Plastic surgery plan generation method, device and storage medium based on medical images

Abstract

The present application relates to the field of data processing technology, and discloses a method, device and storage medium for generating a plastic surgery plan based on medical images. The method comprises: preprocessing and adaptively segmenting multimodal medical image data to obtain a three-dimensional image data set; reconstructing the three-dimensional image data set by multi-scale feature fusion to obtain a three-dimensional anatomical model; interactively defining the target of the three-dimensional anatomical model to obtain a quantitative surgical target parameter set; performing multi-target optimization processing on the quantitative surgical target parameter set to obtain an optimized surgical plan; based on the optimized surgical plan, performing a virtual surgical simulation driven by a physical engine on the three-dimensional anatomical model to obtain a postoperative appearance prediction model; dynamically aligning and performing real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generating a target plastic surgery plan based on the surgical navigation data stream. The present application improves the efficiency and accuracy of generating plastic surgery plans based on medical images.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a method, device and storage medium for generating a plastic surgery plan based on medical images. Background Art

[0002] Currently, surgical planning technology in the field of plastic surgery is undergoing a transformation from traditional experience-based to digital and intelligent. Advanced medical institutions have begun to use technologies such as 3D imaging, computer-aided design (CAD), and virtual reality (VR) to assist in surgical planning. These methods provide doctors with a more intuitive anatomical visualization and surgical simulation platform by building a digital 3D model of the patient. At the same time, some research institutions are exploring the application of artificial intelligence algorithms in surgical planning, such as using deep learning for facial feature recognition and aesthetic assessment, and using machine learning algorithms to optimize surgical plans.

[0003] However, existing technologies still face many challenges and limitations. First, the fusion and utilization efficiency of multimodal medical images is not high, and it is difficult to fully extract and integrate the complementary information provided by different imaging modalities (such as CT, MRI, 3D scanning, etc.). Secondly, the existing surgical goal definition methods are often oversimplified, making it difficult to accurately quantify and express complex aesthetic and functional requirements. Furthermore, most systems lack an effective multi-objective optimization mechanism, making it difficult to find the best balance between multiple dimensions such as aesthetics, safety, and recovery speed. In addition, existing virtual surgical simulation technologies usually use simplified physical models, which make it difficult to accurately simulate the nonlinear deformation behavior of soft tissues, resulting in low accuracy and efficiency in generating plastic surgery plans using medical imaging data. Summary of the invention

[0004] The present application provides a method, device and storage medium for generating a plastic surgery plan based on medical images, which are used to improve the efficiency and accuracy of generating a plastic surgery plan based on medical images.

[0005] In a first aspect, the present application provides a method for generating a plastic surgery plan based on medical images, the method comprising: preprocessing and adaptively segmenting the acquired multimodal medical image data to obtain a three-dimensional image data set; performing multi-scale feature fusion reconstruction on the three-dimensional image data set to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data; performing interactive target definition based on facial features on the three-dimensional anatomical model to obtain a quantitative surgical target parameter set; performing multi-target optimization processing on the quantitative surgical target parameter set to obtain an optimized surgical plan; based on the optimized surgical plan, performing a physics engine-driven virtual surgical simulation on the three-dimensional anatomical model to obtain a postoperative appearance prediction model; performing dynamic registration and real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generating a target plastic surgery plan based on the surgical navigation data stream.

[0006] In combination with the first aspect, in a first implementation of the first aspect of the present application, the preprocessing and adaptive segmentation of the acquired multimodal medical image data to obtain a three-dimensional image data set includes: performing adaptive histogram equalization processing on the multimodal medical image data to obtain enhanced medical image data; performing non-local mean filtering processing on the enhanced medical image data to obtain denoised medical image data; performing spatial registration based on maximizing mutual information on the denoised medical image data to obtain medical image data in a unified coordinate system; inputting the medical image data in the unified coordinate system into a U-Net deep learning network for segmentation to obtain a preliminary segmented image; The preliminary segmented image is processed by morphological closing operation to obtain an optimized segmented image; boundary information of the optimized segmented image is extracted by an edge detection algorithm to obtain anatomical structure boundary information; surface reconstruction is performed according to the anatomical structure boundary information by an iso-rectangular surface extraction algorithm to obtain a preliminary three-dimensional pixel grid; grid simplification and subdivision operations are performed on the preliminary three-dimensional pixel grid to obtain a first candidate three-dimensional pixel grid; texture mapping based on the multimodal medical imaging data is performed on the candidate three-dimensional pixel grid to obtain a second candidate three-dimensional pixel grid with surface features; biomechanical parameter fusion is performed on the second candidate three-dimensional pixel grid to obtain the three-dimensional image data set.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the three-dimensional image dataset is subjected to multi-scale feature fusion reconstruction to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data, including: performing wavelet transform on the three-dimensional image dataset to obtain a multi-scale feature map; performing deep feature extraction on the multi-scale feature map through a convolutional neural network to obtain a deep feature representation; performing multi-level feature fusion on the deep feature representation through a feature pyramid network to obtain a multi-resolution feature map; performing spatial context enhancement on the multi-resolution feature map through a self-attention mechanism to obtain enhanced spatial context information ; Perform a three-dimensional deconvolution operation on the enhanced spatial context information to obtain reconstructed voxel data; perform surface reconstruction on the reconstructed voxel data using the March cube algorithm to obtain a three-dimensional surface mesh; perform Laplace smoothing on the three-dimensional surface mesh to obtain the high-resolution geometric structure; perform surface feature mapping on the high-resolution geometric structure using a texture mapping algorithm to obtain surface data containing multi-level tissue representation information; perform mechanical property calculations on the surface data containing multi-level tissue representation information using a finite element analysis algorithm to obtain a data structure containing biomechanical data; perform data compression and index construction on the data structure containing biomechanical data to obtain the three-dimensional anatomical model.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present application, the interactive target definition based on facial features of the three-dimensional anatomical model to obtain a quantitative surgical target parameter set includes: performing key point detection on the three-dimensional anatomical model to obtain a facial feature point set; performing dimensionality reduction processing on the facial feature point set through a principal component analysis algorithm to obtain a facial morphology principal component; performing statistical analysis on the facial morphology principal component to obtain a facial aesthetic scoring index; performing feature extraction on the three-dimensional anatomical model through a convolutional neural network to obtain facial depth feature representation information; performing cluster analysis on the facial depth feature representation information to obtain a facial region segmentation image; performing topological relationship analysis on the facial region segmentation image through a graph convolutional network to obtain a facial structure graph representation; performing graph matching on the facial structure graph representation based on the facial aesthetic scoring index to obtain an expected aesthetic standard deviation value; performing parameterized encoding on the expected aesthetic standard deviation value through a variational autoencoder to obtain an aesthetic parameter space; interactively adjusting the aesthetic parameter space to obtain a personalized aesthetic target; performing shape encoding on the personalized aesthetic target through a preset Fourier descriptor to obtain the quantitative surgical target parameter set.

[0009] In combination with the first aspect, in a fourth implementation method of the first aspect of the present application, the multi-objective optimization processing of the quantitative surgical target parameter set to obtain an optimized surgical plan includes: decomposing the quantitative surgical target parameter set to obtain a sub-target set; assigning weights to the sub-target set through a hierarchical analysis method to obtain a weighted sub-target set; defining constraints for the weighted sub-target set to obtain a constraint data set; generating an initial solution for the constraint data set through a genetic algorithm to obtain a candidate solution set; performing fitness evaluation on the candidate solution set to obtain quality score data corresponding to the candidate solution set; performing Pareto optimization analysis on the quality score data through a non-dominated sorting genetic algorithm to obtain a non-inferior solution set; performing cluster analysis on the non-inferior solution set to obtain a representative solution set; performing multi-criteria decision-making on the representative solution set through a fuzzy comprehensive evaluation method to obtain an optimal solution; decomposing the optimal solution into surgical steps to obtain a preliminary surgical plan; and refining the preliminary surgical plan through a Monte Carlo tree search algorithm to obtain the optimized surgical plan.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, based on the optimized surgical plan, a virtual surgical simulation driven by a physical engine is performed on the three-dimensional anatomical model to obtain a postoperative appearance prediction model, including: mesh simplification of the three-dimensional anatomical model to obtain simplified three-dimensional mesh data; discretization of the simplified three-dimensional mesh data to obtain calculation mesh data; material property assignment of the calculation mesh data to obtain calculation data; dynamic parameterization of the calculation data through a particle-spring system algorithm to obtain a deformable body parameter set; boundary condition setting of the deformable body parameter set to obtain deformable body data under constraint conditions; time discretization of the deformable body data under constraint conditions to obtain dynamic simulation data; time series data decomposition of the dynamic simulation data to obtain deformation process time series data; contour alignment of the deformation process time series data through a shape matching algorithm to obtain a standardized deformation data set; statistical analysis of the standardized deformation data set to obtain deformation feature distribution data; appearance synthesis of the deformation feature distribution data through a generative adversarial network to obtain postoperative appearance prediction data.

[0011] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, the postoperative appearance prediction model is dynamically registered and analyzed in real time for feedback to obtain a surgical navigation data stream, and a target plastic surgery plan is generated based on the surgical navigation data stream, including: extracting feature points from the postoperative appearance prediction model to obtain a set of key anatomical marker points; performing real-time position matching on the key anatomical marker point set through an iterative closest point algorithm to obtain a dynamic registration matrix; performing error analysis on the dynamic registration matrix to obtain registration accuracy evaluation data; performing noise suppression on the registration accuracy evaluation data through a Kalman filter algorithm to obtain smoothed registration parameters; performing time series analysis on the smoothed registration parameters to obtain registration trend data; classifying the registration trend data through a support vector machine to obtain a surgical process status identifier; performing decision tree analysis on the surgical process status identifier to obtain a real-time feedback instruction set; optimizing the real-time feedback instruction set through a reinforcement learning algorithm to obtain an adaptive navigation strategy; performing data stream processing on the adaptive navigation strategy to obtain a surgical navigation data stream; and performing comprehensive analysis on the surgical navigation data stream through a deep neural network to obtain a target plastic surgery plan.

[0012] In a second aspect, the present application provides a device for generating a plastic surgery plan based on medical images, the device for generating a plastic surgery plan based on medical images comprising:

[0013] A segmentation module is used to preprocess and adaptively segment the acquired multimodal medical image data to obtain a three-dimensional image data set;

[0014] A reconstruction module, used for performing multi-scale feature fusion reconstruction on the three-dimensional image data set to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data;

[0015] A definition module, used for interactively defining the target based on facial features on the three-dimensional anatomical model to obtain a quantitative surgical target parameter set;

[0016] A processing module, used for performing multi-objective optimization processing on the quantitative surgical target parameter set to obtain an optimized surgical plan;

[0017] A simulation module, configured to perform a virtual surgery simulation driven by a physical engine on the three-dimensional anatomical model based on the optimized surgery plan to obtain a postoperative appearance prediction model;

[0018] The analysis module is used to perform dynamic registration and real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generate a target plastic surgery plan based on the surgical navigation data stream.

[0019] A third aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned method for generating a plastic surgery plan based on medical images.

[0020] In the technical solution provided by the present application, by preprocessing and adaptive segmentation of multimodal medical image data, combined with adaptive histogram equalization, non-local mean filtering and spatial registration based on mutual information maximization and other technologies, the quality and information extraction efficiency of the original medical image are significantly improved, laying a solid foundation for subsequent processing. Secondly, multi-scale feature fusion reconstruction technology, including wavelet transform, deep feature extraction of convolutional neural network, feature pyramid network and self-attention mechanism, is adopted to achieve high-precision reconstruction of three-dimensional anatomical model, which not only retains high-resolution geometric structure, but also contains multi-level tissue representation information and biomechanical data, providing comprehensive and accurate anatomical information for surgical planning. In the target definition stage, an interactive target definition method based on facial features is introduced, and the precise quantification of surgical targets is achieved through principal component analysis, convolutional neural network feature extraction, graph convolution network topology analysis and other technologies, which greatly improves the personalization and accuracy of surgical planning. In the optimization stage, multi-objective optimization processing technology, including hierarchical analysis method, genetic algorithm, non-dominated sorting genetic algorithm and fuzzy comprehensive evaluation method, is adopted to effectively balance multiple goals such as aesthetic effect, safety and recovery speed, and generate the optimal surgical plan. In the virtual surgery simulation, the simulation technology driven by the physical engine, combined with the mass-spring system algorithm and shape matching algorithm, achieved highly realistic soft tissue deformation simulation, greatly improving the accuracy of postoperative effect prediction. Finally, in terms of intraoperative navigation and real-time feedback analysis, through the comprehensive application of iterative nearest point algorithm, Kalman filtering, support vector machine classification and reinforcement learning, high-precision dynamic registration and real-time feedback were achieved, providing accurate navigation support for the surgical process. This series of technological innovations and integrations not only significantly improved the accuracy and reliability of plastic surgery planning, but also realized intelligent support for the entire process from preoperative planning to intraoperative navigation. By comprehensively utilizing multimodal medical imaging, deep learning, computer vision and artificial intelligence technologies, this method effectively overcomes the problems of insufficient information utilization, inaccurate target definition, incomplete optimization, and unrealistic simulation in traditional surgical planning methods, providing doctors with more comprehensive and accurate decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0022] Figure 1 A schematic diagram of an embodiment of a method for generating a plastic surgery plan based on medical images in an embodiment of the present application;

[0023] Figure 2 Schematic diagram of an embodiment of a plastic surgery plan generation device based on medical images in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The present application embodiment provides a method, device and storage medium for generating a plastic surgery plan based on medical images. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of a method for generating a plastic surgery plan based on medical images includes:

[0026] Step S101, preprocessing and adaptively segmenting the acquired multimodal medical image data to obtain a three-dimensional image data set;

[0027] Step S102, performing multi-scale feature fusion reconstruction on the three-dimensional image data set to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes a high-resolution geometric structure, multi-level tissue representation information, and biomechanical data;

[0028] Step S103, performing interactive target definition based on facial features on the three-dimensional anatomical model to obtain a quantitative surgical target parameter set;

[0029] Step S104, performing multi-objective optimization processing on the quantified surgical target parameter set to obtain an optimized surgical plan;

[0030] Step S105: Based on the optimized surgical plan, a virtual surgical simulation driven by a physical engine is performed on the three-dimensional anatomical model to obtain a postoperative appearance prediction model;

[0031] Step S106: dynamically align and perform real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generate a target plastic surgery plan based on the surgical navigation data stream.

[0032] It is understandable that the execution subject of the present application may be a plastic surgery plan generation device based on medical images, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0033] Specifically, by acquiring multimodal medical imaging data such as CT, MRI, and 3D surface scans, taking facial plastic surgery as an example, high-resolution CT scans of patients are obtained to understand the bone structure, MRI images to observe the distribution of soft tissues, and 3D surface scans to capture skin surface details. These multimodal data are then preprocessed and adaptively segmented, where the preprocessing includes image enhancement and noise removal, while the adaptive segmentation uses deep learning algorithms such as U-Net to automatically identify and segment different anatomical structures. For example, in facial plastic surgery, structures at different levels such as facial bones, muscles, fat tissue, and skin can be automatically segmented. The output is an accurate three-dimensional image dataset that contains detailed anatomical information of the patient's face.

[0034] Next, this 3D image dataset is reconstructed by multi-scale feature fusion, a process that utilizes technologies such as wavelet transform, convolutional neural network, and feature pyramid network. Through these technologies, a full range of information from large-scale structures (such as overall facial contours) to tiny details (such as wrinkles and pores) can be captured. For example, in rhinoplasty planning, the curve of the bridge of the nose, the shape of the tip of the nose, and the subtle structure of the nose wing can be accurately reconstructed. The result of the reconstruction is a highly accurate 3D anatomical model that contains not only high-resolution geometric structures, but also multi-level tissue representation information (such as layered information of skin, fat, muscle, and bone) and biomechanical data (such as tissue elasticity and hardness).

[0035] With this detailed 3D anatomical model, interactive goal definition based on facial features can be performed with the patient. Techniques such as principal component analysis, deep feature extraction, and graph convolutional networks are used to allow the patient and the patient to jointly define and adjust surgical goals. For example, in mandibular plastic surgery, the angle of the mandibular angle, the protrusion of the chin, and other parameters can be precisely adjusted through an interactive interface, and the impact of these adjustments on the overall facial aesthetics will be displayed in real time. The output of this process is a quantified set of surgical target parameters that accurately defines the desired facial features.

[0036] Subsequently, this set of quantitative surgical target parameters was subjected to multi-objective optimization. Techniques such as the analytic hierarchy process, genetic algorithm, and non-dominated sorting genetic algorithm were used to balance multiple objectives such as aesthetic effect, surgical safety, and recovery time. For example, in the planning of a full facial rejuvenation surgery, multiple objectives such as improving skin firmness, optimizing facial contours, and minimizing surgical trauma are weighed to generate an optimal surgical plan.

[0037] Based on this optimized surgical plan, a physics-driven virtual surgery simulation is then performed on the 3D anatomical model. Techniques such as the mass-spring algorithm and finite element analysis are used to simulate tissue deformation and mechanical response during surgery. For example, in the simulation of facial filler surgery, the deformation of the surrounding soft tissue after the filler is injected, as well as the sedimentation effect over time, can be accurately predicted. The result of this simulation is a postoperative appearance prediction model that provides a highly realistic preview of the surgical effect for the patient. Finally, the postoperative appearance prediction model is dynamically registered and analyzed for real-time feedback.

[0038] In the embodiment of the present application, by preprocessing and adaptive segmentation of multimodal medical image data, combined with adaptive histogram equalization, non-local mean filtering and spatial registration based on mutual information maximization and other technologies, the quality and information extraction efficiency of the original medical image are significantly improved, laying a solid foundation for subsequent processing. Secondly, multi-scale feature fusion reconstruction technology, including wavelet transform, deep feature extraction of convolutional neural network, feature pyramid network and self-attention mechanism, is adopted to achieve high-precision three-dimensional anatomical model reconstruction, which not only retains high-resolution geometric structure, but also contains multi-level tissue representation information and biomechanical data, providing comprehensive and accurate anatomical information for surgical planning. In the target definition stage, an interactive target definition method based on facial features is introduced, and the precise quantification of surgical targets is achieved through principal component analysis, convolutional neural network feature extraction, graph convolution network topology analysis and other technologies, which greatly improves the personalization and accuracy of surgical planning. In the optimization stage, multi-objective optimization processing technology, including hierarchical analysis method, genetic algorithm, non-dominated sorting genetic algorithm and fuzzy comprehensive evaluation method, is adopted to effectively balance multiple goals such as aesthetic effect, safety and recovery speed, and generate the optimal surgical plan. In the virtual surgery simulation, the simulation technology driven by the physical engine, combined with the mass-spring system algorithm and shape matching algorithm, achieved highly realistic soft tissue deformation simulation, greatly improving the accuracy of postoperative effect prediction. Finally, in terms of intraoperative navigation and real-time feedback analysis, through the comprehensive application of iterative nearest point algorithm, Kalman filtering, support vector machine classification and reinforcement learning, high-precision dynamic registration and real-time feedback were achieved, providing accurate navigation support for the surgical process. This series of technological innovations and integrations not only significantly improved the accuracy and reliability of plastic surgery planning, but also realized intelligent support for the entire process from preoperative planning to intraoperative navigation. By comprehensively utilizing multimodal medical imaging, deep learning, computer vision and artificial intelligence technologies, this method effectively overcomes the problems of insufficient information utilization, inaccurate target definition, incomplete optimization, and unrealistic simulation in traditional surgical planning methods, providing doctors with more comprehensive and accurate decision support.

[0039] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0040] (1) Performing adaptive histogram equalization processing on multimodal medical image data to obtain enhanced medical image data;

[0041] (2) performing non-local mean filtering on the enhanced medical image data to obtain denoised medical image data;

[0042] (3) Perform spatial registration of the denoised medical image data based on maximizing mutual information to obtain medical image data in a unified coordinate system;

[0043] (4) Input the medical image data in the unified coordinate system into the U-Net deep learning network for segmentation to obtain a preliminary segmented image;

[0044] (5) Perform morphological closing operation on the preliminary segmented image to obtain an optimized segmented image;

[0045] (6) Extracting boundary information from the optimized segmented image using an edge detection algorithm to obtain boundary information of the anatomical structure;

[0046] (7) Using the equal straight surface extraction algorithm to perform surface reconstruction based on the anatomical structure boundary information, a preliminary three-dimensional pixel grid is obtained;

[0047] (8) performing mesh simplification and subdivision operations on the preliminary three-dimensional pixel grid to obtain a first candidate three-dimensional pixel grid;

[0048] (9) performing texture mapping on the candidate three-dimensional pixel grid based on the multimodal medical imaging data to obtain a second candidate three-dimensional pixel grid having surface features;

[0049] (10) Perform biomechanical parameter fusion on the second candidate three-dimensional pixel grid to obtain a three-dimensional image data set.

[0050] Specifically, adaptive histogram equalization is performed on multimodal medical image data. This step aims to enhance image contrast and make image details clearer. For example, in facial plastic surgery planning, applying adaptive histogram equalization to CT scan images can highlight subtle differences in facial bone structure, while processing MRI images can enhance the contrast of soft tissue. The enhanced medical image data obtained by this processing provides clearer visual information for subsequent analysis. The enhanced medical image data is processed by non-local mean filtering to reduce image noise. The non-local mean filtering algorithm removes noise by finding similar regions in the image and performing weighted averaging while retaining image details and edge information. In facial plastic surgery planning, this step can effectively remove random noise in CT or MRI images while retaining details of key anatomical features such as nose bridge contour and orbital structure. The denoised medical image data then undergoes spatial registration based on maximizing mutual information. This step aims to align images of different modalities (such as CT, MRI, 3D scans) to the same coordinate system. For example, in maxillofacial plastic surgery planning, this step can accurately align CT images showing bone structures with MRI images showing soft tissue distribution, allowing doctors to fully evaluate the patient's anatomical structure in the same spatial reference system. The registered medical image data is input into the U-Net deep learning network for segmentation. U-Net is a convolutional neural network architecture specifically used for medical image segmentation, which can identify and segment different anatomical structures with high accuracy. In facial plastic surgery planning, U-Net can automatically segment different tissue layers such as skin, fat, muscle, and bone on the face, providing detailed anatomical information for subsequent surgical planning. The preliminary segmented image output by U-Net is processed by morphological closing operations to fill small holes and smooth the boundaries. This step can repair small defects that may occur during the segmentation process, such as filling pores or breaks caused by fine wrinkles in facial skin segmentation.

[0051] Subsequently, the boundary information of the optimized segmented image is extracted by edge detection algorithms (such as Canny edge detector) to obtain clear anatomical boundary information. This step can accurately depict the boundaries of key features such as facial contours, nose bridge lines, and lip shapes. The surface is reconstructed based on the anatomical boundary information using an isosurface extraction algorithm (such as Marching Cubes algorithm) to generate a preliminary three-dimensional pixel grid. This step converts the anatomical structure information in the two-dimensional image into a three-dimensional surface model. For example, a three-dimensional nose shape, mandibular contour, etc. can be reconstructed. The preliminary three-dimensional pixel grid is meshed and subdivided to optimize the mesh quality and computational efficiency. Mesh simplification can reduce unnecessary details, while subdivision can increase accuracy in key areas. For example, in rhinoplasty surgery planning, complex areas such as the nose tip can be meshed, while the relatively flat nose bridge area can be appropriately simplified. Next, the optimized three-dimensional pixel grid is texture mapped based on multimodal medical imaging data. This step maps the surface feature information in the original medical image to the three-dimensional model, so that the model not only has accurate geometric shapes but also contains real surface details. For example, information such as skin texture and pigmentation can be accurately mapped onto a 3D facial model.

[0052] Finally, the biomechanical parameters are fused to the three-dimensional pixel grid with surface features. This step integrates the biomechanical properties of the tissue, such as elasticity and hardness, into the model. For example, in the facial filler surgery plan, the corresponding elastic parameters can be assigned to the soft tissues in different areas to accurately predict the tissue deformation after the filler is injected in the subsequent virtual surgery simulation.

[0053] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0054] (1) Perform wavelet transform on the three-dimensional image data set to obtain a multi-scale feature map;

[0055] (2) Deep feature extraction is performed on multi-scale feature maps through convolutional neural networks to obtain deep feature representation;

[0056] (3) Multi-level feature fusion is performed on the deep feature representation through the feature pyramid network to obtain a multi-resolution feature map;

[0057] (4) The spatial context of the multi-resolution feature map is enhanced through the self-attention mechanism to obtain enhanced spatial context information;

[0058] (5) performing a three-dimensional deconvolution operation on the enhanced spatial context information to obtain reconstructed voxel data;

[0059] (6) reconstructing the surface of the reconstructed voxel data using the March cube algorithm to obtain a three-dimensional surface mesh;

[0060] (7) Perform Laplace smoothing on the three-dimensional surface mesh to obtain a high-resolution geometric structure;

[0061] (8) Mapping the surface features of high-resolution geometric structures through texture mapping algorithms to obtain surface data containing multi-level tissue representation information;

[0062] (9) calculating mechanical properties of surface data containing multi-level tissue representation information by a finite element analysis algorithm to obtain a data structure containing biomechanical data;

[0063] (10) Data compression and index construction are performed on the data structure containing the biomechanical data to obtain a three-dimensional anatomical model.

[0064] Specifically, wavelet transform of 3D image datasets is a critical first step. Wavelet transform can decompose images into features of different scales, thereby generating multi-scale feature maps. For example, in facial plastic surgery planning, wavelet transform can simultaneously capture large-scale contours of the face (such as the jawline) and tiny details (such as wrinkle texture), providing a comprehensive information basis for subsequent analysis.

[0065] Deep feature extraction is performed on these multi-scale feature maps through convolutional neural networks. The multi-layer structure of convolutional neural networks can learn and extract increasingly abstract features layer by layer. In rhinoplasty surgery planning, shallow networks may recognize simple edges and textures, while deep networks can capture complex structural features, such as the curvature of the bridge of the nose or the shape of the nose wing. The output of this step is a set of deep feature representations, which contains rich anatomical information. Multi-level feature fusion is performed on these deep feature representations through feature pyramid networks. Feature pyramid networks can integrate features of different scales to generate multi-resolution feature maps. In the planning of comprehensive facial rejuvenation surgery, this step can effectively integrate the overall features of large facial areas (such as forehead and cheeks) and local details (such as eye wrinkles), laying the foundation for subsequent precise planning.

[0066] In order to further enhance the expressive power of features, the self-attention mechanism is applied to the multi-resolution feature map for spatial context enhancement. The self-attention mechanism can capture the long-range dependencies between features, thereby obtaining enhanced spatial context information. For example, in the planning of mandibular plastic surgery, the self-attention mechanism can help understand the relationship between the mandibular angle and the overall facial contour, ensuring that local modifications are coordinated with the overall aesthetics. The enhanced spatial context information is subjected to a three-dimensional deconvolution operation. This step aims to remap the abstract feature information into three-dimensional space to obtain reconstructed voxel data. In the planning of facial filling surgery, this step can accurately reconstruct the volume information of each area of ​​the face, providing a basis for the subsequent filling volume planning. The reconstructed voxel data is surface reconstructed by the March cube algorithm to generate a three-dimensional surface mesh. The March cube algorithm can effectively extract isosurfaces from voxel data to form a continuous surface mesh. For example, in the planning of zygomatic plastic surgery, this step can accurately reconstruct the three-dimensional surface morphology of the zygomatic bone.

[0067] The generated 3D surface mesh is Laplace smoothed to obtain high-resolution geometry. Laplace smoothing can remove noise and small fluctuations on the mesh surface while retaining important geometric features. In lip plastic surgery planning, this step can smooth the lip contour while retaining key features such as the lip peak. The surface features of the high-resolution geometry are mapped by the texture mapping algorithm to obtain surface data containing multi-level tissue representation information. This step accurately maps the texture information in the original medical image to the 3D model. For example, in facial skin rejuvenation surgery planning, information such as skin texture and pigment distribution can be mapped to the facial model to provide an intuitive reference for subsequent treatment plan design. The mechanical properties of the surface data containing multi-level tissue representation information are calculated by the finite element analysis algorithm to obtain a data structure containing biomechanical data. Finite element analysis can simulate the deformation behavior of different tissues under external forces. In breast plastic surgery planning, this step can predict the impact of implants on surrounding soft tissues and help choose the most appropriate implant size and location. The data structure containing biomechanical data is compressed and indexed to obtain the final 3D anatomical model. Data compression can reduce storage space and improve data transmission efficiency, while index construction helps to quickly retrieve and operate specific parts of the model. For example, in complex maxillofacial reconstruction surgery planning, this step can ensure that any part of the model can be quickly accessed and modified during the interactive planning process, improving planning efficiency.

[0068] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0069] (1) Detect key points on the 3D anatomical model to obtain a set of facial feature points;

[0070] (2) Using principal component analysis algorithm to reduce the dimension of facial feature point set and obtain the principal components of facial morphology;

[0071] (3) Statistical analysis of the principal components of facial morphology was performed to obtain facial aesthetic scoring indicators;

[0072] (4) Extract features from the three-dimensional anatomical model through a convolutional neural network to obtain facial deep feature representation information;

[0073] (5) Perform cluster analysis on facial depth feature representation information to obtain a facial region segmentation image;

[0074] (6) Perform topological relationship analysis on the facial region segmentation image through a graph convolutional network to obtain a facial structure graph representation;

[0075] (7) Based on the facial aesthetic scoring index, the facial structure map representation is matched to obtain the expected aesthetic standard deviation value;

[0076] (8) parameterizing the expected aesthetic standard deviation value through a variational autoencoder to obtain an aesthetic parameter space;

[0077] (9) Interactively adjust the aesthetic parameter space to obtain personalized aesthetic goals;

[0078] (10) The shape of the personalized aesthetic target is encoded through a preset Fourier descriptor to obtain a quantitative surgical target parameter set.

[0079] Specifically, key point detection is performed on the three-dimensional anatomical model to obtain a set of facial feature points. These feature points usually include anatomical landmarks such as the corners of the eyes, the tip of the nose, and the corners of the lips. For example, in a rhinoplasty surgery plan, accurately locating key points such as the root of the nose, the tip of the nose, and the wing of the nose is crucial for evaluating the morphology of the nose. Subsequently, the facial feature point set is subjected to dimensionality reduction processing using a principal component analysis algorithm to obtain the principal components of facial morphology. Principal component analysis can extract the main features that best represent facial changes. In a facial contouring plan, these principal components may include key factors such as the mandibular angle and cheek fullness. The principal components of facial morphology are statistically analyzed to obtain facial aesthetic scoring indicators. This step combines the principal components with aesthetic standards to establish an objective scoring system. For example, in a comprehensive facial rejuvenation surgery plan, these indicators may include quantitative indicators such as facial symmetry and golden ratio compliance.

[0080] The feature extraction of the three-dimensional anatomical model is performed through a convolutional neural network to obtain the facial deep feature representation information. The convolutional neural network can automatically learn and extract complex facial features. In the eye plastic surgery plan, these deep features may include subtle features such as eyelid shape and eye angle inclination. The facial deep feature representation information is clustered to obtain a facial region segmentation image. Cluster analysis can classify similar features to divide different facial regions. In the zygomatic plastic surgery plan, this step can accurately divide the boundary between the zygomatic region and the surrounding soft tissue. The topological relationship analysis of the facial region segmentation image is performed through a graph convolutional network to obtain a facial structure graph representation. The graph convolutional network can capture the spatial relationship between different facial regions. In the mandibular angle plastic surgery plan, this step can analyze the mutual influence of the mandibular angle and adjacent facial structures.

[0081] Based on the facial aesthetic scoring index obtained previously, the facial structure graph representation is matched to obtain the expected aesthetic standard deviation value. This step compares the patient's facial structure with the aesthetic standard and quantifies the aspects that need to be improved. For example, in the lip plastic surgery plan, the degree of deviation between the current lip shape and the ideal lip shape can be calculated. The expected aesthetic standard deviation value is parameterized and encoded by the variational autoencoder to obtain the aesthetic parameter space. The variational autoencoder can compress complex aesthetic features into a continuous parameter space. In the facial filling surgery plan, these parameters may represent the filling amount and shape changes in different areas. The aesthetic parameter space is interactively adjusted to obtain personalized aesthetic goals. This step allows doctors and patients to explore and adjust in the parameter space to reach a consensus on the surgical goal. For example, in the rhinoplasty surgery plan, the rhinoplasty effects of different degrees can be previewed by adjusting the parameters. Finally, the personalized aesthetic goal is shape-encoded by the preset Fourier descriptor to obtain a quantitative surgical target parameter set. The Fourier descriptor can accurately describe complex shape changes with limited parameters. In chin surgery planning, these parameters may describe changes in the curvature and projection of the chin contour.

[0082] Through this series of steps, the plastic surgery plan starts from the objective three-dimensional anatomical model, and after multi-level feature extraction, analysis and optimization, it finally obtains a highly personalized and quantified surgical target parameter set. This parameter set not only takes into account the individual characteristics of the patient, but also incorporates aesthetic standards and medical expertise, providing a precise mathematical basis for subsequent surgical simulation and planning. For example, in a complex facial reshaping surgery plan, this parameter set may include multi-dimensional information such as shape change parameters of multiple facial areas, soft tissue filling amount, bone adjustment angle, etc., making the entire surgical plan more accurate, predictable and personalized.

[0083] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0084] (1) Decomposing the quantitative surgical target parameter set to obtain a sub-target set;

[0085] (2) Using the analytic hierarchy process to assign weights to the sub-goal set, a weighted sub-goal set is obtained;

[0086] (3) Define constraints on the weighted sub-goal set to obtain a constraint data set;

[0087] (4) Generate an initial solution for the constraint condition data set through a genetic algorithm to obtain a candidate solution set;

[0088] (5) Perform fitness evaluation on the candidate solution set to obtain quality score data corresponding to the candidate solution set;

[0089] (6) Perform Pareto optimization analysis on the quality score data using a non-dominated sorting genetic algorithm to obtain a non-inferior solution set;

[0090] (7) Perform cluster analysis on the non-inferior solution set to obtain the representative solution set;

[0091] (8) Perform multi-criteria decision-making on the representative solution set through fuzzy comprehensive evaluation method to obtain the optimal solution;

[0092] (9) Decomposing the optimal solution into surgical steps to obtain a preliminary surgical plan;

[0093] (10) The preliminary surgical plan is refined through the Monte Carlo tree search algorithm to obtain the optimized surgical plan.

[0094] Specifically, the quantitative surgical target parameter set is decomposed to obtain a sub-target set. For example, in a comprehensive facial plastic surgery plan, the overall goal can be decomposed into specific sub-targets such as nose morphology improvement, zygomatic bone contour adjustment, and mandibular angle modification. This decomposition makes complex plastic surgery goals easier to analyze and optimize. The sub-target set is weighted by the analytic hierarchy process to obtain a weighted sub-target set. The analytic hierarchy process takes into account the relative importance of each sub-target. In the above facial plastic surgery example, a higher weight may be given to nose morphology improvement, while a medium weight is given to zygomatic bone contour adjustment based on patient needs and medical considerations. Constraints are defined for the weighted sub-target set to obtain a constraint data set. These constraints include medical safety restrictions, anatomical feasibility, etc. For example, in rhinoplasty, constraints may include the maximum allowable increase in nose bridge height to ensure that respiratory function is not affected.

[0095] Genetic algorithms are used to generate initial solutions for the constraint condition data set to obtain a candidate solution set. Genetic algorithms simulate the biological evolution process and generate multiple possible surgical plans that meet the constraints. In facial plastic surgery planning, each candidate solution represents a specific combination of surgical parameters, such as the degree of nose tip protrusion, the amount of cheekbone reduction, etc. The candidate solution set is evaluated for fitness and the quality score data corresponding to the candidate solution set is obtained. The evaluation criteria may include multiple dimensions such as aesthetic effect, surgical risk, and recovery time. For example, a candidate solution may score high in aesthetic effect but low in surgical risk assessment. The quality score data is subjected to Pareto optimization analysis by a non-dominated sorting genetic algorithm to obtain a non-inferior solution set. Pareto optimization considers the trade-offs between multiple objectives and finds a solution set in which improvement in any one objective will lead to the deterioration of other objectives. In facial plastic surgery planning, this may produce a series of solutions, each of which strikes a different balance between aesthetic effect and surgical risk.

[0096] Cluster analysis is performed on the non-inferior solution set to obtain a representative solution set. Cluster analysis groups similar solutions into a group, and selects representative solutions from each group to reduce the number of solutions that need to be further evaluated. In the facial plastic surgery example, several typical solutions may be obtained, such as conservative solutions, radical solutions, and compromise solutions. The fuzzy comprehensive evaluation method is used to make multi-criteria decisions on the representative solution set to obtain the optimal solution. The fuzzy comprehensive evaluation method can handle the uncertainty and subjectivity in the decision-making process. In plastic surgery planning, this step comprehensively considers factors such as doctor experience and patient preferences to select the best solution from the representative solution set. The optimal solution is decomposed into surgical steps to obtain a preliminary surgical plan. This step converts abstract parameters into a specific sequence of surgical steps. For example, in facial plastic surgery, it may include detailed steps such as "first perform nasal bone cutting and raise the height of the nose bridge by 2mm". Finally, the preliminary surgical plan is refined by the Monte Carlo tree search algorithm to obtain the optimized surgical plan. Monte Carlo tree search optimizes the specific parameters and execution order of each step by simulating a large number of possible surgical processes. In facial plastic surgery planning, this may involve adjusting the precise location and depth of each incision, or optimizing the order and dosage of filler injections. Medical imaging-based plastic surgery plan generation methods transform complex surgical goals into an optimized, executable surgical plan. This process not only considers the balance between multiple goals, but also incorporates medical constraints and professional judgment, ultimately producing a personalized surgical plan that meets both aesthetic needs and ensures medical safety. For example, in a comprehensive facial rejuvenation surgery plan, the final optimized surgical plan may contain facial filler injection locations and dosages accurate to the millimeter, precise skin stretching parameters, and detailed surgical step timing, all of which have been optimized multiple times to achieve the best balance between aesthetic effects, safety, and recovery speed.

[0097] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0098] (1) Simplifying the mesh of the three-dimensional anatomical model to obtain simplified three-dimensional mesh data;

[0099] (2) discretizing the simplified three-dimensional grid data to obtain computational grid data;

[0100] (3) Assigning material properties to the computational grid data to obtain computational data;

[0101] (4) Dynamically parameterize the calculated data through the mass-spring system algorithm to obtain the deformable body parameter set;

[0102] (5) setting boundary conditions for the deformable body parameter set to obtain deformable body data under constraint conditions;

[0103] (6) Time discretization of the deformable body data under constraint conditions to obtain dynamic simulation data;

[0104] (7) Decomposing the dynamic simulation data into time series data to obtain the time series data of the deformation process;

[0105] (8) Using a shape matching algorithm to align the contours of the deformation process time series data to obtain a standardized deformation data set;

[0106] (9) Statistically analyze the standardized deformation data set to obtain deformation feature distribution data;

[0107] (10) The appearance of the deformation feature distribution data is synthesized through a generative adversarial network to obtain the postoperative appearance prediction data.

[0108] Specifically, the mesh of the 3D anatomical model is simplified to obtain simplified 3D mesh data. This step reduces the computational complexity while maintaining the main features of the model by reducing the number of vertices and faces in the mesh. For example, in facial plastic surgery planning, the fine mesh of key facial feature points (such as the tip of the nose and the corner of the lip) can be retained, while relatively flat areas can be moderately simplified.

[0109] The simplified 3D mesh data is discretized to obtain computational mesh data. The discretization process converts the continuous geometric model into discrete computational units in preparation for subsequent numerical analysis. In facial filler surgery planning, this step may divide the facial soft tissue into thousands of tetrahedral units, each unit representing a small piece of tissue. Material properties are assigned to the computational mesh data to obtain computational data. This step assigns corresponding physical properties, such as elastic modulus, density, etc., to each unit in the mesh. In rhinoplasty surgery planning, different material properties may be assigned to the tip cartilage, bridge of the nose bone, and surrounding soft tissue to accurately simulate their behavior during surgery.

[0110] The calculated data is dynamically parameterized by the mass-spring system algorithm to obtain a deformable body parameter set. The mass-spring system models the tissue as a network of mass points and springs, which can simulate the elastic deformation of the tissue. In the facelift surgery planning, this step can set the elastic parameters of the facial skin and subcutaneous tissue to predict the morphological changes of the skin after stretching. The deformable body parameter set is subjected to boundary conditions to obtain the deformable body data under constraint conditions. The boundary conditions define the fixed points and external force application points of the model during the simulation process. For example, in the mandibular angle plastic surgery planning, the connection between the jaw and the skull may be set as a fixed point, and the resection force is applied to the mandibular angle area. The deformable body data under constraint conditions is time discretized to obtain dynamic simulation data. Time discretization decomposes the continuous deformation process into a series of discrete time steps, which facilitates the step-by-step calculation of tissue deformation. In the facial filler surgery planning, this may involve simulating the process of soft tissue changes over time after filler injection.

[0111] The dynamic simulation data is decomposed into time series data to obtain the deformation process time series data. This step decomposes the entire deformation process into a series of states at key time points. In the rhinoplasty surgery plan, it may include data at multiple time points such as the initial state, the immediate state after implant placement, the peak of postoperative swelling, and the final stable state. The time series data of the deformation process are aligned by shape matching algorithm to obtain a standardized deformation data set. Shape matching ensures that the data at different time points correspond to each other in space, which is convenient for subsequent analysis. In the facial contour plastic surgery plan, this step can ensure the correspondence of the facial contour throughout the deformation process, so as to accurately track the changes in each part. The standardized deformation data set is statistically analyzed to obtain the deformation feature distribution data. Statistical analysis can reveal the laws and characteristics of the deformation process. For example, in the full facial rejuvenation surgery plan, the distribution of soft tissue thickness changes in different regions can be analyzed to understand the uniformity and durability of the filling effect. The appearance of the deformation feature distribution data is synthesized by the generative adversarial network to obtain the postoperative appearance prediction data. The generative adversarial network can generate realistic visual effects based on statistical features. In facial plastic surgery planning, this step can generate images of the patient's facial appearance at different times after surgery, including details such as skin texture and color changes. The plastic surgery plan generation method based on medical imaging realizes the transition from static three-dimensional models to dynamic deformation prediction, providing comprehensive simulation and prediction support for surgical planning. For example, in a complex facial reshaping surgery plan, this process can accurately predict the combined effects of multiple surgical steps (such as zygomatic bone reduction, mandibular angle reduction, facial filling, etc.), including not only the final appearance changes, but also the recovery process at different times after surgery. This detailed simulation and prediction capability greatly improves the accuracy and reliability of surgical planning, enabling doctors to better assess surgical risks, optimize surgical strategies, and provide patients with more accurate expectations of postoperative effects.

[0112] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0113] (1) Extract feature points from the postoperative appearance prediction model to obtain a set of key anatomical landmark points;

[0114] (2) Using the iterative closest point algorithm, the key anatomical landmark points are matched in real time to obtain a dynamic registration matrix;

[0115] (3) Perform error analysis on the dynamic registration matrix to obtain registration accuracy evaluation data;

[0116] (4) Suppressing the noise of the registration accuracy evaluation data by using the Kalman filter algorithm to obtain smoothed registration parameters;

[0117] (5) Perform time series analysis on the smoothed registration parameters to obtain registration trend data;

[0118] (6) Classifying the registration trend data through support vector machine to obtain the surgical process status identification;

[0119] (7) Perform decision tree analysis on the surgical process status identifier to obtain a real-time feedback instruction set;

[0120] (8) Optimizing the real-time feedback instruction set through reinforcement learning algorithm to obtain an adaptive navigation strategy;

[0121] (9) performing data stream processing on the adaptive navigation strategy to obtain a surgical navigation data stream;

[0122] (10) The surgical navigation data stream is comprehensively analyzed through a deep neural network to obtain the target plastic surgery plan.

[0123] Specifically, feature points are extracted from the postoperative appearance prediction model to obtain a set of key anatomical marker points. These marker points may include key facial positions such as the nose tip, lip peak, and eye corners. For example, in a rhinoplasty surgery plan, 20 feature points such as the root of the nose, nose tip, and nose wing may be extracted to track changes in the nose morphology. The key anatomical marker point set is matched in real time by the iterative closest point algorithm to obtain a dynamic registration matrix. The iterative closest point algorithm finds the best match between two sets of points through continuous iterations, thereby aligning the model with the actual situation. In the facial contour plastic surgery plan, this step can track the movement of facial contour points during the operation and generate a 4x4 transformation matrix to describe the rotation and translation required by the model at each time point. Error analysis is performed on the dynamic registration matrix to obtain registration accuracy evaluation data. Error analysis calculates the deviation between the predicted position and the actual position to quantify the accuracy of the registration. For example, in the mandibular angle plastic surgery plan, the average error and maximum error of the mandibular angle position may be calculated to evaluate the accuracy of the surgical process.

[0124] The registration accuracy assessment data is noise-reduced by the Kalman filter algorithm to obtain smoothed registration parameters. Kalman filtering can effectively remove measurement noise and provide more stable estimates. In facial filler surgery planning, this step can smooth tracking fluctuations caused by tissue deformation during filler injection and provide more continuous and reliable position information. The smoothed registration parameters are analyzed in time series to obtain registration trend data. Time series analysis can reveal the changing patterns and trends in the surgical process. In a comprehensive facial rejuvenation surgery plan, this may involve analyzing the changing trends of various facial regions over time, such as the degree of skin stretching, filler distribution changes, etc. The registration trend data is classified by support vector machine to obtain the surgical process status identification. The support vector machine can classify the surgical process into different states based on the registration trend data. For example, in a rhinoplasty surgery plan, the surgical process may be divided into several key stages such as initial cutting, bone reshaping, and soft tissue adjustment, each of which has its own specific registration trend characteristics. The surgical process status identification is analyzed by decision tree analysis to obtain a real-time feedback instruction set. The decision tree analysis generates guidance for the next operation based on the current state and the predetermined goal. In the planning of zygomatic bone surgery, the decision tree may generate specific instructions such as "move 2 mm medially" or "rotate 3 degrees upward" based on the current zygomatic bone position and target morphology.

[0125] The real-time feedback instruction set is optimized through reinforcement learning algorithms to obtain adaptive navigation strategies. Reinforcement learning learns the optimal decision-making strategy by simulating a large number of surgical scenarios. In complex facial reshaping surgery plans, reinforcement learning can optimize the sequence and parameters of multiple surgical steps to achieve the best overall effect.

[0126] The adaptive navigation strategy is processed into a data stream to obtain a surgical navigation data stream. Data streaming converts discrete strategies into a continuous data stream for real-time processing and transmission. In an actual surgical planning simulation, this may be manifested as generating 60 frames of navigation data per second, including information such as tool position, force feedback, and next step suggestions.

[0127] Finally, the surgical navigation data stream is comprehensively analyzed by a deep neural network to obtain the target plastic surgery plan. Deep neural networks can learn complex patterns and relationships from a large amount of data to generate the final surgical plan. In a comprehensive facial plastic surgery plan, a deep neural network may comprehensively consider multiple factors such as the patient's facial structure, expected results, and surgical risks to generate a detailed list of surgical steps, including the precise parameters, expected results, and potential risks of each step. The plastic surgery plan generation method based on medical images has achieved a transition from static prediction to dynamic navigation, providing comprehensive planning and navigation support for plastic surgery. For example, in a complex facial reconstruction surgery plan, this process can not only accurately predict the final effect of the surgery, but also provide real-time navigation and adjustment suggestions during the simulated operation, making the surgical plan more accurate, safe, and personalized. This dynamic and adaptive plan generation method greatly improves the accuracy and success rate of plastic surgery and provides doctors with a powerful decision support tool.

[0128] The above describes the method for generating a plastic surgery plan based on medical images in the embodiment of the present application. The following describes the device for generating a plastic surgery plan based on medical images in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of a device for generating a plastic surgery plan based on medical images includes:

[0129] The segmentation module 201 is used to pre-process and adaptively segment the acquired multi-modal medical image data to obtain a three-dimensional image data set;

[0130] A reconstruction module 202 is used to perform multi-scale feature fusion reconstruction on the three-dimensional image data set to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data;

[0131] A definition module 203, used for performing interactive target definition based on facial features on the three-dimensional anatomical model to obtain a quantitative surgical target parameter set;

[0132] A processing module 204 is used to perform multi-objective optimization processing on the quantitative surgical target parameter set to obtain an optimized surgical plan;

[0133] A simulation module 205 is used to perform a virtual surgery simulation driven by a physical engine on the three-dimensional anatomical model based on the optimized surgical plan to obtain a postoperative appearance prediction model;

[0134] The analysis module 206 is used to perform dynamic registration and real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generate a target plastic surgery plan based on the surgical navigation data stream.

[0135] Through the collaborative cooperation of the above components, the quality and information extraction efficiency of the original medical images are significantly improved by preprocessing and adaptive segmentation of multimodal medical image data, combined with adaptive histogram equalization, non-local mean filtering and spatial registration based on mutual information maximization, laying a solid foundation for subsequent processing. Secondly, multi-scale feature fusion reconstruction technology, including wavelet transform, deep feature extraction of convolutional neural network, feature pyramid network and self-attention mechanism, is used to achieve high-precision reconstruction of three-dimensional anatomical models, which not only retains high-resolution geometric structure, but also contains multi-level tissue representation information and biomechanical data, providing comprehensive and accurate anatomical information for surgical planning. In the target definition stage, an interactive target definition method based on facial features is introduced. Through principal component analysis, convolutional neural network feature extraction, graph convolutional network topology analysis and other technologies, the precise quantification of surgical targets is achieved, which greatly improves the personalization and accuracy of surgical planning. In the optimization stage, multi-objective optimization processing technology, including hierarchical analysis, genetic algorithm, non-dominated sorting genetic algorithm and fuzzy comprehensive evaluation method, is used to effectively balance multiple goals such as aesthetic effect, safety and recovery speed, and generate the optimal surgical plan. In the virtual surgery simulation, the simulation technology driven by the physical engine, combined with the mass-spring system algorithm and shape matching algorithm, achieved highly realistic soft tissue deformation simulation, greatly improving the accuracy of postoperative effect prediction. Finally, in terms of intraoperative navigation and real-time feedback analysis, through the comprehensive application of iterative nearest point algorithm, Kalman filtering, support vector machine classification and reinforcement learning, high-precision dynamic registration and real-time feedback were achieved, providing accurate navigation support for the surgical process. This series of technological innovations and integrations not only significantly improved the accuracy and reliability of plastic surgery planning, but also realized intelligent support for the entire process from preoperative planning to intraoperative navigation. By comprehensively utilizing multimodal medical imaging, deep learning, computer vision and artificial intelligence technologies, this method effectively overcomes the problems of insufficient information utilization, inaccurate target definition, incomplete optimization, and unrealistic simulation in traditional surgical planning methods, providing doctors with more comprehensive and accurate decision support.

[0136] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the steps of the method for generating a plastic surgery plan based on medical images.

[0137] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0138] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0139] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating a plastic surgery plan based on medical images, characterized in that: The method for generating a plastic surgery plan based on medical images comprises: The acquired multimodal medical image data are preprocessed and adaptively segmented to obtain a three-dimensional image data set, including: performing adaptive histogram equalization processing on the multimodal medical image data to obtain enhanced medical image data; performing non-local mean filtering processing on the enhanced medical image data to obtain denoised medical image data; performing spatial registration based on mutual information maximization on the denoised medical image data to obtain medical image data in a unified coordinate system; inputting the medical image data in the unified coordinate system into a U-Net deep learning network for segmentation to obtain a preliminary segmented image; performing morphological closing operation processing on the preliminary segmented image to obtain an optimized segmented image; performing boundary information extraction on the optimized segmented image by an edge detection algorithm to obtain anatomical structure boundary information; performing surface reconstruction according to the anatomical structure boundary information by an isosurface extraction algorithm to obtain a preliminary three-dimensional pixel grid; performing grid simplification and subdivision operations on the preliminary three-dimensional pixel grid to obtain a first candidate three-dimensional pixel grid; performing texture mapping based on the multimodal medical image data on the candidate three-dimensional pixel grid to obtain a second candidate three-dimensional pixel grid with surface features; performing biomechanical parameter fusion on the second candidate three-dimensional pixel grid to obtain the three-dimensional image data set; The three-dimensional image data set is reconstructed by multi-scale feature fusion to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data, including: performing wavelet transform on the three-dimensional image data set to obtain a multi-scale feature map; performing deep feature extraction on the multi-scale feature map through a convolutional neural network to obtain a deep feature representation; performing multi-level feature fusion on the deep feature representation through a feature pyramid network to obtain a multi-resolution feature map; performing spatial context enhancement on the multi-resolution feature map through a self-attention mechanism to obtain enhanced spatial context information; and performing spatial context enhancement on the enhanced spatial context information. Performing a three-dimensional deconvolution operation to obtain reconstructed voxel data; performing surface reconstruction on the reconstructed voxel data by using a March cube algorithm to obtain a three-dimensional surface mesh; performing Laplace smoothing on the three-dimensional surface mesh to obtain the high-resolution geometric structure; performing surface feature mapping on the high-resolution geometric structure by using a texture mapping algorithm to obtain surface data containing multi-level tissue representation information; performing mechanical property calculations on the surface data containing multi-level tissue representation information by using a finite element analysis algorithm to obtain a data structure containing biomechanical data; performing data compression and index construction on the data structure containing biomechanical data to obtain the three-dimensional anatomical model; The three-dimensional anatomical model is interactively defined based on facial features to obtain a quantitative surgical target parameter set, including: performing key point detection on the three-dimensional anatomical model to obtain a facial feature point set; performing dimensionality reduction processing on the facial feature point set through a principal component analysis algorithm to obtain a facial morphology principal component; performing statistical analysis on the facial morphology principal component to obtain a facial aesthetic scoring index; performing feature extraction on the three-dimensional anatomical model through a convolutional neural network to obtain facial depth feature representation information; performing cluster analysis on the facial depth feature representation information to obtain a facial region segmentation image; performing topological relationship analysis on the facial region segmentation image through a graph convolutional network to obtain a facial structure graph representation; performing graph matching on the facial structure graph representation based on the facial aesthetic scoring index to obtain an expected aesthetic standard deviation value; performing parameterized encoding on the expected aesthetic standard deviation value through a variational autoencoder to obtain an aesthetic parameter space; performing interactive adjustment on the aesthetic parameter space to obtain a personalized aesthetic target; performing shape encoding on the personalized aesthetic target through a preset Fourier descriptor to obtain the quantitative surgical target parameter set; The quantified surgical target parameter set is subjected to multi-objective optimization processing to obtain an optimized surgical plan, including: decomposing the quantified surgical target parameter set to obtain a sub-target set; assigning weights to the sub-target set by means of a hierarchical analysis method to obtain a weighted sub-target set; defining constraints on the weighted sub-target set to obtain a constraint data set; generating an initial solution for the constraint data set by means of a genetic algorithm to obtain a candidate solution set; performing fitness evaluation on the candidate solution set to obtain quality score data corresponding to the candidate solution set; performing Pareto optimization analysis on the quality score data by means of a non-dominated sorting genetic algorithm to obtain a non-inferior solution set; performing cluster analysis on the non-inferior solution set to obtain a representative solution set; performing multi-criteria decision-making on the representative solution set by means of a fuzzy comprehensive evaluation method to obtain an optimal solution; performing surgical step decomposition on the optimal solution to obtain a preliminary surgical plan; refining the preliminary surgical plan by means of a Monte Carlo tree search algorithm to obtain the optimized surgical plan; Based on the optimized surgical plan, a virtual surgical simulation driven by a physical engine is performed on the three-dimensional anatomical model to obtain a postoperative appearance prediction model, including: mesh simplification of the three-dimensional anatomical model to obtain simplified three-dimensional mesh data; discretization of the simplified three-dimensional mesh data to obtain calculation mesh data; material property assignment of the calculation mesh data to obtain calculation data; dynamic parameterization of the calculation data by a particle-spring system algorithm to obtain a deformable body parameter set; boundary condition setting of the deformable body parameter set to obtain deformable body data under constraint conditions; time discretization of the deformable body data under constraint conditions to obtain dynamic simulation data; time series data decomposition of the dynamic simulation data to obtain deformation process time series data; contour alignment of the deformation process time series data by a shape matching algorithm to obtain a standardized deformation data set; statistical analysis of the standardized deformation data set to obtain deformation feature distribution data; appearance synthesis of the deformation feature distribution data by a generative adversarial network to obtain postoperative appearance prediction data; Dynamic registration and real-time feedback analysis are performed on the postoperative appearance prediction model to obtain a surgical navigation data stream, and a target plastic surgery plan is generated based on the surgical navigation data stream, including: extracting feature points from the postoperative appearance prediction model to obtain a key anatomical marker point set; performing real-time position matching on the key anatomical marker point set through an iterative closest point algorithm to obtain a dynamic registration matrix; performing error analysis on the dynamic registration matrix to obtain registration accuracy evaluation data; performing noise suppression on the registration accuracy evaluation data through a Kalman filter algorithm to obtain smoothed registration parameters; performing time series analysis on the smoothed registration parameters to obtain registration trend data; classifying the registration trend data through a support vector machine to obtain a surgical process status identifier; performing decision tree analysis on the surgical process status identifier to obtain a real-time feedback instruction set; optimizing the real-time feedback instruction set through a reinforcement learning algorithm to obtain an adaptive navigation strategy; performing data stream processing on the adaptive navigation strategy to obtain a surgical navigation data stream; and performing comprehensive analysis on the surgical navigation data stream through a deep neural network to obtain a target plastic surgery plan.

2. A device for generating a plastic surgery plan based on medical images, characterized in that: The plastic surgery plan generating device based on medical images comprises: a segmentation module, used for preprocessing and adaptively segmenting the acquired multimodal medical image data to obtain a three-dimensional image data set, including: performing adaptive histogram equalization processing on the multimodal medical image data to obtain enhanced medical image data; performing non-local mean filtering processing on the enhanced medical image data to obtain denoised medical image data; performing spatial registration based on mutual information maximization on the denoised medical image data to obtain medical image data in a unified coordinate system; inputting the medical image data in the unified coordinate system into a U-Net deep learning network for segmentation to obtain a preliminary segmented image; performing morphological closing operation processing on the preliminary segmented image to obtain an optimized segmented image; performing boundary information extraction on the optimized segmented image by an edge detection algorithm to obtain anatomical structure boundary information; performing surface reconstruction according to the anatomical structure boundary information by an isosurface extraction algorithm to obtain a preliminary three-dimensional pixel grid; performing grid simplification and subdivision operations on the preliminary three-dimensional pixel grid to obtain a first candidate three-dimensional pixel grid; performing texture mapping based on the multimodal medical image data on the candidate three-dimensional pixel grid to obtain a second candidate three-dimensional pixel grid with surface features; performing biomechanical parameter fusion on the second candidate three-dimensional pixel grid to obtain the three-dimensional image data set; A reconstruction module is used to perform multi-scale feature fusion reconstruction on the three-dimensional image data set to obtain a three-dimensional anatomical model, wherein the three-dimensional anatomical model includes high-resolution geometric structure, multi-level tissue representation information and biomechanical data, including: performing wavelet transform on the three-dimensional image data set to obtain a multi-scale feature map; performing deep feature extraction on the multi-scale feature map through a convolutional neural network to obtain a deep feature representation; performing multi-level feature fusion on the deep feature representation through a feature pyramid network to obtain a multi-resolution feature map; performing spatial context enhancement on the multi-resolution feature map through a self-attention mechanism to obtain enhanced spatial context information; The following information is subjected to a three-dimensional deconvolution operation to obtain reconstructed voxel data; the reconstructed voxel data is subjected to surface reconstruction by a March cube algorithm to obtain a three-dimensional surface mesh; the three-dimensional surface mesh is subjected to Laplace smoothing to obtain the high-resolution geometric structure; the high-resolution geometric structure is subjected to surface feature mapping by a texture mapping algorithm to obtain surface data containing multi-level tissue representation information; the surface data containing multi-level tissue representation information is subjected to mechanical property calculation by a finite element analysis algorithm to obtain a data structure containing biomechanical data; the data structure containing biomechanical data is subjected to data compression and index construction to obtain the three-dimensional anatomical model; A definition module is used to perform interactive target definition based on facial features on the three-dimensional anatomical model to obtain a quantitative surgical target parameter set, including: performing key point detection on the three-dimensional anatomical model to obtain a facial feature point set; performing dimensionality reduction processing on the facial feature point set through a principal component analysis algorithm to obtain a facial morphology principal component; performing statistical analysis on the facial morphology principal component to obtain a facial aesthetic scoring index; performing feature extraction on the three-dimensional anatomical model through a convolutional neural network to obtain facial depth feature representation information; performing cluster analysis on the facial depth feature representation information to obtain a facial region segmentation image; performing topological relationship analysis on the facial region segmentation image through a graph convolutional network to obtain a facial structure graph representation; performing graph matching on the facial structure graph representation based on the facial aesthetic scoring index to obtain an expected aesthetic standard deviation value; performing parameterized encoding on the expected aesthetic standard deviation value through a variational autoencoder to obtain an aesthetic parameter space; performing interactive adjustment on the aesthetic parameter space to obtain a personalized aesthetic target; performing shape encoding on the personalized aesthetic target through a preset Fourier descriptor to obtain the quantitative surgical target parameter set; a processing module, for performing multi-objective optimization processing on the quantitative surgical target parameter set to obtain an optimized surgical plan, including: decomposing the quantitative surgical target parameter set to obtain a sub-target set; assigning weights to the sub-target set by using a hierarchical analysis method to obtain a weighted sub-target set; defining constraints on the weighted sub-target set to obtain a constraint data set; generating an initial solution for the constraint data set by using a genetic algorithm to obtain a candidate solution set; performing fitness evaluation on the candidate solution set to obtain quality score data corresponding to the candidate solution set; performing Pareto optimization analysis on the quality score data by using a non-dominated sorting genetic algorithm to obtain a non-inferior solution set; performing cluster analysis on the non-inferior solution set to obtain a representative solution set; performing multi-criteria decision-making on the representative solution set by using a fuzzy comprehensive evaluation method to obtain an optimal solution; performing surgical step decomposition on the optimal solution to obtain a preliminary surgical plan; refining the preliminary surgical plan by using a Monte Carlo tree search algorithm to obtain the optimized surgical plan; A simulation module is used to perform a virtual surgery simulation driven by a physical engine on the three-dimensional anatomical model based on the optimized surgical plan to obtain a postoperative appearance prediction model, including: mesh simplification of the three-dimensional anatomical model to obtain simplified three-dimensional mesh data; discretization of the simplified three-dimensional mesh data to obtain calculation mesh data; material property assignment of the calculation mesh data to obtain calculation data; dynamic parameterization of the calculation data by a particle-spring system algorithm to obtain a deformable body parameter set; boundary condition setting of the deformable body parameter set to obtain deformable body data under constraint conditions; time discretization of the deformable body data under constraint conditions to obtain dynamic simulation data; time series data decomposition of the dynamic simulation data to obtain deformation process time series data; contour alignment of the deformation process time series data by a shape matching algorithm to obtain a standardized deformation data set; statistical analysis of the standardized deformation data set to obtain deformation feature distribution data; appearance synthesis of the deformation feature distribution data by a generative adversarial network to obtain postoperative appearance prediction data; The analysis module is used to perform dynamic registration and real-time feedback analysis on the postoperative appearance prediction model to obtain a surgical navigation data stream, and generate a target plastic surgery plan based on the surgical navigation data stream, including: extracting feature points from the postoperative appearance prediction model to obtain a key anatomical marker point set; performing real-time position matching on the key anatomical marker point set through an iterative closest point algorithm to obtain a dynamic registration matrix; performing error analysis on the dynamic registration matrix to obtain registration accuracy evaluation data; performing noise suppression on the registration accuracy evaluation data through a Kalman filter algorithm to obtain smoothed registration parameters; performing time series analysis on the smoothed registration parameters to obtain registration trend data; classifying the registration trend data through a support vector machine to obtain a surgical process status identifier; performing decision tree analysis on the surgical process status identifier to obtain a real-time feedback instruction set; optimizing the real-time feedback instruction set through a reinforcement learning algorithm to obtain an adaptive navigation strategy; performing data stream processing on the adaptive navigation strategy to obtain a surgical navigation data stream; and performing comprehensive analysis on the surgical navigation data stream through a deep neural network to obtain a target plastic surgery plan.

3. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the method for generating a plastic surgery plan based on medical images as described in any one of claim 1 is implemented.

Images