image grayscale as v' and the variance as v d ; Then perform the following calculation for all pixels v in the image: v←v / v d - v', obtain the final high-resolution three-dimensional CBCT maxillofacial image I1, that is, the high-resolution three-dimensional CBCT maxillofacial image in the jaw feature point training data set.

[0074] During training, the average of the differences between the predicted coordinates of all feature points and the coordinates of their corresponding manually marked feature points is used as the loss function. The AdamW gradient descent optimizer is used to perform backpropagation and update parameters to complete the training. During prediction, the forward propagation calculation method is the same as the above method.

[0075] Step 2: After training the mandibular segmentation model using the jaw segmentation training dataset, a trained mandibular segmentation model is obtained;

[0076] The mandibular feature point training dataset consists of original 3D CBCT maxillofacial images of at least 150 mm x 150 mm x 150 mm in size and their corresponding labeled images. The original 3D CBCT maxillofacial images contain the complete maxillary and mandibular structures, the anterior end of the nasal bones, the left and right infraorbital points, and the complete C4 vertebral structure. Large-field CBCT images (in NIfTI or DICOM format) were read in, and the mandible was manually segmented using 3D Slicer software. Each voxel in the labeled image has a value of 0 or 1, representing the non-mandibular region and the mandibular region, respectively.

[0077] The mandibular segmentation model includes a first mandibular segmentation sub-model and a second mandibular segmentation sub-model. The original three-dimensional CBCT maxillofacial image serves as the input of the first mandibular segmentation sub-model, and the first mandibular segmentation sub-model is used to output the rough segmentation result of the mandible. After data reorganization of the original three-dimensional CBCT maxillofacial image, the rough segmentation result of the mandible output by the first mandibular segmentation sub-model, and the label image, the reorganized dataset is obtained and used as the input of the second mandibular segmentation sub-model, and the second mandibular segmentation sub-model is used to output the fine segmentation result of the mandible.

[0078] Both the first and second mandibular segmentation sub-models are composed of multiple modules with an encoder-decoder architecture and incorporate a self-attention mechanism to enhance the modeling of global feature relationships. The coarse segmentation sub-network outputs a two-channel probability map, representing the probability of a voxel belonging to the background and the mandible, respectively. This stage aims to achieve a rough localization of the mandibular region.

[0079] Data reconstruction is performed on the high-resolution 3D CBCT maxillofacial image, the mandibular coarse segmentation results output by the first mandibular segmentation sub-model, and the label image, including:

[0080] Each original 3D CBCT maxillofacial image and the mandibular coarse segmentation result (i.e., 2-channel probability map) output by the corresponding first mandibular segmentation sub-model are resampled to a spatial resolution of r mm / pixel, where r=0.3; then, the resampled 3D CBCT maxillofacial image and the mandibular coarse segmentation result are spliced ​​to form a mandibular fusion image; considering that the volume of the 3D image after resolution unification is large and may exceed the GPU memory capacity limit, the mandibular fusion image is then cut into several local image blocks of the same size, and a reconstructed dataset is composed of all local image blocks and label images of corresponding sizes, that is, each local image block is used as the input of the second mandibular segmentation sub-model. The mandibular fusion image cutting process of the present invention adopts a moving window cropping strategy. Specifically, a small sliding window of size 128*128*128 is constructed, starting from (0,0,0), and the sliding window is moved to extract the local area of ​​the corresponding position from the mandibular fusion image, and the fine segmentation neural network is used to predict the local area of ​​each position in turn until the sliding window covers the entire image space. If the mandibular region does not exist in the local area of ​​the position in the coarse mandibular segmentation result, it will be skipped directly. The prediction results of the local area of ​​each position are integrated to obtain the full-size, high-precision mandibular segmentation result.

[0081] During training, a binary cross-entropy loss function is used to compare the error between the output of the first or second mandibular segmentation sub-model and the manually labeled segmented regions. The first and second mandibular segmentation sub-models are trained separately using stochastic gradient descent. The second mandibular segmentation sub-model is trained after the first is trained.

[0082] Step 3: Figure 2 The 3D CBCT maxillofacial image to be processed is input into a trained jaw feature point positioning model, which outputs a jaw feature point set. The 3D CBCT maxillofacial image to be processed is input into a trained mandibular segmentation model, which outputs a mandibular precise segmentation result. Based on the mandibular precise segmentation result, surface reconstruction is performed using the Marching Cubes Algorithm to obtain a high-precision triangular mesh mandibular model.

[0083] Step 4: Perform symmetry analysis on the mandible based on the mandibular feature point set and the triangular mesh mandibular model to obtain the symmetry analysis results.

[0084] The symmetry analysis of the mandible includes the detection of mandibular dimensions, the symmetry analysis of the jaw contour, the symmetry analysis of the jaw surface distance and surface area, and the symmetry analysis of the jaw volume.

[0085] Mandibular dimensions include but are not limited to bilateral mandibular width, bilateral mandibular body length, and bilateral mandibular height. The left / right mandibular width is the perpendicular distance from the left / right condylar lateral point (CDL) to the midsagittal plane. The left / right mandibular body length is the distance from the left / right mandibular angle (Go) to the promental point (Pog). The left / right mandibular height is the distance from the left / right mandibular angle (Go) to the highest point of the mandible (Co) on the corresponding side.

[0086] Symmetry analysis of the jaw contour includes:

[0087] Based on the 15 points in the mandibular feature point set used for contour symmetry analysis and the triangular mesh mandibular model, the mandibular contour is partitioned, such as Figure 3The mandible is divided into six functional zones: the articular zone, the mandibular ramus Ra, the mandibular angle Ma, the mandibular body Mb, the alveolar process Ap, and the chin Ch. The functional zones are defined as follows: the articular zone is composed of the condyle and coracoid process (Cr), with its lower boundary being the perpendicular line to the RP1-RP2 line passing through point Sg; the mandibular ramus Ra is bounded above by the perpendicular line to the RP1-RP2 line, and its lower boundaries are the RP2-B line and the Ri-Ag line; the mandibular angle Ma is bounded above by the RP2-B line and the Ri-Ag line; the mandibular body Mb is bounded posteriorly by the Ri-Ag line, above by the RP2-B line, and anteriorly by the B-MP2 line; the alveolar process Ap is bounded posteriorly by the Ri-Ag line and below by the RP2-B line; and the chin Ch is bounded posteriorly by the B-MP line.

[0088] Symmetry analysis of jaw surface distances and surface areas includes:

[0089] According to the midsagittal plane of the mandibular feature point set, the triangular mesh mandibular model is mirrored along the midsagittal plane. The surface distances of each contour partition of the mandibular are compared in sequence. The contour partitions are: the left articular part and the mirrored right articular part, the left mandibular ascending part and the mirrored right mandibular ascending part, the left mandibular angle and the mirrored right mandibular angle, the left mandibular body and the mirrored right mandibular body, and the left alveolar process and the mirrored right alveolar process. For each pair of partitions, the first point set A is composed of the coordinates of the surface points in the original left partition, and the second point set B is composed of the coordinates of the surface points in the mirrored right partition. The Hausdorff surface distance HD(A,B) of the pair of partitions is calculated as follows:

[0090]

[0091] in, represents the Euclidean distance, a is the coordinate of the surface point in the first point set A, and b is the coordinate of the surface point in the second point set B. This indicator reflects the maximum local asymmetry between the two surfaces and can be used to evaluate the degree of symmetry of the jaw morphology between the left and right sides.

[0092] Figure 4 This is a visualization of the mandibular symmetry assessment results. Each point on the original mandibular model surface is colored according to its minimum distance from the mirrored mandibular model surface. The relationship between distance and color is represented by the color bar on the right. A reddish color indicates that the mirrored model protrudes from the original model at that point, while a blueish color indicates that the mirrored model is recessed into the original model at that point.

[0093] In addition, to further quantify the symmetry difference, the surface area of ​​each anatomical region was calculated, and the surface area difference and relative difference rate between the corresponding regions on the left and right sides were calculated using the following formula:

[0094] △S=|Sleft -S mirror-right |

[0095] RelDiff(%)=2△S / (S left +S mirror-right )×100%

[0096] Among them, S left and S mirror-right are the surface areas of the original left partition and the mirrored right partition, △S is the surface area difference, RelDiff(%) is the relative difference rate, and || indicates the absolute value operation.

[0097] Symmetry analysis of jaw volume includes:

[0098] The triangular mesh mandibular model was used to compare the volumes of each symmetrical contour partition, including the left and right articular regions, the left and right mandibular ascending ramus, the left and right mandibular angles, the left and right mandibular bodies, and the left and right alveolar processes. The volume △V of each symmetrical contour partition was calculated as follows:

[0099] △V=|V left -V right |

[0100] Among them, V left is the volume of the left contour partition, V right The volume of the right contour partition.

[0101] The present invention also proposes a mandibular symmetry detection system, which includes:

[0102] The jaw feature point positioning unit is used to store the feature point position prediction model and train the feature point position prediction model using the jaw feature point training data set; the 3D CBCT maxillofacial image to be processed is input into the trained jaw feature point positioning model in the jaw feature point positioning unit, and the model outputs the jaw feature point set;

[0103] a mandibular segmentation unit, configured to store a mandibular segmentation model and train the mandibular segmentation model using a mandibular segmentation training dataset;

[0104] The mandibular model reconstruction unit is used to input the 3D CBCT maxillofacial image to be processed into the trained mandibular segmentation model. The model outputs the mandibular segmentation result, and then performs surface reconstruction based on the mandibular segmentation result to obtain a triangular mesh mandibular model.

[0105] The mandibular symmetry analysis unit is used to perform symmetry analysis on the mandibular according to the mandibular feature point set corresponding to the three-dimensional CBCT maxillofacial image to be processed and the triangular mesh mandibular model to obtain a symmetry analysis result.

[0106] The present invention also proposes a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the steps of a mandibular symmetry detection method when executing the computer program.

[0107] Example 2

[0108] like Figure 1 As shown, a mandibular symmetry detection method proposed by the present invention includes the following steps:

[0109] Step 1: After training the feature point position prediction model using the mandibular feature point training dataset, a trained mandibular feature point positioning model is obtained;

[0110] The mandibular landmark training dataset consists of low-resolution 3D CBCT maxillofacial images (at least 150 mm x 150 mm x 150 mm) and their corresponding high-resolution images, along with a landmark set. The low- and high-resolution 3D CBCT maxillofacial images contain the complete maxillary and mandibular structures, the anterior end of the nasal bones, the left and right infraorbital points, and the complete C4 vertebral structure. The 3D CBCT maxillofacial images (in NIfTI or DICOM format) were read in and landmark measurements were performed on the images using 3DSlicer software using manual marking. The landmark set includes the coordinates of landmark points (Type I points) used to determine the midsagittal plane (point A, ANS, and N), 16 feature points used for cephalometric measurements (Type II points), and 16 feature points used for jaw morphology zoning and symmetry analysis (Type III points).

[0111] The feature point position prediction model includes a first position prediction model and a second position prediction model. The low-resolution three-dimensional CBCT maxillofacial image is used as the input of the first position prediction model, and the first position prediction model is used to predict the rough position of the jaw feature points; the high-resolution three-dimensional CBCT maxillofacial image and the rough position set of the jaw feature points predicted by the first position prediction model are used as the input of the second position prediction model, and the second position prediction model is used to predict the precise position of the jaw feature points.

[0112] The first position prediction model includes a first image encoder, a multi-layer perceptron, and a vector transformation module connected in sequence. Specifically, the first image encoder is a Swin autoencoder (Swin Transformer) model, which constructs a multi-level feature pyramid through a sliding window self-attention mechanism (Shifted Window Self-Attention), downsamples the input image step by step, and extracts multi-scale semantic features. The window division strategy reduces the computational complexity while retaining spatial locality; the feature maps of different levels of the encoder are cross-scale fused through jump connections and deformable convolutions to enhance the collaborative representation of fine-grained boundaries and global context information. The first image encoder can also be a multi-layer 3D convolutional neural network or a multi-layer self-attention network. The multi-layer perceptron converts the output of the first image encoder into a one-dimensional vector, that is, the multi-layer perceptron outputs a 117-dimensional feature vector. The multi-layer perceptron includes multiple fully connected layers connected in sequence. The input of the vector transformation module also includes the template coordinate parameter C p , which is of size 35×3 and is a learnable variable.

[0113] In the vector transformation module, first, the 117-dimensional feature vector output by the multilayer perceptron is split into a 35-dimensional x-direction coordinate vector △x^, a 35-dimensional y-direction coordinate vector △y^, a 35-dimensional z-direction coordinate vector △z^, a 9-dimensional linear transformation vector △R^, and a 3-dimensional displacement vector t^; then, after the 9-dimensional linear transformation vector △R^ is deformed, a 3×3 linear transformation matrix R^ is obtained, satisfying R^=E+△R^, where E is a 3×3 unit matrix; finally, according to the 35-dimensional x-direction coordinate vector △x^, the 35-dimensional y-direction coordinate vector △y^, the 35-dimensional z-direction coordinate vector △z^, the 3-dimensional displacement vector t^, the 3×3 transformation matrix R^, and the learnable template coordinate parameter C p Generate the predicted rough position of the jaw feature points using the following formula:

[0114] C pred =C p ·R^ T +1 35 ·t^ T +[△x^ T △y^ T △z^ T ]

[0115] Among them, C pred The rough position of the jaw feature points predicted by the first position prediction model, with a size of 35×3, 1 35 is a column vector of size 35*1, 1 35 ·t^ T The size of is 35×3, and T represents the matrix transpose operation.

[0116] During the experiment, the original three-dimensional CBCT maxillofacial image (in the format specified by NIfTI or DICOM) is first read in and converted into a tensor form that can be read by the first position prediction model used in the present invention. The three-dimensional CBCT maxillofacial image is then spatially normalized, and the image is scaled to a low-resolution three-dimensional CBCT maxillofacial image of uniform size (128*128*128 pixels) using cubic interpolation. Then, the low-resolution three-dimensional CBCT maxillofacial image is grayscale normalized. The normalization method is: first calculate the average value of the image grayscale as v' and the variance as v d ; Then perform the following calculation for all pixels v in the image: v←v / v d - v', obtain the final low-resolution three-dimensional CBCT maxillofacial image I1, that is, the low-resolution three-dimensional CBCT maxillofacial image in the jaw feature point training data set.

[0117] During model training, the average of the differences between the predicted coordinates of all feature points and the coordinates of their corresponding manually marked feature points is used as the loss function. The AdamW gradient descent optimizer is used to perform backpropagation and update parameters to complete the training. During prediction, the forward propagation calculation method is the same as described above.

[0118] The first position prediction model and the second position prediction model are trained separately. After the first position prediction model is trained, the second position prediction model is trained.

[0119] The second position prediction model includes a position preprocessing module, a relative position relationship generation module, a second image encoder, a feature extraction network and a self-attention module. The rough position set of the jaw feature points predicted by the first position prediction model is input into the position preprocessing module. The position preprocessing module outputs the transformed rough position set of the jaw feature points and the image subspace corresponding to all feature points. The relative position relationship generation module generates a relative position matrix M based on the transformed rough position set of the jaw feature points. rel , the second image encoder is connected to the feature extraction network, the image subspace corresponding to all feature points is used as the input of the second image encoder, and the feature extraction network outputs the intermediate features corresponding to all feature points.

[0120] Specifically, the second image encoder is a Swin autoencoder (Swin Transformer) model. This model uses a shifted window self-attention mechanism to construct a multi-level feature pyramid, progressively downsampling the input image and extracting multi-scale semantic features. This window partitioning strategy reduces computational complexity while preserving spatial locality. Feature maps from different encoder layers are fused across scales using skip connections and deformable convolutions to enhance the collaborative representation of fine-grained boundaries and global contextual information. The feature extraction network is a multi-layer perceptron. This feature extraction network takes the output of the second image encoder as input and outputs a one-dimensional vector with a fixed length of 512. The self-attention module positionally encodes the n input features, then uses the self-attention mechanism to mine feature correlations and output 35 three-dimensional numerical values. The self-attention module consists of multiple sequentially connected self-attention blocks.

[0121] In the position preprocessing module, the rough position set of the jaw feature points predicted by the first position prediction model is first spatially transformed to obtain the transformed rough position set of the jaw feature points. The formula is as follows:

[0122] C pred [i]=(C pred [i,1],C pred [i,2],C pred [i,3])

[0123] C L pred [i]= (C pred [i,1]×S x / 128, C pred [i,2]×S y / 128, C pred [i,3]×S z / 128)

[0124] Among them, C pred [i] is the coordinate of the i-th feature point predicted by the first position prediction model, with a size of 35×3, C pred [i,1], C pred [i,2], C pred [i,3] are the rough position set C of the jaw feature points predicted by the first position prediction model pred (which is in matrix form) the values ​​of row i, column 1, row i, column 2, and row i, column 3, C L pred [i] is the initial coordinate of the i-th feature point in the high-resolution space after spatial transformation, S x 、S y 、Sz The resolution of high-resolution three-dimensional CBCT maxillofacial images in the x, y, and z dimensions is the same;

[0125] Next, for each feature point in the transformed rough jaw feature point location set, a corresponding image subspace is extracted from the high-resolution 3D CBCT maxillofacial image, centered at that feature point's rough location. This subspace is denoted as subimg[i], where i = 1, 2, …, 35 and the length, width, and height of the image subspace are 64 pixels. After traversing the coarse locations of all feature points in the transformed rough jaw feature point location set, the image subspace corresponding to each feature point is obtained. Each image subspace serves as the input to the second image encoder, which outputs the corresponding intermediate feature, feature[i], from the feature extraction network.

[0126] In the relative position relationship generation module, the relative position matrix M is established based on the relative position relationship of each feature point in the rough position set of the transformed jaw feature points. rel , the size is 35*105. Relative position matrix M rel The values ​​of each element in satisfy the following formula:

[0127] M rel [i, j]=C L pred [i,0]-C L pred [j,0]

[0128] M rel [i, j+35]=C L pred [i,1]-C L pred [j,1]

[0129] M rel [i, j+70]=C L pred [i,2]-C L pred [j,2]

[0130] Among them, M rel [i,y] is the relative position matrix M rel The value of the i-th row and y-th column in the matrix, y=j, j+35, or j+70, C L pred [z,w] is the rough position set C of the transformed jaw feature points L pred The value in the zth row and wth column, where z=i or j, and w=0, 1, or 2.

[0131] The intermediate features corresponding to all feature points output by the feature extraction network form the feature matrix F, and the feature matrix F and the relative position matrix M rel After splicing, the data are input into the self-attention module, which outputs a 35*3 offset matrix N, which contains the position offsets corresponding to the 35 feature points. Finally, the rough positions of each feature point in the rough position set of the transformed mandibular feature points are adjusted according to the offset matrix N to obtain the precise position of each feature point, thereby outputting the precise position set of the mandibular feature points.

[0132] During the experiment, the original three-dimensional CBCT maxillofacial image (in the format specified by NIfTI or DICOM) is first read in and converted into a tensor form that can be read by the second position prediction model used in the present invention. The three-dimensional CBCT maxillofacial image is then spatially normalized, and the image is scaled to a uniform size and isotropic high-definition (0.3mm / px*0.3mm / px*0.3mm / px) image using cubic interpolation. Then, the low-resolution three-dimensional CBCT maxillofacial image is grayscale normalized. The normalization method is: first calculate the average value of the image grayscale as v' and the variance as v d ; Then perform the following calculation for all pixels v in the image: v←v / v d - v', obtain the final high-resolution three-dimensional CBCT maxillofacial image I1, that is, the high-resolution three-dimensional CBCT maxillofacial image in the jaw feature point training data set.

[0133] During training, the average of the differences between the predicted coordinates of all feature points and the coordinates of their corresponding manually marked feature points is used as the loss function. The AdamW gradient descent optimizer is used to perform backpropagation and update parameters to complete the training. During prediction, the forward propagation calculation method is the same as the above method.

[0134] Step 2: After training the mandibular segmentation model using the jaw segmentation training dataset, a trained mandibular segmentation model is obtained;

[0135] The mandibular feature point training dataset consists of original 3D CBCT maxillofacial images of at least 150 mm x 150 mm x 150 mm in size and their corresponding labeled images. The original 3D CBCT maxillofacial images contain the complete maxillary and mandibular structures, the anterior end of the nasal bones, the left and right infraorbital points, and the complete C4 vertebral structure. Large-field CBCT images (in NIfTI or DICOM format) were read in, and the mandible was manually segmented using 3D Slicer software. Each voxel in the labeled image has a value of 0 or 1, representing the non-mandibular region and the mandibular region, respectively.

[0136] The mandibular segmentation model includes a first mandibular segmentation sub-model and a second mandibular segmentation sub-model. The original three-dimensional CBCT maxillofacial image serves as the input of the first mandibular segmentation sub-model, and the first mandibular segmentation sub-model is used to output the rough segmentation result of the mandible. After data reorganization of the original three-dimensional CBCT maxillofacial image, the rough segmentation result of the mandible output by the first mandibular segmentation sub-model, and the label image, the reorganized dataset is obtained and used as the input of the second mandibular segmentation sub-model, and the second mandibular segmentation sub-model is used to output the fine segmentation result of the mandible.

[0137] Both the first and second mandibular segmentation sub-models are composed of multiple encoder-decoder modules and incorporate a self-attention mechanism to enhance their ability to model global feature relationships. The coarse segmentation sub-network outputs a two-channel probability map, representing the probability of a voxel belonging to the background and the mandible, respectively. This stage aims to achieve a rough localization of the mandibular region.

[0138] Data reconstruction is performed on the high-resolution 3D CBCT maxillofacial image, the mandibular coarse segmentation results output by the first mandibular segmentation sub-model, and the label image, including:

[0139] Each original 3D CBCT maxillofacial image and the corresponding mandibular coarse segmentation result (i.e., 2-channel probability map) output by the first mandibular segmentation sub-model were resampled to a spatial resolution of r mm / pixel, where r = 0.3. Next, the resampled 3D CBCT maxillofacial image and the mandibular coarse segmentation result were concatenated to form a mandibular fusion image. The reconstructed dataset consisted of all mandibular fusion images and labeled images of corresponding sizes, i.e., each mandibular fusion image served as the input to the second mandibular segmentation sub-model.

[0140] During training, a binary cross-entropy loss function is used to compare the error between the output of the first or second mandibular segmentation sub-model and the manually labeled segmented regions. The first and second mandibular segmentation sub-models are trained separately using stochastic gradient descent. The second mandibular segmentation sub-model is trained after the first is trained.

[0141] Step 3: Input the 3D CBCT maxillofacial image to be processed into the trained jaw feature point positioning model, which outputs a set of jaw feature points. Also, input the 3D CBCT maxillofacial image to be processed into the trained mandibular segmentation model, which outputs a precise mandibular segmentation result. Based on this precise mandibular segmentation result, surface reconstruction is performed using the Marching Cubes algorithm to obtain a high-precision triangular mesh mandibular model.

[0142] Step 4: Perform symmetry analysis on the mandible based on the mandibular feature point set and the triangular mesh mandibular model to obtain the symmetry analysis results.

[0143] The symmetry analysis of the mandible includes the detection of mandibular dimensions, the symmetry analysis of the jaw contour, the symmetry analysis of the jaw surface distance and surface area, and the symmetry analysis of the jaw volume.

[0144] Mandibular dimensions include but are not limited to bilateral mandibular width, bilateral mandibular body length, and bilateral mandibular height. The left / right mandibular width is the perpendicular distance from the left / right condylar lateral point (CDL) to the midsagittal plane. The left / right mandibular body length is the distance from the left / right mandibular angle (Go) to the promental point (Pog). The left / right mandibular height is the distance from the left / right mandibular angle (Go) to the highest point of the mandible (Co) on the corresponding side.

[0145] Symmetry analysis of the jaw contour includes:

[0146] The mandibular contour was partitioned based on 15 points in the mandibular feature set used for contour symmetry analysis and a triangular mesh mandibular model. The mandible was divided into six functional areas: the articular region, the ascending ramus Ra, the mandibular angle Ma, the mandibular body Mb, the alveolar process Ap, and the chin Ch. The definitions of each functional area are as follows: the articular part is composed of the condyle and the coracoid process (Cr), and its lower boundary is the perpendicular line of the RP1-RP2 line passing through point Sg; the upper boundary of the mandibular ascending ramus Ra is the perpendicular line of the RP1-RP2 line, and its lower boundaries are the RP2-B line and the Ri-Ag line; the upper boundary of the mandibular angle Ma is the RP2-B line and the Ri-Ag line; the posterior boundary of the mandibular body Mb is the Ri-Ag line, the upper boundary is the RP2-B line, and the anterior boundary is the B-MP2 line; the posterior boundary of the alveolar process Ap is the Ri-Ag line, and its lower boundary is the RP2-B line; the posterior boundary of the chin Ch is the B-MP line.

[0147] Symmetry analysis of jaw surface distances and surface areas includes:

[0148] According to the midsagittal plane of the mandibular feature point set, the triangular mesh mandibular model is mirrored along the midsagittal plane. The surface distances of each contour partition of the mandibular are compared in sequence. The contour partitions are: the left articular part and the mirrored right articular part, the left mandibular ascending part and the mirrored right mandibular ascending part, the left mandibular angle and the mirrored right mandibular angle, the left mandibular body and the mirrored right mandibular body, and the left alveolar process and the mirrored right alveolar process. For each pair of partitions, the first point set A is composed of the coordinates of the surface points in the original left partition, and the second point set B is composed of the coordinates of the surface points in the mirrored right partition. The Hausdorff surface distance HD(A,B) of the pair of partitions is calculated as follows:

[0149]

[0150] in, represents the Euclidean distance, a is the coordinate of the surface point in the first point set A, and b is the coordinate of the surface point in the second point set B. This indicator reflects the maximum local asymmetry between the two surfaces and can be used to evaluate the degree of symmetry of the jaw morphology between the left and right sides.

[0151] In addition, to further quantify the symmetry difference, the surface area of ​​each anatomical region was calculated, and the surface area difference and relative difference rate between the corresponding regions on the left and right sides were calculated using the following formula:

[0152] △S=|S left -S mirror-right |

[0153] RelDiff(%)=2△S / (S left +S mirror-right )×100%

[0154] Among them, S left and S mirror-right are the surface areas of the original left partition and the mirrored right partition, △S is the surface area difference, RelDiff(%) is the relative difference rate, and || indicates the absolute value operation.

[0155] Symmetry analysis of jaw volume includes:

[0156] The triangular mesh mandibular model was used to compare the volumes of each symmetrical contour partition, including the left and right articular regions, the left and right mandibular ascending ramus, the left and right mandibular angles, the left and right mandibular bodies, and the left and right alveolar processes. The volume △V of each symmetrical contour partition was calculated as follows:

[0157] △V=|V left -V right |

[0158] Among them, V left is the volume of the left contour partition, V right The volume of the right contour partition.

Claims

1. A method for detecting mandibular symmetry, characterized in that: The following steps are involved: Step 1: After training the feature point position prediction model using the mandibular feature point training dataset, a trained mandibular feature point positioning model is obtained; Step 2: After training the mandibular segmentation model using the jaw segmentation training dataset, a trained mandibular segmentation model is obtained; Step 3: Input the 3D CBCT maxillofacial image to be processed into the trained jaw feature point positioning model, and the model outputs a set of jaw feature points; The three-dimensional CBCT maxillofacial image to be processed is input into the trained mandibular segmentation model, and the model outputs the mandibular segmentation result. Then, the surface is reconstructed based on the mandibular segmentation result to obtain a triangular mesh mandibular model. Step 4: Perform symmetry analysis on the mandible based on the mandibular feature point set and the triangular mesh mandibular model to obtain the symmetry analysis results; Mandibular symmetry analysis includes the detection of mandibular dimensions, symmetry analysis of jaw contours, symmetry analysis of jaw surface distances and surface areas, and symmetry analysis of jaw volume. Symmetry analysis of the jaw contour includes: The mandibular contour is partitioned based on the points in the mandibular feature point set used for contour symmetry analysis and the triangular mesh mandibular model; Symmetry analysis of jaw surface distances and surface areas includes: Based on the midsagittal plane where the jaw feature points are concentrated, the triangular mesh mandibular model is mirrored along the midsagittal plane; the surface distance of each contour partition of the mandibular is compared in sequence; The surface area of ​​each anatomical subregion was also calculated, and the surface area difference and relative difference rate between the corresponding areas on the left and right sides were obtained; Symmetry analysis of jaw volume includes: The volume of each symmetrical contour partition was compared sequentially using a triangulated mandibular model.

2. The method for detecting mandibular symmetry according to claim 1, wherein: In step 1, the mandibular feature point training data set includes a low-resolution three-dimensional CBCT maxillofacial image and a corresponding high-resolution three-dimensional CBCT maxillofacial image, and a landmark point set, wherein the low-resolution and high-resolution three-dimensional CBCT maxillofacial images include complete upper and lower mandibles, the most anterior end of the nasal bone, left and right infraorbital points, and a complete C4 vertebral structure, and the landmark point set includes feature point coordinates for determining the midsagittal plane, feature point coordinates for cephalometric measurement, and feature point coordinates for mandibular morphological zoning and symmetry analysis.

3. The method for detecting mandibular symmetry according to claim 1, wherein: In step 1, the feature point position prediction model includes a first position prediction model and a second position prediction model. The low-resolution three-dimensional CBCT maxillofacial image is used as input to the first position prediction model, and the first position prediction model is used to predict the rough position of the jaw feature points; the high-resolution three-dimensional CBCT maxillofacial image and the rough position set of the jaw feature points predicted by the first position prediction model are used together as input to the second position prediction model, and the second position prediction model is used to predict the precise position of the jaw feature points.

4. The method for detecting mandibular symmetry according to claim 3, wherein: The first position prediction model includes a first image encoder, a multi-layer perceptron and a vector transformation module connected in sequence.

5. The method for detecting mandibular symmetry according to claim 4, characterized in that: In the vector transformation module, first, the feature vector output by the multilayer perceptron is split into an x-direction coordinate vector △x^, a y-direction coordinate vector △y^, a z-direction coordinate vector △z^, a linear transformation vector △R^, and a displacement vector t^; then, the linear transformation vector △R^ is deformed to obtain a linear transformation matrix R^; finally, according to the x-direction coordinate vector △x^, the y-direction coordinate vector △y^, the z-direction coordinate vector △z^, the displacement vector t^ and the transformation matrix R^ and the learnable template coordinate parameter C p Generate the predicted rough position of the jaw feature points using the following formula: C pred =C p ·R^ T +1 35 ·t^ T +[△x^ T △y^ T △z^ T ] Among them, C pred is the rough position of the jaw feature point predicted by the first position prediction model, 1 35 is a column vector of size 35*1, and T represents the matrix transpose operation.

6. The method for detecting mandibular symmetry according to claim 3, wherein: The second position prediction model includes a position preprocessing module, a relative position relationship generation module, a second image encoder, a feature extraction network and a self-attention module; the rough position set of the jaw feature points predicted by the first position prediction model is input into the position preprocessing module, the position preprocessing module outputs the transformed rough position set of the jaw feature points and the image subspace corresponding to all feature points, and the relative position relationship generation module generates a relative position matrix M according to the transformed rough position set of the jaw feature points. rel The second image encoder is connected to the feature extraction network. The image subspace corresponding to all feature points is used as the input of the second image encoder. The feature extraction network outputs the intermediate features corresponding to all feature points. The intermediate features corresponding to all feature points output by the feature extraction network form a feature matrix F. The feature matrix F and the relative position matrix M are combined. rel After splicing, the data are input into the self-attention module, which outputs the offset matrix N. Finally, the rough positions of each feature point in the rough position set of the transformed mandibular feature points are adjusted according to the offset matrix N to obtain the precise position of each feature point, thereby outputting the precise position set of the mandibular feature points.

7. The method for detecting mandibular symmetry according to claim 6, wherein: In the position preprocessing module, the rough position set of the jaw feature points predicted by the first position prediction model is first spatially transformed to obtain the transformed rough position set of the jaw feature points. The formula is as follows: C pred [i]=(C pred [i,1],C pred [i,2],C pred [i,3]) C L pred [i]= (C pred [i,1]×S x / 128,C pred [i,2]×S y / 128,C pred [i,3]×S z / 128) Among them, C pred [i] is the coordinate of the i-th feature point predicted by the first position prediction model, C pred [i,1], C pred [i,2], C pred [i,3] are the rough position set C of the jaw feature points predicted by the first position prediction model pred The values ​​of row i, column 1, row i, column 2, and row i, column 3, C L pred [i] is the initial coordinate of the i-th feature point in the high-resolution space after spatial transformation, S x 、S y 、S z is the resolution size of high-resolution three-dimensional CBCT maxillofacial images in the x, y, and z dimensions; Then, for the coarse position of each feature point in the transformed rough position set of mandibular feature points, the corresponding image subspace is intercepted in the high-resolution three-dimensional CBCT maxillofacial image with the coarse position of the feature point as the center. After traversing the coarse positions of all feature points in the transformed rough position set of mandibular feature points, the image subspace corresponding to all feature points is obtained.

8. The method for detecting mandibular symmetry according to claim 6, wherein: In the relative position relationship generation module, a relative position matrix M is established based on the relative position relationship of each feature point in the rough position set of the transformed jaw feature points. rel , relative position matrix M rel The values ​​of each element in satisfy the following formula: M rel [i, j]=C L pred [i,0]-C L pred [j,0] M rel [i, j+35]=C L pred [i,1]-C L pred [j,1] M rel [i, j+70]=C L pred [i,2]-C L pred [j,2] Among them, M rel [i,y] is the relative position matrix M rel The value of the i-th row and y-th column in the matrix, y=j, j+35 or j+70, C L pred [z,w] is the rough position set C of the transformed jaw feature points L pred The value in the zth row and wth column, where z=i or j, and w=0, 1, or 2.

9. A mandibular symmetry detection system, characterized in that: include: A mandibular feature point positioning unit, used to store a feature point position prediction model and train the feature point position prediction model using a mandibular feature point training data set; a mandibular segmentation unit, configured to store a mandibular segmentation model and train the mandibular segmentation model using a mandibular segmentation training dataset; The mandibular model reconstruction unit is used to input the 3D CBCT maxillofacial image to be processed into the trained mandibular segmentation model. The model outputs the mandibular segmentation result, and then performs surface reconstruction based on the mandibular segmentation result to obtain a triangular mesh mandibular model. A mandibular symmetry analysis unit is used to perform symmetry analysis on the mandibular according to the mandibular feature point set and the triangular mesh mandibular model to obtain a symmetry analysis result; Mandibular symmetry analysis includes the detection of mandibular dimensions, symmetry analysis of jaw contours, symmetry analysis of jaw surface distances and surface areas, and symmetry analysis of jaw volume. Symmetry analysis of the jaw contour includes: The mandibular contour is partitioned based on the points in the mandibular feature point set used for contour symmetry analysis and the triangular mesh mandibular model; Symmetry analysis of jaw surface distances and surface areas includes: Based on the midsagittal plane where the jaw feature points are concentrated, the triangular mesh mandibular model is mirrored along the midsagittal plane; the surface distance of each contour partition of the mandibular is compared in sequence; The surface area of ​​each anatomical subregion was also calculated, and the surface area difference and relative difference rate between the corresponding areas on the left and right sides were obtained; Symmetry analysis of jaw volume includes: The volume of each symmetrical contour partition was compared sequentially using a triangulated mandibular model.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the mandibular symmetry detection method according to any one of claims 1 to 8 are implemented.

Images