A three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities

Abstract

This invention discloses a three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities, relating to the fields of medical image processing and computer vision technology. The method includes: constructing a three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities; utilizing piecewise rigid transformation constraints based on weakly supervised learning to avoid predicting geometric deformations of bone tissue, enhancing the semantic correspondence of predicted maxillofacial bone tissue structures after orthognathic surgery; decomposing the non-rigid deformation field between images before and after orthognathic surgery into partially truncated rigid transformations of maxillary and mandibular bone tissues; generating a voxel-dense deformation field with good interpretability for the maxillofacial morphological changes brought about by the orthognathic process; and improving the accuracy of three-dimensional maxillofacial reconstruction during the orthognathic process of dentofacial deformities.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer vision technology, specifically to a three-dimensional maxillofacial reconstruction method based on cone-beam CT images of maxillofacial deformities. Background Technology

[0002] Orthognathic surgery for dentofacial deformities involves partially truncating and rigidly repositioning the upper and lower jawbones to eliminate the deformities and restore the functionality and aesthetics of the facial structure. Automatic maxillofacial reconstruction during orthognathic surgery can predict the morphology of bone and facial soft tissues after orthognathic surgery. Among these methods, CT image-based computer-aided modeling for maxillofacial reconstruction has gained attention in recent years.

[0003] Existing computer-aided modeling techniques for maxillofacial reconstruction require careful selection of specific templates or the construction of prior statistical shape models, but template selection relies heavily on the user's professional experience. Among existing methods, mirror registration-based methods select mirror-symmetric images of the normal side region as template images, but cannot handle bilateral defects and deformities. Methods based on sparse representation and correlation analysis can generate specific landmarks and reference models, but require time-consuming online iterative optimization processes. Computer-aided modeling of orthognathic procedures for dentofacial deformities based on 3D free deformation methods produces unnatural structural distortions, with partially truncated maxillary and mandibular structures exhibiting inappropriate non-rigid deformations. Rigid transformations can avoid distortion of bone tissue shape, but using only rigid transformations makes it difficult to establish a consistent correspondence between the maxillofacial structures before and after orthognathic surgery in specific situations. Therefore, existing technologies lack efficient 3D maxillofacial reconstruction methods specific to cone-beam computed tomography (CBCT) images for orthognathic procedures of dentofacial deformities. Summary of the Invention

[0004] This invention provides a three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities. The purpose is to overcome the shortcomings of the existing technology and to perform efficient three-dimensional reconstruction of cone-beam CT images of dentofacial deformities for the orthognathic process.

[0005] To achieve the above objectives, the three-dimensional maxillofacial reconstruction method proposed in this invention is a data-driven three-dimensional maxillofacial reconstruction method based on weakly supervised learning and segmented rigid constraints for the orthognathic process of dentofacial deformities. It introduces segmented rigid transformation constraints to enhance the semantic correspondence of maxillofacial structures in the three-dimensional image deformation before and after orthognathic surgery. It decomposes the non-rigid deformation field between cone-beam CT images of dentofacial deformities before and after orthognathic surgery into rigid transformations of partially truncated maxilla and mandible to reduce the need for dense voxel correspondence annotation, effectively overcoming the dependence on template selection and the problem of predicting bone tissue distortion. Considering that the changes in three-dimensional maxillofacial morphology caused by orthognathic surgery for dentofacial deformities are determined by the rigid movement of the partially truncated maxilla and mandible during the orthognathic process, the method proposed in this invention divides the skull into a static background region unaffected by the orthognathic process and a dynamic foreground region composed of a series of partially truncated maxillofacial bone tissue blocks. A rigid transformation prediction module is introduced to predict the rigid transformation parameters of each truncated bone tissue block during the orthognathic process. A piecewise rigid transformation constraint consistent with the predicted rigid transformation is applied to the global non-rigid deformation field, enhancing the semantic correspondence between maxillofacial bone tissue structures and avoiding inappropriate bone tissue distortion. This results in a voxel-dense deformation field with good interpretability and a reasonable prediction of the maxillofacial morphology after orthognathic surgery. In the testing phase, the three-dimensional maxillofacial reconstruction network for dentofacial deformities provided by this invention enables end-to-end maxillofacial reconstruction of the orthognathic process from cone-beam CT images of dentofacial deformities.

[0006] The technical solution provided by this invention is:

[0007] This invention provides a three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities. Utilizing piecewise rigid transformation constraints based on weakly supervised learning, it effectively avoids predicting geometric deformations of bone tissue, enhances the semantic correspondence of predicted maxillofacial bone structures after orthognathic surgery, and decomposes the non-rigid deformation field between images before and after orthognathic surgery into partially truncated rigid transformations of the maxilla and mandible. This generates a voxel-dense deformation field with good interpretability for the maxillofacial morphological changes brought about by the orthognathic process, improving the accuracy of three-dimensional maxillofacial reconstruction during the orthognathic process. The method includes the following steps:

[0008] 1) Construct a three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities;

[0009] The three-dimensional maxillofacial reconstruction network model for the orthognathic process of dentofacial deformity constructed in this invention includes a bone tissue segmentation module, an image feature extraction module, a deformation field inference module, and a rigidity transformation prediction module.

[0010] The bone tissue segmentation module employs a multi-level (four-level) convolutional 3D-Unet network structure with skip connections to extract bone tissue segmentation images from orthognathic anterior cone-beam CT images and orthognathic posterior cone-beam CT images.

[0011] The image feature extraction module and deformation field inference module specifically employ a three-dimensional convolutional neural network model architecture, which includes an image feature extraction module as the encoder and a deformation field inference module as the decoder. The current input of the model is only a single orthognathic anterior cone CT image. Given the input orthognathic anterior cone CT image, the image feature extraction module extracts image feature embeddings; the deformation field inference module generates a global voxel dense deformation field from the image feature embeddings. The resulting global voxel dense deformation field contains the displacement information of each voxel from the orthognathic anterior cone CT image to the orthognathic posterior cone CT image.

[0012] The rigid transformation prediction module consists of multiple rigid transformation prediction heads with identical structures; the number of rigid transformation prediction heads corresponds to the total number of mandibular and maxillofacial bone tissue blocks that are truncated and moved during orthognathic surgery for dentofacial deformities. Specifically, each rigid transformation prediction head comprises six 3D convolutional layers and one fully connected layer, used to predict the rotation and translation parameters of a corresponding truncated bone tissue block. Image features are embedded as input to the rigid transformation prediction head, which outputs a vector of predicted rotation and translation parameters for the rigid transformation of the truncated bone tissue block. It also generates a rigid transformation matrix and a rigid transformation deformation field equivalent to the predicted rigid transformation parameters to apply piecewise rigid transformation constraints to the global voxel-dense deformation field.

[0013] 2) Training a three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities;

[0014] The parameters of the three-dimensional maxillofacial reconstruction network model constructed in this invention are the weight parameters of all three-dimensional convolutional layers and fully connected layers contained in the bone tissue segmentation module, image feature extraction module, deformation field inference module, and rigidity transformation prediction module. The parameters of the three-dimensional maxillofacial reconstruction network model are trained and optimized using a weakly supervised learning method.

[0015] The training data includes paired premaxillary and postmaxillary cone-beam CT images and their corresponding bone tissue segmentation images.

[0016] The total loss function L is:

[0017] L = L rec +γ1L rig +γ2L mask (Formula 1)

[0018] Among them, L mask L represents the segmentation loss function. rec L represents the reconstruction loss function. rig Let γ1 and γ2 represent the piecewise rigid transformation loss function, where γ1 and γ2 are the weight hyperparameters for different loss functions.

[0019] The cross-entropy loss function is used for segmentation of multi-class bone tissue, and is expressed as:

[0020]

[0021] In Formula 2, k is the total number of bone tissue blocks of the maxilla and mandible truncated, N is the total number of voxels in the cone-beam CT image, and s i,j This represents the predicted probability that the j-th voxel in the cone-beam CT image belongs to the i-th truncated maxillary or mandibular bone tissue block, while Label the corresponding truth values.

[0022] The reconstruction loss function, used to train the network to predict the voxel dense deformation field, is composed of an image similarity metric based on mean square error and a smoothing constraint, and is expressed as:

[0023]

[0024] In Formula 3, V X To input orthognathic anterior cone tract CT images, V Y The image shows a true-valued orthognathic posterior cone-beam CT image, where α is the smoothing constraint coefficient and φ is the model-predicted global non-rigid deformation field, i.e., the global voxel dense deformation field. The symbol is... Indicates image distortion; The gradient of the global non-rigid deformation field φ; ||·|| F This represents the Frobenius norm.

[0025] A piecewise rigid transformation constraint is applied to the global non-rigid deformation field using a piecewise rigid transformation loss function, which is expressed as follows:

[0026]

[0027] In Formula 4, k represents the total number of bone fragments truncated from the maxilla and mandible, and V X and V Y These are cone-beam CT images taken before and after orthognathic surgery, respectively. and φ represents the binary mask images of the i-th truncated maxillary and mandibular bone tissue blocks in the premaxillary and postmaxillary bone tissue segmentation images, respectively; φ is the global non-rigid deformation field predicted by the deformation field inference module. β1 and β2 are the rigid deformation fields of the i-th truncated maxillary and mandibular bone tissue blocks predicted by the rigid transformation prediction module; β1 and β2 are the weight hyperparameters for balancing the loss function terms. The term represents the Hadamard product; 'd' in the third term represents the Dice loss function. In Equation 4, the first term constrains the model's predicted rigid transformation, accurately transforming the maxillary and mandibular bone tissue blocks truncated in the anterior orthognathic cone-beam CT image to their corresponding regions in the posterior orthognathic cone-beam CT image. The second term constrains the predicted global voxel dense deformation field to maintain consistency with the predicted rigid transformation in each truncated bone tissue block region. The combination of these two constraints ensures that the model's predicted global voxel dense deformation field performs a rigid transformation on each truncated maxillary and mandibular bone tissue block, thus preserving the semantic correspondence of the bone tissue structure without distortion during the transformation. The third term in Equation 4 constrains the anterior orthognathic bone tissue region to be consistent with the posterior orthognathic bone tissue region after the predicted transformation, further enhancing the model's prediction of the global voxel dense deformation field's maintenance of the semantic consistency of the bone tissue structure.

[0028] 3) Three-dimensional maxillofacial reconstruction for online orthognathic surgery for dentofacial deformities

[0029] During the online testing phase, the given anterior cone-beam CT image of the maxillofacial deformity is input into the three-dimensional maxillofacial reconstruction network model of the orthognathic process of dentofacial deformity trained in step 2). The network model output includes: the corresponding bone tissue segmentation image, the rigidity transformation parameters of each truncated maxillary and mandibular bone tissue block, the predicted global voxel dense deformation field, and the posterior cone-beam CT image of the maxillofacial deformity, thereby realizing three-dimensional maxillofacial reconstruction based on the cone-beam CT image of dentofacial deformity.

[0030] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0031] This invention provides a three-dimensional maxillofacial reconstruction method based on cone-beam computed tomography (CBCT) images of dentofacial deformities. In the three-dimensional maxillofacial reconstruction, a piecewise rigid transformation constraint is introduced to enhance the semantic correspondence of maxillofacial structures in the deformation of the three-dimensional images before and after orthognathic surgery, effectively overcoming the dependence on template selection and the problem of predicted bone tissue distortion. The skull is divided into a static background region unaffected by the orthognathic process and a dynamic foreground region composed of a series of partially truncated maxillary and mandibular bone tissue blocks during the orthognathic process. A rigid transformation prediction module is introduced to predict the rigid transformation parameters of each truncated bone tissue block during the orthognathic process. A piecewise rigid constraint consistent with the predicted rigid transformation is applied to the global non-rigid deformation field, enhancing the semantic correspondence between maxillofacial bone tissue structures and avoiding inappropriate bone tissue distortion. This results in a dense voxel deformation field with good interpretability and a reasonable prediction of the maxillofacial morphology after orthognathic surgery. This invention can improve the reconstruction accuracy of the three-dimensional maxillofacial structure during the orthognathic process of dentofacial deformities. Attached Figure Description

[0032] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0034] This invention provides a three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities. By establishing a three-dimensional maxillofacial reconstruction network model based on these images, three-dimensional maxillofacial reconstruction of dentofacial deformities is achieved. In the three-dimensional maxillofacial reconstruction, segmented rigid constraints are introduced to enhance the semantic correspondence of maxillofacial structures in the deformation of pre- and post-orthognathic 3D images, effectively overcoming the dependence on template selection and the problem of predicted bone tissue distortion. The skull is divided into a static background region unaffected by the orthognathic process and a dynamic foreground region composed of a series of partially truncated maxillary and mandibular bone tissue blocks during the orthognathic process. A rigid transformation prediction module is introduced to predict the rigid transformation parameters of each truncated bone tissue block during the orthognathic process. Segmented rigid constraints consistent with the predicted rigid transformation are applied to the global non-rigid deformation field, enhancing the semantic correspondence between maxillofacial bone tissue structures and avoiding inappropriate bone tissue distortion deformation, resulting in a dense voxel deformation field with good interpretability and reasonable post-orthognathic maxillofacial morphology prediction. This invention can improve the reconstruction accuracy of the three-dimensional maxillofacial structure during the orthognathic process of dentofacial deformities.

[0035] Addressing the problem of three-dimensional maxillofacial reconstruction during orthognathic surgery for dentofacial deformities, this invention utilizes piecewise rigid constraints based on weakly supervised learning to effectively avoid geometric deformation of predicted bone tissue, enhance the semantic correspondence of predicted maxillofacial bone tissue structure after orthognathic surgery, decompose the non-rigid deformation field between dentofacial images before and after orthognathic surgery into rigid transformations of partially truncated maxillary and mandibular bone tissue, and generate a voxel-dense deformation field with good interpretability for the changes in maxillofacial morphology brought about by the orthognathic process, thereby improving the image reconstruction accuracy of dentofacial deformities before and after orthognathic surgery for dentofacial deformities.

[0036] Figure 1 The diagram illustrates the overall flow of the three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities provided by this invention. The invention will be further explained below with reference to this diagram through an embodiment. This embodiment uses cone-beam CT images of dentofacial deformities, with a size of 128. 3 The physical size of the voxel is 1.2. 3 mm 3 In specific implementation, this invention first constructs a three-dimensional maxillofacial reconstruction network model based on cone-beam CT images of dentofacial deformities. Then, it optimizes the network model parameters using weakly supervised learning on a dataset consisting of paired pre- and post-orthognathic cone-beam CT images and their corresponding bone tissue segmentation images. After training, the resulting three-dimensional maxillofacial reconstruction network model can be used to perform three-dimensional maxillofacial reconstruction on any given cone-beam CT image of a dentofacial deformity. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities specifically includes the following steps:

[0037] 1) Construct a three-dimensional maxillofacial reconstruction network model based on cone-beam CT images of dentofacial deformities;

[0038] Considering that the changes in maxillofacial morphology caused by orthognathic surgery can be determined by a series of rigid transformations of the partially truncated maxillary and mandibular bone tissues during the orthognathic process, this invention proposes a three-dimensional reconstruction network model for orthognathic surgery of maxillofacial deformities, using truncated bone tissue blocks as basic units. To overcome the problem of bone tissue distortion caused by the voxel dense deformation field in free deformation, the network infers the rigid transformation parameters of each truncated and moved bone tissue block and applies piecewise rigid constraints on the voxel dense deformation field. The reconstruction network model constructed in this invention includes a bone tissue segmentation module, an image feature extraction module, a deformation field inference module, and a rigid transformation prediction module. The bone tissue segmentation module adopts a four-level convolutional 3D-Unet network structure with skip connections to extract bone tissue segmentation images from anterior and posterior orthognathic cone-beam CT images.

[0039] This invention employs a three-dimensional convolutional neural network architecture similar to the unsupervised medical image registration model VoxelMorph, which includes an image feature extraction module as the encoder and a deformation field inference module as the decoder. Unlike the registration model, which takes a volume image to be registered as input, the current input is only a single orthognathic anterior cone CT image. Given the input orthognathic anterior cone CT image, the image feature extraction module extracts image feature embeddings, and the deformation field inference module generates a global voxel-dense deformation field from the image feature embeddings, which contains displacement information of each voxel from the orthognathic anterior cone CT image to the orthognathic posterior cone CT image.

[0040] During orthognathic surgery, partially truncated maxillary and mandibular bone tissue blocks undergo rigidity transformations. Therefore, this invention constructs a rigidity transformation prediction module composed of multiple structurally identical rigidity transformation prediction heads. The number of rigidity transformation prediction heads corresponds to the total number of truncated and moved maxillary and mandibular bone tissue blocks during the orthognathic procedure. Each rigidity prediction head consists of six 3D convolutional layers and one fully connected layer, used to predict the rotation and translation parameters of a corresponding truncated bone tissue block. Given image feature embeddings, the rigidity transformation prediction head outputs the predicted rotation parameters and translation vectors of the rigidity transformation of the truncated bone tissue block, generating the corresponding rigidity transformation matrix and rigidity transformation deformation field.

[0041] 2) Train a three-dimensional reconstruction network model for orthognathic surgery of dentofacial deformities;

[0042] The proposed three-dimensional reconstruction network model for orthognathic deformities in this invention employs weakly supervised learning to train and optimize network parameters. Training data includes paired pre- and post-orthognathic cone-beam CT images and their corresponding bone tissue segmentation images. The cross-entropy loss function is selected as the training loss function for multi-class bone tissue segmentation.

[0043]

[0044] In Formula 2, k is the total number of bone tissue blocks of the maxilla and mandible truncated, N is the total number of voxels in the cone-beam CT image, and s i,j This represents the predicted probability that the j-th voxel in the cone-beam CT image belongs to the i-th truncated maxillary or mandibular bone tissue block, while The corresponding ground truth labels are provided. To train the network to predict the voxel dense deformation field, the reconstruction loss function, which is a combination of image similarity metrics based on mean square error and smoothing constraints, is optimized:

[0045]

[0046] In Formula 3, V X To input orthognathic anterior cone tract CT images, V Y The image shows a true-valued posterior cone-beam CT image of the maxilla, where α is the smoothing constraint coefficient and φ is the model-predicted global non-rigid deformation field. Indicates image distortion. The gradient of the global non-rigid deformation field φ; ||·|| F This represents the Frobenius norm. Since an unrestricted free deformation field can lead to unreasonable geometric distortions in bone tissue regions, including truncated maxillary and mandibular bone blocks, this method applies piecewise rigid transformation constraints to the global non-rigid deformation field by introducing a piecewise rigid transformation loss function.

[0047]

[0048] In Formula 4, k represents the total number of bone fragments truncated from the maxilla and mandible, and V X and V Y These are cone-beam CT images taken before and after orthognathic surgery. and φ represents the binary mask images of the i-th truncated maxillary and mandibular bone tissue blocks in the pre- and post-orthognathic bone tissue segmentation images, respectively, and φ is the global non-rigid deformation field predicted by the deformation field inference module. Let β1 and β2 be the rigid deformation field of the i-th truncated maxillary and mandibular bone tissue block predicted by the rigidity prediction module, and let β1 and β2 be the weight hyperparameters balancing the loss function terms. This represents the Hadamard product. The first term in Equation 3 constrains the rigid transformation parameters predicted by the model, enabling the rigid transformation of the truncated bone tissue block in the anterior conical CT image of the orthognathic jaw to the corresponding position after orthognathic surgery. The second term constrains the global non-rigid deformation field to be consistent with the predicted rigid deformation field in the corresponding truncated bone tissue block region. In the third term, d represents the Dice loss function, calculated as follows:

[0049]

[0050] Binary segmentation images of truncated maxillary and mandibular bone tissue blocks are used to constrain the bone tissue structure of the anterior cone-beam CT image of the orthognathic jaw to maintain semantic consistency with the posterior orthognathic jaw after transformation.

[0051] The final loss function L is:

[0052] L = L rec +γ1L rig +γ2L mask (Formula 1)

[0053] Where L mask L represents the segmentation loss function. rec L represents the reconstruction loss function. rig Let L represent the piecewise rigid transformation loss function, and γ1 and γ2 be the weight hyperparameters for different loss functions. The model is trained by optimizing the final loss function L.

[0054] 3) Online reconstruction of orthognathic posterior cone-beam CT images for dentofacial deformities;

[0055] During the online testing phase, the given anterior cone-beam CT image of the maxilla is input into the three-dimensional reconstruction network model of the orthognathic deformity obtained in step 2). The network outputs the corresponding bone tissue segmentation image, the rigidity transformation parameters of each truncated maxillary and mandibular bone tissue block, the predicted global deformation field, and the predicted posterior cone-beam CT image of the maxilla.

[0056] In specific implementation, the given embodiment size is 128. 3 The voxel size is 1.2. 3 mm 3 cone-beam CT images of dentofacial deformities V X The data is input into a three-dimensional reconstruction network model for orthognathic surgery of dentofacial deformities. The rigid transformation prediction module outputs the rotation and translation parameters of each truncated maxillary and mandibular bone tissue block, while the deformation field inference module outputs the predicted global non-rigid deformation field φ. The predicted global non-rigid deformation field is then applied to the input cone-beam CT image of dentofacial deformities V. X The upper deformation yields the predicted orthognathic posterior cone-beam CT image.

[0057] To verify the effectiveness of the method of the present invention, experiments were conducted on the acquired cone-beam CT images of dentofacial deformities. The three-dimensional maxillofacial reconstruction of the orthognathic process was tested using cone-beam CT images of dentofacial deformities. It can generate a voxel dense deformation field with good interpretability and effectively predict the changes in maxillofacial morphology caused by orthognathic surgery and the three-dimensional maxillofacial morphology after orthognathic surgery from the acquired cone-beam CT images of dentofacial deformities.

[0058] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A three-dimensional maxillofacial reconstruction method based on cone-beam computed tomography (CBCT) images of dentofacial deformities, characterized in that, A three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities was constructed. Using piecewise rigid transformation constraints based on weakly supervised learning, the geometric deformation of predicted bone tissue was avoided, enhancing the semantic correspondence of predicted maxillofacial bone tissue structure after orthognathic surgery. The non-rigid deformation field between images before and after orthognathic surgery was decomposed into rigid transformations of partially truncated maxillary and mandibular bone tissues, generating a global voxel-dense deformation field with good interpretability for the maxillofacial morphological changes brought about by the orthognathic process, thus improving the accuracy of three-dimensional maxillofacial reconstruction during the orthognathic process of dentofacial deformities. The model includes the following steps: 1) Construct a three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities; The three-dimensional maxillofacial reconstruction network model includes a bone tissue segmentation module, an image feature extraction module, a deformation field inference module, and a rigidity transformation prediction module. The bone tissue segmentation module employs a multi-level convolutional network structure with skip connections to extract bone tissue segmentation images from orthognathic anterior cone-beam CT images and orthognathic posterior cone-beam CT images. The image feature extraction module and the deformation field inference module adopt a three-dimensional convolutional neural network model architecture. The encoder in the network model is the image feature extraction module, and the decoder is the deformation field inference module. The model takes only a single orthognathic anterior cone CT image as input. Image feature embedding is extracted by the image feature extraction module. A global voxel dense deformation field is generated from the image feature embedding by the deformation field inference module. The resulting global voxel dense deformation field contains displacement information of each voxel from the orthognathic anterior cone CT image to the orthognathic posterior cone CT image. The rigid transformation prediction module includes multiple rigid transformation prediction heads with the same structure; the number of rigid transformation prediction heads is the total number of maxillary and mandibular bone tissue blocks that are cut off and moved during the orthognathic surgery for dentofacial deformities; the rigid transformation prediction head is used to predict the rotation and translation parameters of a corresponding cut bone tissue block. Image features are embedded as input to the rigid transformation prediction head, which outputs a vector of rotation and translation parameters for predicting the rigid transformation of the truncated bone tissue block. It also generates a rigid transformation matrix and a rigid transformation deformation field equivalent to the predicted rigid transformation parameters to apply piecewise rigid transformation constraints to the global voxel dense deformation field. 2) Train a three-dimensional maxillofacial reconstruction network model for orthognathic surgery of dentofacial deformities; use weakly supervised learning to train and optimize the parameters of the three-dimensional maxillofacial reconstruction network model; the model parameters include: the weight parameters of all three-dimensional convolutional layers and fully connected layers contained in the model; The training data consisted of paired premaxillary and postmaxillary cone-beam CT images and their corresponding bone tissue segmentation images. The total loss function L is expressed as: L = L rec +γ1L rig +γ2L mask Formula 1 Among them, L mask L represents the segmentation loss function. rec L represents the reconstruction loss function. rig Let γ1 and γ2 represent the piecewise rigid transformation loss function, where γ1 and γ2 are the weight hyperparameters for different loss functions. The cross-entropy loss function is selected as the segmentation loss function for multi-class bone tissue. The reconstruction loss function is used to train the network to predict the global voxel dense deformation field, and is composed of an image similarity measure based on mean square error and a smoothing constraint. A piecewise rigid transformation constraint is applied to the global non-rigid deformation field using a piecewise rigid transformation loss function. The piecewise rigid transformation loss function consists of three terms: the first term constrains the rigid transformation predicted by the model to transform the maxillary and mandibular bone tissue block regions truncated in the anterior orthognathic cone-beam CT image to the corresponding regions in the posterior orthognathic cone-beam CT image; the second term constrains the predicted global voxel dense deformation field to maintain consistency with the predicted rigid transformation in each truncated bone tissue block region; and the third term constrains the anterior orthognathic bone tissue region to be consistent with the posterior orthognathic bone tissue region after the predicted transformation, thereby enhancing the semantic consistency of the model's predicted global voxel dense deformation field with respect to bone tissue structure. 3) Three-dimensional maxillofacial reconstruction for online orthognathic surgery for dentofacial deformities During the online testing phase, the anterior cone-beam CT images of the maxilla were input into the three-dimensional maxillofacial reconstruction network model trained in step 2). The network model output included: rigid transformation parameters of each truncated maxillary and mandibular bone tissue block, the predicted global voxel dense deformation field, and the posterior cone-beam CT images of the maxilla. Through the above steps, three-dimensional maxillofacial reconstruction based on cone-beam CT images of dentofacial deformities is achieved.

2. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 1, characterized in that, The bone tissue segmentation module specifically adopts a four-level convolutional 3D-Unet network structure with skip connections.

3. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 1, characterized in that, Each rigid transformation prediction head consists of 6 three-dimensional convolutional layers and 1 fully connected layer.

4. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 1, characterized in that, In step 2), the segmentation loss function for multi-class bone tissue is expressed as: In Formula 2, k is the total number of bone tissue blocks of the maxilla and mandible truncated, N is the total number of voxels in the cone-beam CT image, and s i,j The predicted probability that the j-th voxel in the cone-beam CT image belongs to the i-th truncated maxillary and mandibular bone tissue block is represented. Label the corresponding truth values.

5. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 4, characterized in that, In step 2), the reconstruction loss function is expressed as: In Formula 3, V X To input orthognathic anterior cone tract CT images, V Y The image shows a true-valued orthognathic posterior cone-beam CT image, where α is the smoothing constraint coefficient and φ is the model-predicted global non-rigid deformation field, i.e., the global voxel dense deformation field. The symbol is... Indicates image distortion; The gradient of the global non-rigid deformation field φ; ‖·‖ F This represents the Frobenius norm.

6. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 5, characterized in that, In step 2), the piecewise rigid transformation loss function is expressed as: k is the total number of bone fragments cut from the maxilla and mandible, V X and V Y These are cone-beam CT images taken before and after orthognathic surgery, respectively. and These are binary mask images representing the i-th truncated maxillary and mandibular bone tissue blocks in the premaxillary and postmaxillary bone tissue segmentation images, respectively. φ represents the global non-rigid deformation field predicted by the deformation field inference module; β1 and β2 are the rigid deformation fields of the i-th truncated maxillary and mandibular bone tissue blocks predicted by the rigid transformation prediction module; β1 and β2 are the weight hyperparameters for balancing the loss function terms. It represents the Hadamardi (or Hadama) stack.

7. The three-dimensional maxillofacial reconstruction method based on cone-beam CT images of dentofacial deformities as described in claim 1, characterized in that, The first and second terms in the piecewise rigid transformation loss function are combined so that the model predicts a global voxel dense deformation field that performs a rigid transformation on each truncated maxillary and mandibular bone tissue block, thereby maintaining the semantic correspondence of the bone tissue structure without distortion during the transformation.

Images