Postoperative facial soft tissue morphology prediction method and system based on conditional deformation field completion

Abstract

The application discloses a postoperative facial soft tissue shape prediction method and system based on a conditional deformation field completion, first acquires bone and soft tissue data, including dense preoperative soft tissue point cloud, sparse preoperative soft tissue landmark points and sparse preoperative bone landmark points, and the three are combined to construct a data set; an initial displacement vector of each point in the data set is extracted; then a deformation field completion network containing an information propagator and a prediction head is constructed, the information propagator is used for propagating the known initial displacement vector on the bone landmark points to the soft tissue point cloud and the soft tissue landmark points with zero initial displacement along the inherent geometric manifold of the facial point cloud; the prediction head predicts a three-dimensional displacement vector for each soft tissue point cloud and soft tissue landmark point based on the information propagator result, and further predicts a postoperative coordinate. The application more effectively learns the spatial topological relationship between bones and soft tissues, improves the personalized fidelity of prediction, and can support real-time and interactive surgical plan planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided design, and in particular to a postoperative facial soft tissue morphology prediction method and system based on conditional deformation field completion. BACKGROUND

[0002] Orthognathic surgery aims to correct dental and maxillofacial deformities by precisely reshaping craniofacial bone structures, thereby restoring the patient's physiological functions (such as occlusion, breathing) and balanced facial proportions. Accurate prediction of the final impact of the surgical plan on facial soft tissue morphology has become a core link and key challenge in modern computer-aided surgical planning (CASS).

[0003] A precise prediction model is of vital clinical value in assisting doctors in developing optimal surgical plans, facilitating doctor-patient communication, and managing patient postoperative expectations. Currently, digital virtual surgical planning (VSP) has become a clinical standard, and its process usually involves simulating bone cutting and moving on a three-dimensional model reconstructed from CBCT (cone beam CT) data, and simulating soft tissue response.

[0004] However, there is a highly nonlinear, individualized complex relationship between bone movement and soft tissue response, which is dominated by biomechanics. The complexity of this relationship stems from the nonlinear response of soft tissue, tissue heterogeneity, and the non-rigid linkage effect between muscles, making accurate prediction extremely challenging.

[0005] In recent years, based on deep learning, especially geometric deep learning models with point cloud as input, has become the mainstream research direction in this field, and has formed a dual-flow guided technical paradigm. Its typical technical solution is as follows:

[0006] 1. Input data: This paradigm usually requires three independent, dense point cloud models as input: (1) preoperative soft tissue point cloud; (2) preoperative bone point cloud; (3) dense postoperative bone point cloud generated according to the surgical plan.

[0007] 2. Network architecture: Use a dual-branch (dual-flow) neural network architecture:

[0008] Main Stream: Use a powerful point cloud encoder (such as PointNet++, DGCNN) to process the input dense preoperative soft tissue point cloud to extract its deep-level geometric features.

[0009] Conditional Stream: Use an independent encoder to process the preoperative and postoperative dense bone point clouds to calculate the bone displacement information and encode it into a conditional guide vector.

[0010] 3. Feature fusion: After the feature extraction of the two branches is completed respectively, the guiding information of the skeleton and the feature information of the soft tissue are combined through an independent fusion module (such as feature splicing or cross attention mechanism).

[0011] 4. Displacement prediction: finally, the fused features are sent into a prediction head (usually MLP), and the three-dimensional displacement vector of each soft tissue point is regressed to obtain the predicted postoperative morphology.

[0012] However, the above technology has the following two major shortcomings:

[0013] 1. Design bottleneck, limited prediction accuracy: this paradigm artificially separates the soft tissue and the skeleton, which are in the same physical space, and independently encodes the features in their respective streams. This causes the valuable spatial topological relationship between the two to be lost early. The subsequent fusion module needs to painstakingly reconstruct this complex relationship in two highly abstract feature spaces, which not only increases the learning burden, but also often leads to insufficient fusion, making the guiding signal of the skeleton ambiguous and generalized. The final prediction result tends to be average, that is, the postoperative facial morphology of each patient is similar, and it is difficult to capture the fine and personalized soft tissue response caused by specific bone movement.

[0014] 2. Violate the clinical workflow, low efficiency: this paradigm relies on a dense, simulation-generated postoperative bone point cloud as a key input. In clinical practice, surgeons need to iteratively and finely adjust the virtual surgical plan (for example, fine-tune the mandibular movement from 5mm to 6mm). Under the existing paradigm, each fine-tuning must return to the CASS software to re-operate and generate a new dense postoperative bone model before prediction can be performed again. This "adjust-rebuild-predict" cycle is cumbersome and time-consuming, and cannot support real-time and interactive plan exploration and optimization for doctors, which is contrary to the core demand of high efficiency in clinical practice. SUMMARY

[0015] The present application aims to solve the problems of the prior art, and proposes a postoperative facial soft tissue morphology prediction method and system based on conditional deformation field completion, which solves the fusion bottleneck problem caused by information separation coding in the "double flow guiding" paradigm and the low efficiency of the clinical workflow caused by the dependence of the existing method on the dense postoperative bone model as input, aims to avoid the subsequent difficult feature fusion step, so as to more effectively learn the spatial topological relationship between bones and soft tissues, improve the personalized fidelity of prediction, and propose a prediction method that only needs sparse and easily obtained clinical surgical parameters to drive, to support real-time and interactive surgical plan planning.

[0016] The application aims to realize the technical scheme of a postoperative facial soft tissue morphology prediction method based on a conditional deformation field completion, which comprises the following steps:

[0017] (1) Obtain bone and soft tissue data, including dense preoperative soft tissue point cloud, sparse preoperative soft tissue landmark points and sparse preoperative bone landmark points, and combine the three to construct a data set;

[0018] (2) Extract the initial displacement vector of each point in the data set, and set a mask for each point to represent the type of the point;

[0019] (3) Construct a deformation field completion network comprising an information propagator and a prediction head, wherein the information propagator is used to propagate the known initial displacement vector on the bone landmark points along the inherent geometric manifold of the facial point cloud to the soft tissue point cloud and the soft tissue landmark points with zero initial displacement; and the prediction head predicts a three-dimensional displacement vector for the soft tissue point cloud and the soft tissue landmark points based on the information propagator result;

[0020] (4) Calculate the predicted postoperative coordinates of the soft tissue point cloud and the soft tissue landmark points in the data set based on the predicted three-dimensional displacement vector.

[0021] Further, in step (2), for the bone landmark points, the initial displacement vector is obtained based on the difference between the postoperative coordinates and the preoperative coordinates.

[0022] Further, in step (2), for all soft tissue point clouds and soft tissue landmark points, the initial displacement vector is set to zero.

[0023] Further, in step (2), the mask is 0 for the preoperative soft tissue points, 1 for the bone landmark points, and 2 for the soft tissue landmark points.

[0024] Further, in step (3), the information propagator adopts an information propagation mechanism of a dynamic graph convolution network as a deformation field propagator,

[0025] Further, the information propagator adopts a stacked more EdgeConv module to construct a deeper information propagation path.

[0026] Further, the bone landmark points are used as the known condition source of the deformation field, and the soft tissue landmark points are used as the supervision or evaluation target.

[0027] Further, the deformation field completion network adopts a composite loss function for supervised training, including:

[0028] (1) Overall shape supervision: adopt chamfer distance loss to ensure that the predicted point cloud is aligned with the real postoperative point cloud in macro shape;

[0029] (2) Geometric regularization loss: Laplacian smoothing loss is introduced to promote surface smoothness, and edge length consistency loss is introduced to maintain the consistency of edge length in the local neighborhood.

[0030] In another aspect, the present application also provides a postoperative facial soft tissue morphology prediction system based on conditional deformation field completion, which comprises:

[0031] A point cloud acquisition module is configured to acquire bone and soft tissue data, including dense preoperative soft tissue point cloud, sparse preoperative soft tissue landmark points and sparse preoperative bone landmark points, which are combined to construct a data set;

[0032] A displacement extraction module is configured to extract an initial displacement vector of each point in the data set, and set a mask for each point to represent the type of the point;

[0033] A network construction module is configured to construct a deformation field completion network comprising an information propagator and a prediction head, the information propagator is configured to propagate the known initial displacement vector on the bone landmark points along the inherent geometric manifold of the facial point cloud to the soft tissue point cloud and the soft tissue landmark points with zero initial displacement, and the prediction head is configured to predict a three-dimensional displacement vector for the soft tissue point cloud and the soft tissue landmark points based on the result of the information propagator;

[0034] A coordinate prediction module is configured to calculate the predicted postoperative coordinates of the soft tissue point cloud and the soft tissue landmark points in the data set based on the predicted three-dimensional displacement vector.

[0035] The present application has the following advantages:

[0036] 1. Higher precision and better fidelity: The present application replaces the "separated guidance" with "integrated completion", and processes the bone and soft tissue in a shared geometric space. The displacement signal propagates naturally along the inherent geometric structure, avoiding the bottleneck of external feature fusion. This enables the model to better capture the fine and personalized soft tissue response caused by specific bone movement, effectively avoiding the problem of "averaging" the prediction results.

[0037] 2. High clinical efficiency and real-time interaction support: One of the most prominent advantages of the present application is its clinical practicability. It only needs sparse bone landmark point displacement as input (which is exactly the parameter directly operated by the clinician when planning the operation), and does not need to generate dense postoperative bone point cloud at all. This fundamentally solves the cumbersome "adjust-rebuild-predict" process of the prior art, and changes the planning tool from a static verifier to a dynamic and interactive exploration tool. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to make the technical solutions of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only need to explain the present application, and other drawings can be obtained by those skilled in the art without any creative work.

[0039] Figure 1 The method flowchart of the present application.

[0040] Figure 2 The schematic diagram of the deformation field completion network.

[0041] Figure 3 The comparison diagram of the predicted result and the real postoperative result. DETAILED DESCRIPTION

[0042] In order to make the technical solutions of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only need to explain the present application, and other drawings can be obtained by those skilled in the art without any creative work.

[0043] As Figure 1 shown, the present application provides a postoperative facial soft tissue shape prediction method based on conditional deformation field completion. The postoperative facial soft tissue prediction task is redefined from "separated double-flow guidance" to "integrated conditional deformation field completion". The method no longer regards the bone displacement as an external guidance signal, but regards it as a known initial condition in a unified geometric field, and propagates and completes the signal in the field through a specially designed network. The specific steps are as follows:

[0044] 1. Construction of unified input point cloud

[0045] The present application first constructs a unified geometric entity, integrates the sparse bone landmark points with known displacement and the dense soft tissue point cloud to be predicted into the same shared coordinate space. The dense preoperative soft tissue point cloud containing N points (for example, N=4096 points) and the sparse preoperative soft tissue landmark points (for example, 22) and the sparse preoperative bone landmark points (for example, 21) are combined into a unified input point cloud P containing N+M points (for example, N=4139).

[0046] Wherein:

[0047] Preoperative soft tissue landmark points include: tip of nose point, alar point (left / right), alar lateral point (left / right), alar base point (left / right), nasolabial sulcus middle point (left / right), nasolabial sulcus inner point (left / right), subnasale point, upper lip point, lower lip point, lip aperture point, cheilion point (left / right), soft tissue B point, soft tissue prementon point, soft tissue submenton point, soft tissue gonion point (left), soft tissue gonion point (right) a total of 22. These landmark points are soft tissue landmark points located on the surface of the facial skin, points with clear anatomical significance, which directly determine our facial appearance.

[0048] Preoperative bone landmark points include: piriform aperture lateral margin (left / right), anterior nasal spine point, upper alveolar seat point, upper central incisor midpoint, upper canine point (left / right), upper first molar point (left / right), lower central incisor midpoint, lower canine point (left / right), lower first molar point (left / right), mental foramen point (left / right), lower alveolar seat point, prementon point, submenton point, gonion point (left / right) a total of 21, as known points of the deformation field, the displacement of these bone landmark points can represent the influence of orthognathic surgery on the entire bone morphology.

[0049] 2. Construction of 7-dimensional input features

[0050] To unify each point p in the input point cloud P i , a 7-dimensional feature vector [x, y, z, dx, dy, dz, m] is constructed.

[0051] (x, y, z): preoperative three-dimensional coordinates pos i of the point.

[0052] (dx, dy, dz): initial displacement vector disp i of the point. This is one of the keys of the present application.

[0053] For bone landmark points, the vector is the exact postoperative displacement (i.e. postoperative coordinates - preoperative coordinates) known according to the surgical plan.

[0054] For all soft tissue points (including ordinary point clouds and landmark points), the vector is initialized to a zero vector (0, 0, 0).

[0055] m: a three-value mask mask i used to distinguish the type of point in the network.

[0056] m=0: represents a point in the preoperative soft tissue point cloud.

[0057] m=1: represents a bone landmark point (as a known condition source of the deformation field).

[0058] m=2: represents a soft tissue landmark point (which can be used as a special supervision or evaluation target).

[0059] 3. Processing of Deformation Field Completion Network (DFCN)

[0060] like Figure 2 As shown, the constructed unified input, containing N points and their 7-dimensional features, is fed into the Deformation Field Completion Network (DFCN) designed in this invention. The DFCN consists of two core modules:

[0061] (1) DGCNNPropagator (Information Propagator):

[0062] 1) Core Function: This module is the core of this invention. Its function is no longer general feature extraction, but rather a task-specific "deformation field propagator". Its role is to "propagate" or "interpolate" the known initial displacement signals (dx, dy, dz) on the skeletal landmarks layer by layer, naturally, along the inherent geometric manifold of the facial point cloud (defined by the proximity relationships between points) to all soft tissue point clouds and landmarks with zero initial displacement.

[0063] 2) Network selection: Dynamic Graph Convolutional Network (DGCNN) is adopted as the core information propagation mechanism.

[0064] 3) Key structural modifications:

[0065] Deepen the network: By stacking more (e.g., 4 layers) EdgeConv convolutional modules, a deeper information propagation path is constructed to ensure that even distant skeletal landmarks can have their displacement signals propagated to every soft tissue point through multi-hop neighbor relationships.

[0066] Global feature module removed: The global max pooling and feature broadcasting steps used at the end of the network in traditional DGCNN for classification or segmentation tasks are completely removed. This is because in the "completeness" task of this invention, an averaged, blurry global feature will "dilute" and "interfere" with the localized, refined deformation signals from specific skeleton points that we want to accurately propagate.

[0067] (2) PredictionHead:

[0068] 1) This module receives high-dimensional features output by DGCNNPropagator, which have already incorporated the propagated displacement information.

[0069] 2) It is a simple network consisting of a multilayer perceptron (MLP, which can be implemented by 1D convolution).

[0070] 3) Its function is to regress a three-dimensional displacement correction for each point (including soft tissue point cloud and landmark points) This correction is a fine-tuning adjustment to the initial propagation / interpolation result.

[0071] 4. Final coordinate output

[0072] The invention adopts the residual learning method to calculate the final predicted postoperative coordinates. For each soft tissue point cloud and soft tissue landmark point in the input:

[0073] Initial coordinates: C_pre = (x, y, z)

[0074] Network predicted displacement correction: Deltad

[0075] Final predicted postoperative coordinates: C_post = C_pre + Deltad.

[0076] 5. Supervision of composite loss function

[0077] In order to ensure the accuracy of the predicted macro shape and the physical reasonableness of the local geometric details, the invention adopts a composite loss function for training, including:

[0078] (1) Overall shape supervision: Chamfer Distance Loss is used to ensure that the predicted point cloud is aligned with the true postoperative point cloud in terms of macro shape.

[0079] (2) Geometric regularization loss: Laplacian Loss is introduced to promote surface smoothing and prevent unnatural spikes; and Edge Length Loss is introduced to maintain the consistency of edge length within the local neighborhood, simulate the anti-stretching / compression characteristics of soft tissue, and improve the physical realism.

[0080] Experimental results:

[0081] In order to compare the accuracy of the prediction method provided by the invention based on the condition deformation field completion and the traditional double-flow guided prediction method based on separation in postoperative facial soft tissue shape prediction. Among them, the double-flow guided prediction method based on separation for comparison includes GF-Net which adopts global feature concatenation as the feature fusion method and AG-Net which adopts cross-attention guidance as the feature fusion method.

[0082] The evaluation method is divided into four specific evaluation indexes under two levels, including point cloud level error: chamfer distance (CD) and Hausdorff distance (HD), and surface level set fidelity: average surface distance (ASD) and 2mm success detection rate (SDR@2mm). Among them, the point cloud level error compares the prediction result and the actual postoperative soft tissue form on the form of the point cloud directly output by the prediction method, chamfer distance (CD) is used to reflect the overall shape similarity between the two point clouds; Hausdorff distance (HD) is used to measure the maximum mismatch between the two point clouds, which is sensitive to the maximum deviation. The surface level geometry fidelity further reconstructs the predicted point cloud into a continuous surface, and then quantifies the difference between the prediction result and the actual postoperative form, wherein the average surface distance (ASD) calculates the average distance between the two surfaces, and the unit is millimeter. The smaller the ASD value, the higher the overall fit of the predicted soft tissue surface and the actual postoperative surface. The 2mm success detection rate (SDR@2mm) calculates the proportion of the area on the actual soft tissue surface whose prediction error is less than a clinically acceptable error tolerance threshold, which is selected as 2mm here.

[0083] Table 1

[0084]

[0085] As can be seen from Table 1, the DFCN model under the deformation field completion paradigm proposed by the present application has certain advantages in various evaluation indexes. In the core index directly reflecting the clinical value, the average surface distance (ASD) of the DFCN prediction is 2.08 ± 0.49 millimeters, and the area with an error less than 2mm (SDR@2mm) accounts for 76.61%. These values show that the prediction result of DFCN is overall consistent with the actual postoperative form, and the error of most surface areas is controlled within the clinically acceptable 2mm threshold. At the same time, the chamfer distance (CD) and Hausdorff distance (HD) and other point cloud geometry indexes are also better than the comparative model, reflecting that the generated results have good fidelity and stability in geometric form.

[0086] When compared with the representative model based on the "separated guidance" paradigm, the performance of DFCN is better. For example, the ASD error of DFCN is reduced by about 3.7% and 4.6% compared with AG-Net and GF-Net respectively. This result shows that there may be certain limitations in the traditional "separated guidance" paradigm when dealing with the prediction task driven by sparse bone points, and the "deformation field completion" paradigm proposed by the present application provides an effective idea for accurate prediction using sparse driving signals by modeling the information propagation process by regarding bone-soft tissue as a unified entity.

[0087] As Figure 3The figure shows the comparison between the prediction results of the DFCN network and the real postoperative results on a representative case, where the red point cloud is the real postoperative soft tissue point cloud and the blue point cloud is the postoperative soft tissue point cloud predicted by the DFCN. Figure 3 The left, middle and right are the front view, side view and enlarged local view of the lip and chin part, respectively. From the front view (the left part in FIG. 8) for showing the overall contour, it can be seen that the postoperative soft tissue morphology predicted by the DFCN is basically consistent with the real postoperative soft tissue morphology in overall contour, and there is no enlargement or reduction. The prediction accuracy of the front and back direction movement of the face caused by the orthognathic surgery can be observed from the side view (the middle part in FIG. 8), and it is found that the curve of the prediction result of the DFCN is basically consistent with the side curve of the real postoperative soft tissue. Figure 3 The left, middle and right are the front view, side view and enlarged local view of the lip and chin part, respectively. From the front view (the left part in FIG. 8) for showing the overall contour, it can be seen that the postoperative soft tissue morphology predicted by the DFCN is basically consistent with the real postoperative soft tissue morphology in overall contour, and there is no enlargement or reduction. The prediction accuracy of the front and back direction movement of the face caused by the orthognathic surgery can be observed from the side view (the middle part in FIG. 8), and it is found that the curve of the prediction result of the DFCN is basically consistent with the side curve of the real postoperative soft tissue. Figure 3 The left, middle and right are the front view, side view and enlarged local view of the lip and chin part, respectively. From the front view (the left part in FIG. 8) for showing the overall contour, it can be seen that the postoperative soft tissue morphology predicted by the DFCN is basically consistent with the real postoperative soft tissue morphology in overall contour, and there is no enlargement or reduction. The prediction accuracy of the front and back direction movement of the face caused by the orthognathic surgery can be observed from the side view (the middle part in FIG. 8), and it is found that the curve of the prediction result of the DFCN is basically consistent with the side curve of the real postoperative soft tissue. Figure 3 The left, middle and right are the front view, side view and enlarged local view of the lip and chin part, respectively. From the front view (the left part in FIG. 8) for showing the overall contour, it can be seen that the postoperative soft tissue morphology predicted by the DFCN is basically consistent with the real postoperative soft tissue morphology in overall contour, and there is no enlargement or reduction. The prediction accuracy of the front and back direction movement of the face caused by the orthognathic surgery can be observed from the side view (the middle part in FIG. 8), and it is found that the curve of the prediction result of the DFCN is basically consistent with the side curve of the real postoperative soft tissue.

[0088] The above examples are used to explain and illustrate the present application, but not to limit the present application, and any modification and change made to the present application within the spirit and protection scope of the claims of the present application, fall into the protection scope of the present application.

Claims

1. A method for postoperative facial soft tissue morphology prediction based on conditional deformation field completion, characterized in that, The method comprises the following steps: (1) Obtain bone and soft tissue data, including dense preoperative soft tissue point cloud, sparse preoperative soft tissue landmark points and sparse preoperative bone landmark points, which are combined to construct a data set; (2) Extract the initial displacement vector of each point in the data set, and set a mask for each point to represent the type of the point; (3) Construct a deformation field completion network containing an information propagator and a prediction head, wherein the information propagator is used to propagate the known initial displacement vector on the bone landmark points along the inherent geometric manifold of the facial point cloud to the soft tissue point cloud and the soft tissue landmark points with zero initial displacement; and the prediction head predicts a three-dimensional displacement vector for the soft tissue point cloud and the soft tissue landmark points based on the information propagator result; (4) Calculate the predicted postoperative coordinates of the soft tissue point cloud and the soft tissue landmark points in the data set based on the predicted three-dimensional displacement vector.

2. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, characterized in that, In step (2), for the bone landmark points, the difference between the postoperative coordinates and the preoperative coordinates is obtained. 3.The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, wherein, In step (2), for all soft tissue point clouds and soft tissue landmark points, the initial value is a zero vector.

4. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, characterized in that, In step (2), the mask is 0 for preoperative soft tissue points, 1 for bone landmark points, and 2 for soft tissue landmark points.

5. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, characterized in that, In step (3), the information propagator uses the information propagation mechanism of the dynamic graph convolution network as the deformation field propagator.

6. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, characterized in that, The information propagator uses a stack of more EdgeConv modules to build a deeper information propagation path.

7. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 5, characterized in that, The bone landmark points serve as known condition sources for the deformation field, and the soft tissue landmark points serve as supervision or evaluation targets.

8. The postoperative facial soft tissue morphology prediction method based on condition deformation field completion according to claim 1, characterized in that, The deformation field completion network uses a composite loss function for supervised training, including: (1) Overall shape supervision: using chamfer distance loss to ensure that the predicted point cloud is aligned with the true postoperative point cloud in terms of macro shape; (2) Geometric regularization loss: introducing Laplacian smoothing loss to promote surface smoothness and introducing edge length consistency loss to maintain the consistency of edge length within the local neighborhood.

9. A postoperative facial soft tissue morphology prediction system based on conditional morphable field completion to implement the method of any one of claims 1-8, characterized in that, The system comprises: A point cloud acquisition module for acquiring bone and soft tissue data, including dense preoperative soft tissue point cloud, sparse preoperative soft tissue landmark points and sparse preoperative bone landmark points, which are combined to construct a data set; A displacement extraction module for extracting the initial displacement vector of each point in the data set, and setting a mask for each point to represent the type of the point; A network construction module for constructing a deformation field completion network containing an information propagator and a prediction head, wherein the information propagator is used to propagate the known initial displacement vector on the bone landmark points along the inherent geometric manifold of the facial point cloud to the soft tissue point cloud and the soft tissue landmark points with zero initial displacement; and the prediction head predicts a three-dimensional displacement vector for the soft tissue point cloud and the soft tissue landmark points based on the information propagator result; A coordinate prediction module for calculating the predicted postoperative coordinates of the soft tissue point cloud and the soft tissue landmark points in the data set based on the predicted three-dimensional displacement vector.

Images