A point cloud completion method and system for the medical field

Abstract

The application discloses a point cloud completion method and system for the medical field, and relates to the technical field of three-dimensional medical image processing; in the application, an encoder is used for dynamic construction of a local neighborhood and extraction of multi-scale geometric features through three-stage graph edge convolution and farthest point downsampling, and then multi-scale splicing fusion and global pooling are performed to accurately capture local anatomical details such as high curvature and complex topology of medical point clouds, and loss of local features caused by a general encoder is avoided; a decoder is used for step-by-step upsampling in a manner of point splitting, displacement prediction and feature interaction, and is used for modeling long-range dependence in cooperation with a self-attention or Mamba architecture, so that the point cloud density is gradually improved while the structural continuity is maintained; and the application enables the network to learn global anatomical morphology and focus on local fine structures, so that the generated completed point cloud is not only rich in structure, free of holes and concave-convex distortion, but also strictly conforms to medical anatomical constraints such as a skull and a malar bone.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional medical image processing technology, specifically a point cloud completion method and system for the medical field. Background Technology

[0002] Three-dimensional reconstruction of the skeleton plays a crucial role in preoperative planning, personalized implant design, and postoperative evaluation. However, due to tumor resection or trauma, the obtained three-dimensional images of the skull, cheekbone, and orbit often contain extensive structural defects. Traditionally, surgeons have relied primarily on manual methods to complete the missing bones, such as by mirroring the contralateral structure or by deforming and registering a standard bone model, while adjusting the local morphology based on their own experience. These manual completion methods are cumbersome, time-consuming, and experience-dependent, making it difficult to guarantee consistent results. When the defect boundaries are highly irregular or the shape is complex, manual methods often fail to accurately and promptly restore details, potentially introducing morphological deviations that affect the accuracy of subsequent implant design and surgical outcomes.

[0003] Deep learning-based point cloud completion technology offers a new solution to the aforementioned problems. Point cloud completion networks, by learning from large amounts of data, can automatically infer complete structures from incomplete input point clouds, transforming the completion process from experience-driven to data-driven. However, existing completion methods still have the following shortcomings: 1. Many general-purpose point cloud completion networks (such as PCN, PF-Net, SnowflakeNet, AdaPoinTr, etc.) have achieved certain results in global morphology prediction, but they lack consideration for medical anatomical constraints: for critical areas such as high curvature regions or defect edges, conventional methods are prone to morphological deviations. For example, when the defect is large, the model may not be able to accurately reconstruct the complex surface details at the defect boundary, and the generated completed surface may be too smooth or have discontinuous bumps and holes.

[0004] 2. Existing methods mostly use the standard Chamfer distance or Earth Mover distance as the loss function, treating the error of all points equally and failing to give extra attention to the key anatomical regions, thus making it difficult to guarantee the reconstruction accuracy of these regions.

[0005] 3. Most point cloud completion networks employ a progressively refined, multi-stage output architecture (generating point clouds from coarse to fine). During training, the losses of each stage's output are typically simply added together with fixed weights. If the errors of each stage differ significantly, the fixed weights are insufficient to adequately address training needs: either the early coarse-grained stages have excessively large errors that dominate the loss, hindering the learning of details; or the later fine-grained stages do not receive sufficient weights, resulting in inadequate optimization of details by the model. The lack of an adaptive weight adjustment mechanism may lead to unstable training or suboptimal convergence.

[0006] 4. Since the parts preserved in the input point cloud (the intact anatomical structures) should be fully preserved in the completion result, without special constraints, the network may shift or deform the existing parts of the structure when pursuing global shape matching. For example, the model may slightly move the original point position in order to optimize the overall Chamfer distance, resulting in the completion result not being completely aligned with the known parts of the input.

[0007] This invention provides a point cloud completion method and system for the medical field to solve the above-mentioned technical problems. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a point cloud completion method and system for the medical field.

[0009] To achieve the above objectives, a first aspect of the present invention provides a point cloud completion method for the medical field, comprising: Obtain the residual defect cloud after data preprocessing; wherein, the residual defect cloud is a three-dimensional coordinate matrix; The residual point cloud is input into a pre-trained point cloud completion network to obtain a multi-stage predicted point cloud. The point cloud completion network includes a feature encoder, a seed point generator, and a multi-stage point cloud decoder. The feature encoder extracts the global point cloud feature vector of the residual point cloud. The seed point generator generates an initial point cloud based on the global point cloud feature vector. The multi-stage point cloud decoder upsamples the initial point cloud stage by stage under the guidance of the global point cloud feature vector to obtain a multi-stage predicted point cloud, where each stage's predicted point cloud is a three-dimensional coordinate matrix. Extract the highest-level predicted point cloud from the multi-stage predicted point cloud and use it as the point cloud completion result.

[0010] In one possible implementation, the feature encoder includes a sampling module, an input feature transformation module, a first graph-edge convolution module, a first farthest point downsampling module, a second graph-edge convolution module, a second farthest point downsampling module, a third graph-edge convolution module, a third farthest point downsampling module, a multi-scale feature splicing and fusion module, an output feature transformation module, and a global pooling module.

[0011] In one possible implementation, the data processing flow of the feature encoder includes: The three-dimensional coordinate matrix of the residual point cloud is input to the sampling module, and the three-dimensional coordinate matrix of the sampled point cloud is obtained by sampling using the farthest point sampling strategy; wherein, the dimension of the three-dimensional coordinate matrix of the residual point cloud is... , The number of points in the residual point cloud is given, and the dimension of the three-dimensional coordinate matrix of the sampled point cloud is 5120×3. The three-dimensional coordinate matrix of the sampled point cloud is input to the input feature transformation module to obtain the initial feature matrix; wherein the dimension of the initial feature matrix is ​​5120×8; A local neighborhood of a point is constructed based on K-nearest neighbors; the local neighborhood and the initial feature matrix are input into the first graph edge convolution module to extract edge convolution features, resulting in a first-scale feature matrix; wherein the dimension of the first-scale feature matrix is ​​5120×64, and K in K-nearest neighbors is 16; The first scale feature matrix is ​​input into the first farthest point downsampling module for downsampling to obtain the first downsampled feature matrix, wherein the dimension of the first downsampled feature matrix is ​​512×64; The first downsampled feature matrix is ​​input into the second graph edge convolution module to obtain the second scale feature matrix; the second scale feature matrix is ​​downsampled by the second farthest point downsampled module to obtain the second downsampled feature matrix; wherein, the dimension of the second scale feature matrix is ​​512×128 and the dimension of the second downsampled feature matrix is ​​256×128. The second downsampled feature matrix is ​​input into the third graph edge convolution module to obtain the third scale feature matrix; the third scale feature matrix is ​​downsampled by the third farthest point downsampled module to obtain the high-level feature matrix; wherein, the dimension of the third scale feature matrix is ​​256×256, and the dimension of the high-level feature matrix is ​​128×512. The first-scale feature matrix, the second-scale feature matrix, the third-scale feature matrix, and the high-level feature matrix are input into the multi-scale feature splicing and fusion module for splicing and fusion to obtain the fused feature matrix; wherein, the dimension of the fused feature matrix is ​​128×960; The fused feature matrix is ​​input to the output feature transformation module to obtain the point feature matrix; wherein the dimension of the point feature matrix is ​​128×D, and D is the preset feature dimension; The point feature matrix is ​​input into the global pooling module for max pooling to obtain the global point cloud feature vector; wherein the dimension of the global point cloud feature vector is 1×D.

[0012] In one possible implementation, the seed point generator includes a point feature expansion module, a feature splicing module, a residual mapping module, and a coordinate generation module.

[0013] In one possible implementation, the data processing flow of the seed point generator includes: The global feature vector is input to the point feature expansion module to obtain the point feature matrix; wherein the dimension of the point feature matrix is... , The number of seed points; The point feature matrix and the globally copied point cloud feature vectors are input into the feature stitching module and stitched along the feature dimension to obtain a first stitched feature matrix. The first stitched feature matrix is ​​then nonlinearly mapped by the residual mapping module to obtain an enhanced point feature matrix. The dimension of the first stitched feature matrix is... The dimension of the enhanced point feature matrix is ; The enhanced point feature matrix is ​​input into the coordinate generation module to obtain the three-dimensional coordinate matrix of the seed point cloud; wherein, the dimension of the three-dimensional coordinate matrix of the seed point cloud is... ; The 3D coordinate matrix of the seed point cloud and the 3D coordinate matrix of the residual point cloud are merged to obtain the 3D coordinate matrix of the merged point cloud; the farthest point is sampled from the 3D coordinate matrix of the merged point cloud to obtain the 3D coordinate matrix of the initial point cloud; wherein, the dimension of the 3D coordinate matrix of the merged point cloud is... The dimension of the initial point cloud's three-dimensional coordinate matrix is , To select the number of representative points from the merged point cloud.

[0014] In one possible implementation, the point feature matrix and the point-by-point copied global point cloud feature vector are input to a feature stitching module for stitching along the feature dimension, including: Copy the global point cloud feature vector in the row direction. This yields the global feature replication matrix; where the dimension of the global feature replication matrix is... ; The global feature copy matrix and the point feature matrix are concatenated along the column direction to obtain a first concatenated feature matrix; wherein the dimension of the first concatenated feature matrix is... .

[0015] In one possible implementation, the multi-stage point cloud decoder includes a first-stage point cloud deconvolution module, a second-stage point cloud deconvolution module, a third-stage point cloud deconvolution module, and a fourth-stage point cloud deconvolution module, with each stage's point cloud deconvolution module having the same module structure.

[0016] In one possible implementation, the data processing flow of the multi-stage point cloud decoder includes: The three-dimensional coordinate matrix of the initial point cloud is input into the first-stage point cloud deconvolution module to obtain the three-dimensional coordinate matrix of the first-stage predicted point cloud. The three-dimensional coordinate matrix of the predicted point cloud in the first stage is input into the deconvolution module of the point cloud in the second stage to obtain the three-dimensional coordinate matrix of the predicted point cloud in the second stage. The three-dimensional coordinate matrix of the predicted point cloud in the second stage is input into the deconvolution module of the point cloud in the third stage to obtain the three-dimensional coordinate matrix of the predicted point cloud in the third stage. The three-dimensional coordinate matrix of the predicted point cloud in the third stage is input into the deconvolution module of the third stage point cloud to obtain the three-dimensional coordinate matrix of the predicted point cloud in the fourth stage; wherein, the predicted point cloud in the fourth stage is used as the predicted point cloud in the highest stage.

[0017] In one possible implementation, the point cloud deconvolution module includes a first point feature extraction module, a feature fusion module, a feature interaction module, a point splitting module, a feature upsampling module, a displacement feature generation module, and a displacement prediction module.

[0018] In one possible implementation, the data processing flow of the point cloud deconvolution module includes: Let the dimension of the three-dimensional coordinate matrix of the input point cloud at the current stage be... The three-dimensional coordinate matrix is ​​input into the first point feature extraction module to obtain the first point feature matrix; wherein, the dimension of the first point feature matrix is... , Determined based on the stage of the point cloud deconvolution module. For stage indexing; The first point feature matrix, the global aggregated features, and the point-replicated global point cloud feature vectors are concatenated along the feature dimension to obtain the second concatenated feature matrix; wherein, the global aggregated features are obtained by max pooling the first feature matrix along the point dimension, and the dimension of the second concatenated matrix is... ; The second concatenated matrix is ​​input into the feature fusion module to obtain the query feature matrix; wherein the dimension of the query feature matrix is... ; The query feature matrix is ​​input to the feature interaction module for feature interaction to obtain the interaction feature matrix; wherein, the dimension of the interaction feature matrix is... The feature interaction module adopts a self-attention mechanism or a feature interaction mechanism based on the Mamba architecture; The interaction feature matrix is ​​input into the point splitting module, and then analyzed according to the upsampling factor. Perform point-level splitting to obtain sub-point feature matrices; where the dimension of the sub-point feature matrices is... , The upsampling factor. The deconvolution module of the point cloud is set from low to high as 1, 1, 2, 4. The interaction feature matrix is ​​input into the feature upsampling module, and an upsampling interpolation strategy is used to obtain the upsampled interaction feature matrix; wherein, the dimension of the upsampled interaction feature matrix is... Upsampling interpolation strategies include nearest neighbor interpolation or bilinear interpolation; The sub-point feature matrix and the upsampled interactive feature matrix are concatenated along the feature dimension to obtain a third concatenated feature matrix; wherein the dimension of the third concatenated feature matrix is... ; The third concatenated feature matrix is ​​input into the displacement feature generation module to obtain the displacement feature matrix; wherein the dimension of the displacement feature matrix is... ; The displacement feature matrix is ​​input into the displacement prediction module to obtain a three-dimensional displacement matrix; wherein the dimension of the three-dimensional displacement matrix is... ; The three-dimensional displacement matrix is ​​added point by point to the three-dimensional coordinate matrix of the upsampled input point cloud to obtain the three-dimensional coordinate matrix of the output point cloud at the current stage.

[0019] In one possible implementation, the first point feature matrix, the global aggregated features, and the point-to-point copied global point cloud feature vectors are concatenated along the feature dimension, including: Max pooling is performed on the first point feature matrix along the point dimension to obtain the global aggregated feature; wherein the dimension of the global aggregated feature is... ; Copy the 1×D global point cloud feature vector in the row direction. This yields the global feature replication matrix; where the dimension of the global feature replication matrix is... ; The first feature matrix, the global aggregated feature matrix, and the global feature copy matrix are concatenated along the column direction to obtain the second concatenated feature matrix; the dimension of the second concatenated feature matrix is... .

[0020] In one possible implementation, the displacement prediction module includes an edge constraint submodule, which performs bounded processing on the three-dimensional displacement matrix and scales it according to the stage index.

[0021] In one possible implementation, the training process of the point cloud completion network includes: Medical point cloud data is preprocessed to construct a training dataset of the incomplete point cloud to be completed and the ground truth point cloud corresponding to the incomplete point cloud. The residual point clouds in the training dataset are input into the point cloud completion network to obtain the predicted point clouds in four stages; wherein, the predicted point cloud is a three-dimensional coordinate matrix; Based on the predicted point cloud and the ground truth point cloud, the medical geometric perception quantity and the missing correlation quantity are calculated, and point-level weights are generated based on the medical geometric perception quantity and the missing correlation quantity. The reconstruction error between the predicted point cloud and the ground truth point cloud at each stage is weighted using the point-level weights to obtain the loss at each stage; dynamic stage weights are determined based on the stage loss to sum the weighted losses at each stage to obtain the multi-stage main loss; wherein, the dynamic stage weights at each stage are non-negative and sum to 1. A Partial Matching loss branch is introduced to perform one-way nearest neighbor matching between the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, and the matching error is obtained. The multi-stage main loss and the matching error are summed to obtain the total loss. The point cloud completion network is then trained and updated using backpropagation based on the total loss to complete the training of the point cloud completion network.

[0022] In one possible implementation, the data preprocessing includes coordinate normalization, noise reduction, downsampling, scale standardization, and coordinate system alignment.

[0023] In one possible implementation, the generation of the point-level weights includes: A local neighborhood of any point in the predicted point cloud is constructed based on K-nearest neighbors; where K=16 or K=32. Based on the local neighborhood, medical geometric perception quantities representing curvature and normal changes are calculated, and missing correlation quantities are calculated based on the distance from the point to the residual cloud. Based on the medical geometric perception quantity and the missing correlation quantity, a point-level original weight score is constructed, and the point-level original weight score is bounded normalized to obtain the point-level weight.

[0024] In one possible implementation, the calculation of the loss at each stage includes: Alignment sampling is performed on the ground value point cloud based on the predicted point cloud at each stage to obtain the ground value point set; wherein, the alignment sampling adopts the farthest point sampling strategy, and the number of points in the predicted point cloud and the ground value point set is the same; The chamfer distance loss between the predicted point cloud and the ground truth point set is calculated based on the point-level weights and used as the corresponding stage loss; wherein the distance metric in the chamfer distance loss adopts the L1 norm.

[0025] In one possible implementation, dynamic stage weights are determined based on stage losses, including: A stage evaluation metric is constructed based on the stage loss; wherein the stage evaluation metric is a monotonically decreasing function of the stage loss. The stage evaluation quantity is linearly transformed to obtain the stage score; the stage score is then substituted into a Softmax function with a temperature parameter to obtain the dynamic stage weight.

[0026] In one possible implementation, the stage evaluation metric is constructed using a stage loss derived from a moving average, in order to suppress dynamic stage weight oscillations.

[0027] In one possible implementation, the dynamic stage weights are iteratively updated according to a preset update cycle during the training of the point cloud completion network, and adaptively change with the training process based on weight smoothing constraints and temperature annealing strategies; wherein, the weight smoothing constraints are used to limit the change range of dynamic stage weights within adjacent update cycles, and the temperature annealing strategy is used to gradually reduce the temperature parameter with each training round to improve the discriminativeness of weight allocation in each stage.

[0028] In one possible implementation, performing a one-way nearest neighbor matching between the residual point cloud and the highest-stage prediction point cloud in the prediction point cloud includes: The set of observation points is determined from the residual cloud as the query point set; In the highest stage of the predicted point cloud, a nearest neighbor search is performed for each query point in the query point set to obtain the corresponding nearest neighbor predicted point. The positive matching error is calculated based on the distance metric between the query point and the corresponding nearest neighbor prediction point, and this positive matching error is used as the matching error; wherein, the distance metric is the L1 norm.

[0029] In one possible implementation, after calculating the positive matching error, the predicted points in the predicted point cloud are used as query points, and a nearest neighbor search is performed for each query point in the set of observation points to obtain the corresponding nearest neighbor predicted points. The reverse matching error is calculated based on the distance metric between the query point and the corresponding nearest neighbor prediction point; the forward matching error and the reverse matching error are weighted by a preset weighting coefficient to obtain the matching error; wherein the preset weighting coefficient is used to balance the contribution of the forward matching error and the directional matching error.

[0030] In one possible implementation, the formula for calculating the total loss is: ;in, For multi-stage main losses, These are the weight coefficients for the Partial Matching loss branch. The value range is [0, 1].

[0031] A second aspect of the present invention provides a point cloud completion system for the medical field, comprising: Data acquisition and processing module: used to acquire raw image data through the image acquisition interface, preprocess the raw image data, and obtain the incomplete cloud to be filled; Multi-stage point cloud completion module: Used to perform multi-stage point cloud prediction on the incomplete point cloud to be completed through the point cloud completion network to obtain multi-stage predicted point cloud; extract the highest stage predicted point cloud in the multi-stage predicted point cloud as the point cloud completion result. The medical geometry-defect coupling weight generation module is used to construct a local neighborhood of any point in the predicted point cloud based on K-nearest neighbors; calculate medical geometry perception quantities representing curvature and normal changes based on the local neighborhood; calculate defect-related quantities based on the distance from the point to the defect point cloud; and, Based on the medical geometric perception quantity and the missing correlation quantity, a point-level original weight score is constructed, and the point-level original weight score is bounded normalized to obtain the point-level weight. Multi-stage loss adaptive fusion evaluation module: used to weight the reconstruction error between the predicted point cloud and the ground truth point cloud at each stage using point-level weights to obtain the loss at each stage; and, Dynamic stage weights are determined based on stage losses, and the weighted summation of losses from each stage yields the multi-stage master loss. Locally observable consistency matching loss module: This module introduces the Partial Matching loss branch, performs one-way nearest neighbor matching between the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, and obtains the matching error. Training optimization and parameter update module: used to optimize network parameters of the point cloud completion network based on the total loss; where the total loss is obtained by summing the multi-stage main loss and matching error.

[0032] By employing the above technical solutions, this invention provides a point cloud completion method and system for the medical field. Through a dynamic graph convolution, a long-range dependency modeling encoder-decoder framework, and a loss function design based on medical geometric perception and dynamic weight optimization, it can dynamically balance multi-scale reconstruction errors and strengthen constraints on key geometric regions during neural network training. This invention possesses at least the following beneficial effects: 1. This invention employs a local geometry-aware encoder and a multi-level upsampling decoder architecture to enhance completion performance from two dimensions: feature extraction and structure generation. The encoder dynamically constructs local neighborhoods and extracts multi-scale geometric features through three-level graph-edge convolution and farthest-point downsampling. These features are then multi-scale stitching and global pooling to accurately capture local anatomical details such as high curvature and complex topology in medical point clouds, avoiding the loss of local features by general encoders. The decoder uses a four-stage point cloud deconvolution module, upsampling step-by-step through a combination of point splitting, displacement prediction, and feature interaction. Combined with self-attention or Mamba architecture to model long-range dependencies, it ensures that the point cloud density gradually increases while maintaining structural continuity. This invention enables the network to learn both global anatomical morphology and focus on local fine structures, resulting in a completed point cloud that is not only structurally rich and free of holes and distortions, but also strictly conforms to medical anatomical constraints such as the skull and cheekbone, thus improving the overall integrity and anatomical rationality of the completed morphology.

[0033] 2. This invention constructs a point-level weighting mechanism based on medical geometric perception and defect correlation to specifically improve the reconstruction accuracy of key regions. It constructs a local neighborhood of the predicted point through K-nearest neighbors, calculates medical geometric perception quantities such as curvature and normal changes, and identifies key regions such as high curvature and anatomical transitions. Simultaneously, it calculates the distance from the point to the defect cloud, locates the defect boundary and the core completion region, and fuses and normalizes these two types of indicators to generate point-level weights, amplifying the chamfer distance loss. This invention breaks the limitation of traditional loss functions that treat all point errors equally, allowing the network to focus on optimizing the reconstruction errors of high curvature details and defect boundaries during training, avoiding problems such as over-smoothing and morphological deviations in these areas. Compared to general point cloud completion methods, this invention can accurately restore the subtle features of medical anatomical structures, allowing the completed region to connect naturally with the healthy region, significantly improving the reconstruction accuracy and clinical anatomical rationality of key anatomical parts, and meeting the stringent standards of detail restoration in medical applications.

[0034] 3. This invention achieves adaptive adjustment of multi-stage loss through dynamic stage weights, optimizing the model training process and convergence effect. During training, the weighted chamfer distance loss for each stage is first calculated, and a moving average smoothing loss is used to suppress oscillations. Then, a stage evaluation metric is constructed based on the loss, and dynamic stage weights are generated using a Softmax function with a temperature parameter. Combined with weight smoothing constraints and a temperature annealing strategy, the weights adaptively change with the training process. In the early stages of training, the temperature parameter is higher, and the weights are evenly distributed, with the network focusing on learning the global skeleton morphology to ensure consistency in coarse-grained completion. In the later stages of training, the temperature decreases, and the weights tilt towards higher-precision stages, with the network focusing on local fine-structure optimization. This invention solves the problem of coarse-stage error dominating and insufficient fine-stage optimization in traditional fixed-weight training, achieving coordinated optimization of global morphology and local details. This makes the training process more stable and converges faster, while avoiding the model getting trapped in local optima, thus improving the overall performance and robustness of the multi-stage completion network.

[0035] 4. This invention introduces an independent Partial Matching loss branch, providing dedicated fidelity constraints for healthy anatomical structures. This branch uses the set of observation points in the residual point cloud as the query object, performing one-way / two-way nearest neighbor matching in the highest-stage predicted point cloud, calculating the distance error between the observation points and the predicted points, and incorporating it into the total loss. Independent of the main loss branch, this invention specifically constrains the network from shifting or altering the original healthy anatomical structure, preventing the network from moving existing observation point positions to optimize the global distance. This ensures, from a loss perspective, that the completion result strictly preserves the patient's healthy skeletal morphology. Compared to methods without observation region constraints, the completion result of this invention is completely aligned with the original healthy region, without misalignment or deformation issues, significantly improving the reliability and clinical safety of the completion result. It can be directly used in high-precision clinical applications such as surgical planning and prosthesis fabrication, effectively reducing the medical risks caused by completion deviations. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the training process of the point cloud completion network in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of the point cloud completion network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the data processing flow of the feature encoder in the point cloud completion network according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the data processing flow of the seed point generator in an embodiment of the present invention; Figure 5 This is a schematic diagram of the data processing flow of the point cloud deconvolution module in the multi-stage point cloud decoder in this embodiment of the invention; Figure 6 This is a schematic diagram of the point-level weight generation process in an embodiment of the present invention; Figure 7 This is a schematic diagram of the calculation process for the multi-stage main loss in an embodiment of the present invention; Figure 8 This is a schematic diagram of the process for calculating the matching error using the Partial Matching loss branch in an embodiment of the present invention; Figure 9 This is a schematic diagram of the system structure of a point cloud completion system for the medical field in an embodiment of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1 The first aspect of this invention provides a point cloud completion method for the medical field, comprising: acquiring a damaged point cloud to be completed; inputting the damaged point cloud into a pre-trained point cloud completion network to obtain a multi-stage predicted point cloud; wherein the point cloud completion network includes a feature encoder, a seed point generator, and a multi-stage point cloud decoder; the feature encoder is used to extract a global point cloud feature vector of the damaged point cloud; the seed point generator is used to generate an initial point cloud based on the global point cloud feature vector; the multi-stage point cloud decoder is used to perform stage-by-stage upsampling of the initial point cloud under the guidance of the global point cloud feature vector to obtain a multi-stage predicted point cloud, wherein the predicted point cloud of each stage is a three-dimensional coordinate matrix; and extracting the highest stage predicted point cloud from the multi-stage predicted point cloud as the point cloud completion result.

[0040] In practice, the incomplete point cloud to be completed needs to be acquired first. This incomplete point cloud is usually obtained from raw image data acquired by medical imaging equipment, such as CT or MRI scanners. The raw image data is preprocessed into a three-dimensional coordinate matrix through image segmentation, surface extraction, and other preprocessing. Before being input into the point cloud completion network, the incomplete point cloud undergoes data preprocessing to adapt it to the input requirements of the point cloud completion network. Data preprocessing includes coordinate normalization, noise reduction, downsampling, and scale normalization. After data preprocessing, the incomplete point cloud is input into the point cloud completion network in the form of a fixed-dimensional three-dimensional coordinate matrix to obtain the corresponding point cloud completion result.

[0041] Figure 1 This diagram illustrates the training process of a point cloud completion network, which includes key steps such as data processing, multi-stage prediction, and network parameter optimization. The core of this point cloud completion network lies in two aspects: firstly, it achieves progressive point cloud completion from coarse to fine through a feature encoder based on dynamic graph convolution and a multi-stage point cloud decoder; secondly, it achieves an adaptive balance between error amplification in key regions and multi-stage training through medical geometric perception loss calculation and a dynamic stage weight adjustment mechanism.

[0042] Figure 2 A schematic diagram of the overall structure of the point cloud completion network. (Example) Figure 2The point cloud completion network adopts an encoder-decoder architecture. After the point cloud completion network is trained, the incomplete point cloud to be completed is input into the feature encoder to extract the global point cloud feature vector. This global point cloud feature vector encodes the overall combined shape and topological information of the incomplete point cloud. Subsequently, the seed point generator generates several seed points based on the global point cloud feature vector, and merges them with the incomplete point cloud and performs farthest point sampling to obtain the initial point cloud. The initial point cloud is then passed through four cascaded point cloud deconvolution modules (MPCD) in the multi-stage point cloud decoder, and upsampled stage by stage to output the first-stage predicted point cloud, the second-stage predicted point cloud, the third-stage predicted point cloud, and the fourth-stage predicted point cloud. The number of points in each stage of the predicted point cloud increases with the stage, thereby achieving progressive completion from coarse to fine. Finally, the fourth-stage predicted point cloud, that is, the highest-stage predicted point cloud, serves as the point cloud completion result corresponding to the incomplete point cloud.

[0043] Figure 3 This diagram illustrates the data processing flow of the feature encoder in a point cloud completion network. The feature encoder employs a dynamic graph convolution-based encoding structure, aiming to extract feature vectors with local geometric awareness and global representation capabilities from incomplete point clouds. Dynamic graph convolution can dynamically construct local neighborhoods of points at various stages of feature extraction and extract the relative geometric features between points and their local neighborhoods through edge convolution, thereby enhancing the sensitivity of the point cloud completion network to local curvature changes and detailed structures.

[0044] like Figure 3 As shown, the feature encoder includes a sampling module c11, an input feature transformation module t11, a first graph edge convolution module e11, a first farthest point downsampling module f11, a second graph edge convolution module e12, a second farthest point downsampling module f12, a third graph edge convolution module e13, a third farthest point downsampling module f13, a multi-scale feature splicing and fusion module k11, an output feature transformation module t12, and a global pooling module q11.

[0045] like Figure 3 As shown, the data processing flow of the feature encoder includes: T01: Let the dimension of the three-dimensional matrix of the residual cloud be... ×3, will The residual point cloud 3D coordinate matrix is ​​input to the sampling module c11 for sampling, resulting in a 5120×3 sampling point cloud 3D coordinate matrix; where, The number of points in the defective cloud; The sampling module c11 adopts the farthest point sampling strategy to select representative points from the input residual defect cloud, so that the sampled points can cover the main spatial distribution of the residual defect cloud. T02: Input the 5120×3 sampling point cloud 3D coordinate matrix into the input feature transformation module t11 to obtain the 5120×8 initial feature matrix; The input feature transformation module t11 is usually a one-dimensional convolutional layer that maps three-dimensional coordinates to a higher-dimensional feature space, providing a basic representation for subsequent feature extraction. T03: Construct local neighborhoods of points based on K-nearest neighbors; input the 5120×8 initial feature matrix into the first graph edge convolution module e11 to extract edge features, and obtain a 5120×64 first-scale feature matrix; where the dimension of the first-scale feature matrix is ​​5120×64, and K takes the value of 16 in K-nearest neighbors; The first edge convolution module e11 extracts local geometric structure information by calculating the relative feature differences between a point and its local neighborhood points. For each point, the difference between the point and its local neighborhood points is calculated in its K nearest neighbors (K=16), and nonlinear mapping is performed through a multilayer perceptron. Finally, max pooling is performed on the neighborhood dimension to obtain the edge convolution feature of the point. T04: The 5120×64 first-scale feature matrix is ​​downsampled by the first farthest point downsampling module f11 to obtain a 512×64 first-downsampled feature matrix; the first farthest point downsampling module f11 adopts the farthest point sampling strategy; T05: Input the 512×64 first downsampled feature matrix into the second graph edge convolution module e12 to obtain a 512×128 second scale feature matrix, and then downsample it through the second farthest point downsampled module f12 to obtain a 256×128 second downsampled feature matrix. The second edge convolution module e12 performs local neighborhood feature extraction again on a sparser point set, further abstracting geometric features; T06: Input the 256×128 second downsampled feature matrix into the third graph edge convolution module e13 to obtain a 256×256 third scale feature matrix, and then downsample it through the third farthest point downsampled module f13 to obtain a 128×512 high-level feature matrix. After multiple layers of graph-edge convolution and downsampling, the point cloud completion network gradually aggregates spatial information over a larger range to form high-level abstract features; T07: Align the points of the 5120×64 first-scale feature matrix, the 512×128 second-scale feature matrix, the 256×256 third-scale feature matrix and the 128×512 high-level feature matrix, and input them into the multi-scale feature splicing and fusion module k11 for splicing and fusion to obtain a 128×960 fused feature matrix. The multi-scale feature stitching and fusion module k11 first aligns the first-scale feature matrix, the second-scale feature matrix, the third-scale feature matrix, and the high-level feature matrix to 128 points by sampling the farthest point, and then stitches them together in the feature dimension to obtain the fused feature matrix. T08: Input the 128×960 fused feature matrix into the output feature transformation module t12 to obtain a 128×D point feature matrix; The output feature transformation module t12 is usually a one-dimensional convolutional layer that maps the concatenated high-dimensional features to a preset feature dimension D, further refining the feature representation. It should be noted that the feature dimension specifically refers to the feature dimension of the global encoding output by the encoder. The value of the feature dimension will change the richness of the encoding; the larger the value, the more detailed features are retained, but the time complexity will increase exponentially. The setting of the feature dimension depends on the complexity of the original data. It can be determined that the selectable range is between 128 and 2048. When D is 1024, good experimental results are usually obtained while ensuring moderate efficiency, which can be regarded as the preferred value.

[0046] T09: Input the 128×D point feature matrix into the global pooling module q11 for max pooling to obtain a 1×D global point cloud feature vector.

[0047] The global pooling module q11 performs max pooling on the point dimension, extracting the global maximum value from the features of 128 points to form a single global feature vector.

[0048] Through the above data processing flow, the feature encoder will input... The residual point cloud is gradually converted into a 1×D global point cloud feature vector. This feature vector is repeatedly reused in the subsequent seed point generation and multi-stage decoding process, providing the network with consistent global context information.

[0049] Compared to traditional PointNet or PointNet++ encoders, encoders based on dynamic graph convolution can better capture local geometric details and neighborhood relationships, thereby improving the ability to represent features of high curvature regions and defect boundaries in medical point clouds.

[0050] Figure 4 This is a schematic diagram of the data processing flow for the seed point generator. The seed point generator's role is to generate a set of initial seed points based on the global point cloud feature vector, serving as the starting point for subsequent multi-stage decoding. The seed point generator includes a point feature expansion module e21, a feature concatenation module c21, a residual mapping module d21, and a coordinate generation module g21.

[0051] like Figure 4 As shown, the data processing flow of the seed point generator is as follows: Z01: Input the 1×D global point cloud feature vector into the point feature extension module e21 to obtain... Point feature matrix; where, The number of seed points; The point feature extension module e21 typically employs transposed convolution (also known as deconvolution) to expand a single global point cloud feature vector into... The feature matrix consists of 10 points; in the specific implementation, the kernel size of the transposed convolution is set to 1. Step size set to This expands the 1×D global point cloud feature vector to Point feature matrix; It should be noted that the number of seed points describes the initial point cloud. The global feature vector matrix is ​​gradually reduced in dimensionality through a multilayer perceptron, and then a series of operations such as concatenation, convolution, and activation are used to obtain the initial sparse point cloud. The number of seed points determines the sparsity of the initial point cloud and the decoding difficulty. The default value is 256, which depends on a fixed proportion of the ground truth point cloud.

[0052] Z02: Feature splicing module c21 will... The point feature matrix and the 1×D global point cloud feature vector copied point by point are concatenated along the feature dimension to obtain... The first concatenated feature matrix; In the specific implementation, the 1×D global point cloud feature vector is copied in the row direction. Next, get Copy the global feature matrix, and then Global feature replication matrix and The point feature matrices are concatenated along the column direction (feature dimension) to obtain... First concatenated feature matrix; Z03: Will The first concatenated feature matrix is ​​input to the residual mapping module d21 for nonlinear mapping, resulting in... Enhanced point feature matrix; The residual mapping module d21 typically consists of multiple one-dimensional convolutional layers and non-linear activation functions (such as the ReLU function), used to further refine and enhance the first concatenated feature matrix. The design of the residual mapping module d21 adopts the residual connection idea, which retains some input features during the non-linear mapping process, thus avoiding the gradient vanishing problem in deep network training. Z04: Will The enhanced point feature matrix is ​​input into the coordinate generation module g21 to obtain... The 3D coordinate matrix of the seed point cloud; The coordinate generation module g21 is typically a one- or multi-layer one-dimensional convolution that maps the enhanced point feature matrix to a three-dimensional coordinate space to generate... The three-dimensional coordinates of the seed points; Z05: Will The 3D coordinate matrix of the seed point cloud and By merging the three-dimensional coordinate matrices of the residual cloud, we obtain Merge the 3D coordinate matrix of the point cloud; for By merging the 3D coordinate matrix of the point cloud and sampling the farthest point, we obtain... The 3D coordinate matrix of the initial point cloud; The farthest point sampling strategy here selects from the merged point cloud. A number of representative points are selected to distribute these points as evenly as possible in space, covering both the completed area generated by the seed point and the observation area of ​​the residual cloud.

[0053] Number of representative points for and The number of points sampled from the furthest point after merging. The optimal value is 2048, which ensures a smoother upsampling process.

[0054] Figure 5 This diagram illustrates the data processing flow of the point cloud deconvolution module in a multi-stage point cloud decoder. The multi-stage point cloud decoder is a core component of the point cloud completion network of this invention, used to upsample and refine the initial input point cloud to generate a predicted point cloud with a higher point count. The multi-stage point cloud decoder employs a combination of point splitting and displacement prediction, gradually increasing the point cloud density while maintaining the continuity of the combined structure. The multi-stage point cloud decoder includes a first-stage point cloud deconvolution module, a second-stage point cloud deconvolution module, a third-stage point cloud deconvolution module, and a fourth-stage point cloud deconvolution module, with each stage's point cloud deconvolution module having the same module structure.

[0055] The data processing flow of the multi-stage point cloud decoder includes: D01: Will The initial 3D coordinate matrix of the point cloud is input into the first-stage point cloud deconvolution module, and the output is... The first stage predicts the 3D coordinate matrix of the point cloud; where the upsampling factor of the first stage point cloud deconvolution module is 1. D02: Will The 3D coordinate matrix of the predicted point cloud in the first stage is input into the point cloud deconvolution module in the second stage to obtain... The second stage predicts the 3D coordinate matrix of the point cloud; where the upsampling factor of the second stage point cloud deconvolution module is 1. D03: Will The second-stage predicted point cloud 3D coordinate matrix is ​​input into the third-stage point cloud deconvolution module to obtain... The third stage predicts the 3D coordinate matrix of the point cloud; where the upsampling factor of the third stage point cloud deconvolution module is 2; D04: Will The 3D coordinate matrix of the predicted point cloud in the third stage is input into the deconvolution module of the point cloud in the fourth stage to obtain... The fourth stage predicts the 3D coordinate matrix of the point cloud; where the upsampling factor of the fourth stage point cloud deconvolution module is 4.

[0056] Each stage of the point cloud deconvolution module includes a first point feature extraction module e31, a feature fusion module q31, a feature interaction module i31, a point splitting module m31, a feature upsampling module g31, a displacement feature generation module s31, and a displacement prediction module p31.

[0057] like Figure 5 The data processing flow of the point cloud deconvolution module includes: D01: Let the dimension of the 3D coordinate matrix of the input point cloud in the current stage be... , will The three-dimensional coordinate matrix is ​​input into the first point feature extraction module e31 to obtain... The first feature matrix; where, Determined based on the stage of the point cloud deconvolution module. For stage indexing; The first feature extraction module (e31) is typically a multi-layer one-dimensional convolutional network that performs a non-linear mapping on the three-dimensional coordinate matrix of the input point cloud to extract the initial feature representation of each point. For example, the first-stage point cloud deconvolution module... The initial point cloud's 3D coordinate matrix is ​​nonlinearly mapped to extract the initial feature representation of each point; D02: Will The first feature matrix, its global aggregated features, and the point-by-point copied global point cloud feature vectors are concatenated along the feature dimension to obtain... Second concatenated feature matrix; During the actual assembly, first... The first feature matrix is ​​max-pooled along the point dimension to obtain... Global aggregated features; copying the 1×D global point cloud feature vector along the row direction. Next, get Global feature copy matrix: Concatenate the first-point feature matrix, the global aggregated features, and the global feature copy matrix along the column direction to obtain... The second concatenated feature matrix.

[0058] D03: Will The second concatenated feature matrix is ​​input into the feature fusion module q31 to obtain... Query the feature matrix; the feature fusion module is usually a one-dimensional convolutional layer. D04: Will The feature matrix is ​​input into the feature interaction module i31 for feature interaction, and the result is obtained. Interaction feature matrix; The feature interaction module i31 typically employs a self-attention mechanism or a feature interaction mechanism based on the Mamba architecture to model long-range dependencies between points. The self-attention mechanism calculates the attention weight between the query feature matrix and itself, enabling each point's features to aggregate feature information from other points, thereby enhancing global consistency and structural integrity. In some other preferred embodiments, the feature interaction module i31 adopts a Mamba Gate structure, which combines the advantages of traditional attention mechanisms and Mamba sequence modeling, and can effectively capture long-range dependencies while maintaining computational efficiency, thereby improving the accuracy and stability of point cloud completion. D05-1: Will The interactive feature matrix is ​​input to the point splitting module m31 according to the upsampling factor. Perform point-level splitting to obtain Sub-point feature matrix; upsampling factor The deconvolution module of the point cloud is set from low to high as 1, 1, 2, 4. The point splitting module m31 first performs a one-dimensional convolution on the interaction feature matrix, mapping it to... The intermediate feature matrix is ​​then transposed and convolved to expand the intermediate feature matrix into... Sub-point feature matrix; where the kernel size and stride of the transposed convolution are both set to... Thus achieving the points Double amplification; point splitting operation generates for each input point These child points inherit the feature information of the parent point, but their final spatial location has not yet been determined. D05-2: Will The interactive feature matrix is ​​input into the feature upsampling module g31 to obtain... Upsampled interactive feature matrix; The feature upsampling module g31 uses nearest neighbor interpolation or bilinear interpolation strategies to... The interaction feature matrix is ​​expanded in the point dimension to Upsampled interactive feature matrix; D06: Will Sub-point characteristic matrix and The upsampled interactive feature matrix is ​​concatenated along the feature dimension to obtain Third concatenation feature matrix; During the specific assembly process, Sub-point characteristic matrix and The upsampled interactive feature matrix is ​​concatenated along the column direction to obtain Third concatenation feature matrix; D07: Will The third concatenated feature matrix is ​​input into the displacement feature generation module s31 to obtain... Displacement feature matrix; the displacement feature generation module s31 is usually composed of multiple one-dimensional convolutional layers and residual connections between convolutional layers; the displacement feature matrix encodes the displacement direction and magnitude information of each child point relative to its corresponding parent point, providing a basis for the final displacement prediction; D08: Will The displacement feature matrix is ​​input into the displacement prediction module p31 to obtain... Three-dimensional displacement matrix; The displacement prediction module p31 is usually composed of a one-dimensional convolutional layer and a non-linear activation function, which maps the displacement feature matrix to three-dimensional space and outputs the three-dimensional displacement vector of each sub-point. D09: Will The 3D displacement matrix is ​​added point by point to the 3D coordinate matrix of the upsampled input point cloud to obtain the 3D coordinate matrix of the output point cloud at the current stage.

[0059] In some other preferred embodiments, the displacement prediction module p31 further includes a boundary constraint submodule. The boundary constraint submodule performs bounded processing on the three-dimensional displacement matrix and follows a stage index. Scaling is performed. In practice, the hyperbolic tangent function is used (…). The three-dimensional displacement is constrained within the range of [-1, 1], and then scaled according to the radius parameter r and the stage index i, i. ,in This is the final bounded displacement vector; Output for displacement prediction; The training process for the point cloud completion network is as follows: First, the medical point cloud data is preprocessed to construct a training dataset of the incomplete point cloud to be completed and the corresponding ground truth point cloud. After obtaining the predicted point cloud for each stage, medical geometric perception quantities (such as curvature, normal change, etc.) and missing correlation quantities (such as the distance from the point to the observed part of the point cloud) are calculated for the predicted point cloud and the ground truth point cloud. Point-level weights are generated based on the medical geometric perception quantities and missing correlation quantities. Second, the reconstruction error between the predicted point cloud and the aligned sampled ground truth point cloud at each stage is weighted using the point-level weights to obtain the loss for each stage. Finally, dynamic stage weights are determined based on the loss for each stage, and the weighted summation of the losses for each stage yields the multi-stage main loss.

[0060] In addition, a Partial Matching loss branch is introduced to perform one-way nearest neighbor matching between the observed part of the incomplete point cloud and the highest stage predicted point cloud to obtain the matching error. Finally, the multi-stage main loss and the matching error are summarized to obtain the total loss, and the point cloud completion network is backpropagated and updated based on the total loss. After multiple rounds of iterative training, the network parameters are gradually optimized, enabling it to complete new incomplete point clouds with high accuracy.

[0061] By applying the KDTree nearest neighbor algorithm to the residual point cloud and the ground truth point cloud (the dataset is obtained by combining the ground truth point cloud with the residual point cloud obtained by simulating defects in the ground truth point cloud. During training, the ground truth point cloud only participates in loss calculation and does not participate in the prediction process), the missing local regions in the residual point cloud are obtained, i.e., the observation part.

[0062] The total loss calculation in the above training process is one of the core inventive points of this invention. It aims to improve the reconstruction accuracy of key anatomical regions and achieve adaptive balance in multi-stage training by introducing medical geometric perception point-level weights and dynamic stage weight adjustment mechanisms. The total loss calculation includes point-level weight generation, dynamic stage weight determination, partial matching loss branching, and summarizing to obtain the total loss.

[0063] Figure 6 This is a schematic diagram of the point-level weight generation process. The purpose of generating point-level weights is to assign a weight coefficient to each point in the predicted point cloud, so that key anatomical parts such as high curvature regions and defect boundaries receive greater weight in the total loss calculation, thereby guiding the point cloud completion network to focus on optimizing the reconstruction accuracy of these regions.

[0064] like Figure 6 As shown, the process for generating point-level weights includes: J01: Constructing a local neighborhood of any point in the predicted point cloud based on K-nearest neighbors (KNN), that is, for each point in the predicted point cloud, searching for the K nearest points in the ground truth point cloud to construct a local neighborhood; the selection of K is adjusted according to the point cloud density and the scale of medical anatomical structures, with typical values ​​being K=16 or K=32. J02: Medical geometric perception quantities representing curvature and normal changes based on local neighborhood calculations. Specifically, for each point, firstly, a local plane or local surface is fitted based on its K nearest neighbors to calculate the normal vector of that point. Then, the change in the normal angle between that point and points in its local neighborhood is calculated as an approximate measure of curvature. Regions with greater curvature correspond to regions with drastic changes in normal direction. These regions are often key structural turning points and functional sites in medical anatomy. J03: Calculate the missing correlation quantity based on the distance from points in the predicted point cloud to the residual point cloud (the observed part of the point cloud). The missing correlation quantity is used to measure the distance between points in the predicted point cloud and the residual point cloud (i.e., the observed part), thereby identifying the regions to be filled in. Specifically, for each point in the predicted point cloud, the shortest distance from it to all points in the residual point cloud is calculated; this distance is the missing correlation quantity. Points with larger distances are located within the missing regions generated by the incomplete reconstruction. The reconstruction accuracy of these regions directly affects the completeness of the incomplete result. Points with smaller distances are located near the residual cloud or in overlapping areas. These regions need to be consistent with the observed part to prevent the network from introducing offsets or deformations. (The last part, "missing correlation quantity," appears to be an unrelated fragment and is left untranslated.) This reflects the difficulty and importance of point completion: points inside the missing area are more difficult to reconstruct due to the lack of observation constraints and require higher weights, while points close to the observation part are relatively easier to reconstruct due to the reference of neighboring observation points, but they also need a certain weight to ensure a smooth connection with the observation part. J04: Construct point-level raw weight scores based on medical geometric perception quantities and defect-related quantities, and then perform bounded normalization on the point-level raw weight scores to obtain point-level weights. Specifically, the point-level raw weight scores... It can be represented as: ,in , This is a balancing coefficient used to adjust the contribution ratio of geometrically sensed quantities and defect-related quantities to the weights; then, for Bounded normalization is performed to map the weights to the range [0, 1], resulting in the final point-level weights. Normalization can be achieved using the Sigmoid function: Alternatively, piecewise linear normalization can be used: - ,in and The minimum and maximum values ​​of the original point-level weight scores are defined as the minimum and maximum values ​​in the entire point cloud. Bounded normalization ensures the numerical stability of the point-level weights and avoids the impact of excessively large or small weight values ​​on the training process.

[0065] Through the above process, each point in the predicted point cloud is assigned a point-level weight. The weight reflects the medical geometric importance and completion difficulty of the point. In the subsequent loss function calculation, the reconstruction error of points in high curvature regions and defect boundaries accounts for a larger proportion in the loss function due to their larger point-level weights, thus guiding the network to focus on optimizing the completion accuracy of these key regions.

[0066] Figure 7This is a schematic diagram of the calculation process for multi-stage principal loss. The calculation of multi-stage principal loss involves weighting and summing the reconstruction errors between the predicted point clouds and the ground truth point clouds at each stage, and achieving adaptive balance of losses at each stage through a dynamic stage weight adjustment mechanism.

[0067] like Figure 7 As shown, the calculation process for the multi-stage main loss includes: Z01: Calculate the stage loss for each stage separately: (The...) stage( Predicted point cloud First, examine the complete truth point cloud. Perform alignment sampling to obtain the same as Truth value set with consistent point count Alignment sampling typically employs a farthest point sampling (FPS) strategy, selecting representative points from the ground truth point cloud to ensure that the sampled ground truth point set matches the number of points and spatial distribution of the predicted point cloud as closely as possible. Then, the predicted point cloud is calculated based on point-level weights. With the set of truth points Weighted chamfer distance loss between them; Chamfer distance (CD) is a commonly used error metric in point cloud reconstruction tasks. It is defined as the sum of the bidirectional nearest neighbor distances between two point clouds. Specifically, the first... stage-weighted chamfer distance loss It can be represented as: ; in, Used for traversing and predicting point clouds all points For each point Find the set of truth points The point closest to it Calculation points Time of Norm (i.e., Manhattan distance), multiplied by the point Point-level weights Finally, for all points The weighted average distance is calculated. Used for traversing the set of truth points all points For each point Find the predicted point cloud The point closest to it Calculation points Time of Norm, and multiplied by point Point-level weights Finally, for all points The weighted distance is averaged. For the first The number of points in the predicted point cloud at each stage; Based on the above calculations, the losses in the first stage are obtained respectively. Second phase losses Third stage losses and the fourth phase loss These stage losses measure the reconstruction error between the predicted point cloud and the corresponding ground truth point set at each stage. The introduction of point-level weights amplifies the errors of high curvature regions and missing boundaries in each stage loss. Z02: Construct a stage evaluation metric based on the stage loss: Stage Evaluation Metric Used to evaluate the The quality of reconstruction at each stage serves as the basis for subsequent dynamic stage weight calculations, and is used as a stage evaluation metric. For stage losses The function is a monotonically decreasing function, meaning that the larger the stage loss, the smaller the stage evaluation value. One construction method uses the reciprocal form: ;in, It is a very small integer to avoid division by zero errors and stabilize numerical calculations. In this way, stages with larger stage losses correspond to smaller stage evaluation values, and stages with smaller stage losses correspond to larger stage evaluation values.

[0068] To suppress weight oscillations, the stage evaluation metric can optionally employ a moving average of historical stage losses. Specifically, in each training batch, the stage loss of the current batch is recorded. Then, the smoothed stage loss is updated using the exponential moving average (EMA): ;in, For smoothing coefficients (e.g.) ), The stage evaluation metric is calculated based on the smoothed stage loss from the previous batch: The introduction of moving average can mitigate the drastic fluctuations in stage loss during training, making the stage evaluation quantity and subsequent dynamic stage weights more stable and avoiding training oscillations.

[0069] Z03: Obtaining stage scores by linear transformation of stage evaluation metrics: The purpose of linear transformation is to map stage evaluation metrics to a suitable numerical range, facilitating subsequent Softmax normalization. One method of linear transformation is to multiply by a scaling factor. ,in This is the scaling factor; affine transformation may also be used in some other embodiments: ,in The bias term is the stage fraction after linear transformation. The relative magnitudes of the stage evaluation quantities were maintained, that is, the greater the stage loss, the smaller the stage score, and the smaller the stage loss, the larger the stage score. Z04: Input the stage scores into the Softmax function with temperature parameters to obtain dynamic stage weights: The Softmax function normalizes the stage scores into a probability distribution, making the weights of each stage non-negative and summing to 1. Temperature parameters... Used to adjust the smoothness of the Softmax distribution: when When the value is large, the Softmax distribution tends to be uniform, and the weight differences between stages are small. When the values ​​are small, the Softmax distribution tends to be sharper, with stages having higher stage scores having significantly higher weights and stages having lower stage scores having significantly lower weights. Dynamic stage weights of the stage It can be represented as: ; Among them, in the early stage of training, temperature parameter Set to a larger value (e.g.) This ensures a relatively balanced weight distribution across all stages, allowing the network to simultaneously optimize the output of each stage and establish a progressive completion capability from coarse to fine. As training progresses, the temperature parameter... Gradually reduce according to the predetermined annealing strategy (e.g., reduce at regular intervals). Multiply by 0.9), so that the dynamic stage weights gradually tilt towards higher stages with smaller stage losses, guiding the network to focus on optimizing the output of fine stages; To further suppress the oscillations of dynamic stage weights during training, this invention introduces a weight smoothing constraint in the updating of dynamic stage weights. Specifically, the dynamic stage weights are iteratively updated according to a predetermined update cycle (such as each training batch or every few batches), and the dynamic stage weights of the current cycle are... Dynamic stage weights compared to the previous cycle Fusion is performed using an exponential moving average method: ,in The new weights are calculated based on the loss of the current cycle stage. For smoothing coefficients (e.g.) ); Z05: The stage losses of each stage are weighted and summed according to the dynamic stage weights to obtain the multi-stage main loss. It can be represented as: ; in, , , , The dynamic stage weights for each stage satisfy the following conditions: By weighting and summing the dynamic stage weights, the multi-stage master loss can adaptively balance the training contribution of each stage. In the early stage of training, since the losses of each stage are not much different or the loss of the coarse-grained stage is large, the dynamic stage weights are relatively balanced. The multi-stage master loss focuses on the optimization of each stage at the same time, promoting the network to establish the ability to complete from coarse to fine. As training progresses, the loss of the fine stage gradually decreases, and the dynamic stage weights tilt towards higher stages.

[0070] Figure 8 This is a schematic diagram of the process for calculating the matching error using the Partial Matching loss branch in an embodiment of the present invention. The Partial Matching (PM) loss branch is used to establish local consistency constraints between the observed part of the residual point cloud and the predicted point cloud, ensuring that the generated completed point cloud strictly retains the healthy anatomical structure of the original observation and preventing the network from introducing offsets or deformations to known parts when pursuing global morphological matching.

[0071] like Figure 8 As shown, the process of calculating the matching error in the Partial Matching loss branch includes: P01: Determine the set of observation points from the defect cloud as the query set; the set of observation points contains the skeletal structure of the patient's healthy parts, and these points should be preserved as is during the completion process. The set of observation points is denoted as... Its points are ; It should be noted that the set of observation points may be a residual cloud or a subset of a residual cloud.

[0072] P02: Within the highest-level predicted point cloud, perform a nearest neighbor search for each query point to obtain the corresponding nearest neighbor predicted point. Specifically, traverse the set of observation points. Each point in Predicting point clouds at the highest stage (i.e., the fourth stage predicted point cloud, with 20480 points) search distance points nearest point , recorded as ;in, Traversal For all points in the search, nearest neighbor search typically uses data structures such as KD-trees or Ball trees to accelerate the search and improve search efficiency. P03: Calculate the positive matching error based on the distance metric between the query point and its corresponding nearest neighbor predicted point, and use this positive matching error as the final matching error; positive matching error It can be represented as: ; Among them, summation traverses the set of observation points. all points , For point Predicting point clouds The nearest neighbor in the observation point set is used to measure the shortest distance from each point in the observation point set to the predicted point cloud, reflecting the extent to which the predicted point cloud covers the observed portion.

[0073] In some other preferred embodiments, the calculation of the matching error further includes: P04: Using the predicted point as the query point, perform a nearest neighbor search within the observation point set to obtain the reverse matching error. Specifically, traverse the predicted point cloud. Each point in In the observation point set Search distance point nearest point , recorded as ;in, Traversal All points in the dataset are analyzed, and then the reverse matching error is calculated. : ; Among them, summation traversal predicts point clouds. all points , For point In the observation point set The nearest neighbor in the middle, For predicting point clouds The number of points (i.e., 20480) is used to measure the distance between points in the predicted point cloud and the observation point set, reflecting how many points in the predicted point cloud are located near the observation part. P05: Weight the forward matching error and the backward matching error according to the preset weight coefficient to obtain the matching error. , specifically: ; in, and These are weighting coefficients used to balance the contributions of one-way and reverse matching errors. In a preferred embodiment, =1.0, =0.5, meaning that the weight of the forward matching error is relatively large because preserving the observed portion is the primary objective, while the weight of the backward matching error is relatively small as an auxiliary constraint.

[0074] In some other preferred embodiments, the matching error can also be calculated as follows: Using the predicted point as the query point, a nearest neighbor search is performed within the observation point set to obtain the reverse matching error, which is then used as the final matching error. Specifically, the predicted point cloud is traversed. Each point in In the observation point set Search distance point nearest point , recorded as ;in, Traversal All points in the dataset are analyzed, and then the reverse matching error is calculated. : ; This invention provides three methods for calculating the matching error: The first is to calculate the forward matching error, using it as the final matching error; this method ensures coverage of the observed point cloud. The second is to calculate the reverse matching error, using it as the final matching error; this method can suppress excessive crowding of predicted points in the observation area. The third is to calculate the forward and reverse matching errors, and then weight and fuse them to obtain the final matching error. The method for calculating the matching error can be selected according to specific needs.

[0075] By introducing the Partial Matching loss branch, the point cloud completion network not only optimizes the global matching between the predicted point cloud and the ground truth point cloud during training, but also ensures that the predicted point cloud strictly covers the original observation part to avoid tampering with the healthy structure. This constraint is particularly important in medical point cloud completion tasks, because in clinical applications, no deviation or deformation of the patient's existing healthy anatomical structure is allowed, otherwise it may affect the subsequent implant design and surgical planning.

[0076] The multi-stage main loss is calculated. Matching error with Partial Matching Then, the two are summed according to preset weighting coefficients to obtain the total loss. : ; in, The non-negative PM branch weight coefficient is used to adjust the contribution of the Partial Matching loss to the total loss. In a preferred embodiment, That is, PM losses account for approximately 20% of the total losses, while multi-stage main losses account for approximately 80%. The selection requires a trade-off between global morphological matching and local observation preservation: If the size is too small, the network may not retain enough of the observed data; If the size is too large, the network may focus excessively on matching the observed parts, while ignoring the quality of the completion of the missing areas.

[0077] Based on total loss The backpropagation algorithm is used to calculate the gradients of the parameters of each layer of the point cloud-completed network, and the optimizer updates the network parameters. Specifically, the gradient calculation is performed by backpropagating from the total loss to each layer of the network using the chain rule to obtain the gradient value of each parameter; then, the optimizer updates the parameters based on the gradient values ​​and the learning rate. ; in, For network parameters, For learning rate, The total loss is calculated based on the network parameters. The gradient. During the learning process, the learning rate is usually adjusted using cosine annealing or piecewise decay strategies to ensure training stability and convergence speed.

[0078] Through multiple rounds of iterative training, the network parameters were gradually optimized, resulting in a reduction in both the multi-stage principal loss and the PM loss. After training, the point cloud completion network was able to perform high-precision completion of new incomplete point clouds during the testing phase. The generated completed point cloud not only closely matches the ground truth point cloud in terms of global morphology, but also maintains fine geometric details in key areas such as high curvature regions and defect boundaries. At the same time, it strictly preserves the healthy anatomical structure of the original observed portion, meeting the high standards required for medical clinical applications.

[0079] A second aspect of the present invention provides a point cloud completion system for the medical field, such as... Figure 9 As shown, this system is used to execute the aforementioned point cloud completion method for the medical field, including: Data acquisition and processing module: used to acquire raw image data through the image acquisition interface, preprocess the raw image data, and obtain the incomplete cloud to be filled; Multi-stage point cloud completion module: Used to perform multi-stage point cloud prediction on the incomplete point cloud to be completed through the point cloud completion network to obtain multi-stage predicted point cloud; extract the highest stage predicted point cloud in the multi-stage predicted point cloud as the point cloud completion result. The medical geometry-defect coupling weight generation module is used to construct a local neighborhood of any point in the predicted point cloud based on K-nearest neighbors; calculate medical geometry perception quantities representing curvature and normal changes based on the local neighborhood; calculate defect-related quantities based on the distance from the point to the defect point cloud; and, Based on the medical geometric perception quantity and the missing correlation quantity, a point-level original weight score is constructed, and the point-level original weight score is bounded normalized to obtain the point-level weight. Multi-stage loss adaptive fusion evaluation module: used to weight the reconstruction error between the predicted point cloud and the ground truth point cloud at each stage using point-level weights to obtain the loss at each stage; and, Dynamic stage weights are determined based on stage losses, and the weighted summation of losses from each stage is used to obtain the multi-stage master loss. Locally observable consistency matching loss module: This module introduces the Partial Matching loss branch, performs one-way nearest neighbor matching between the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, and obtains the matching error. Training optimization and parameter update module: used to optimize network parameters of the point cloud completion network based on the total loss; where the total loss is obtained by summing the multi-stage main loss and matching error.

[0080] The data acquisition and processing module includes an image acquisition interface, a coordinate normalization component, and a sampling and shaping component. The image acquisition interface is used to interact with medical imaging equipment (such as CT scanners, MRI scanners, etc.) to acquire the patient's three-dimensional medical image data; the coordinate normalization component is used to convert the coordinates of the three-dimensional medical image data to a uniform scale, such as moving the center of the point cloud to the origin and normalizing the bounding box of the point cloud to within a unit sphere; the sampling and shaping component is used to perform operations such as denoising, downsampling, scale normalization, and coordinate system alignment on the three-dimensional medical image data to meet the format and size requirements of network input.

[0081] The multi-stage point cloud completion module comprises a feature encoder, a stage-by-stage decoder, and a stage output convergence component. The feature encoder employs a dynamic graph convolution-based encoding structure to extract global feature vectors from the residual point cloud. The stage-by-stage decoder includes a seed point generator and a multi-stage point cloud decoder, which consists of four cascaded point cloud deconvolution modules that upsample and refine the initial point cloud stage by stage to generate multi-stage predicted point clouds. The stage output convergence component collects the predicted point clouds from each stage and extracts the highest-stage predicted point cloud as the final completed point cloud result. The multi-stage point cloud completion module adopts a staged deep neural network architecture, where each stage decoder includes a self-attention mechanism or a feature interaction mechanism based on the Mamba architecture.

[0082] The medical geometry-defect coupling weight generation module includes a neighborhood construction component, a geometry extraction component, and a weight normalization component. The neighborhood construction component is used to construct a K-nearest neighbor local neighborhood for each point in the predicted point cloud. The geometry extraction component is used to calculate medical geometric perception quantities representing curvature and normal changes based on the local neighborhood, and to calculate defect-related quantities based on the distance from the point to the defect point cloud (observation point cloud). The weight normalization component is used to construct point-level original weight scores based on medical geometric perception quantities and defect-related quantities, and to perform bounded normalization to obtain point-level weights.

[0083] The multi-stage loss adaptive fusion evaluation module includes an alignment sampling component, a weighted distance metric component, and a stage weight calculation component. The alignment sampling component is used to sample the farthest points from the complete ground truth point cloud to obtain a ground truth point set consistent with the number of points in the predicted point cloud at each stage. The weighted distance metric component is used to calculate the weighted chamfer distance loss between the predicted point cloud and the ground truth point set based on point-level weights to obtain the loss for each stage. The stage weight calculation component is used to construct a stage evaluation metric based on the loss for each stage, obtain a stage score through linear transformation, and input the stage score into a Softmax function with a temperature parameter to obtain the dynamic stage weight.

[0084] The Locally Observable Consistency Matching Loss Module is used to introduce the Partial Matching loss branch, which performs one-way nearest neighbor matching on the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud to obtain the matching error.

[0085] The locally observable consistency matching loss module includes a nearest neighbor search component and a matching error aggregation component. The nearest neighbor search component performs a nearest neighbor search for each point in the observation point set in the predicted point cloud to obtain the corresponding nearest neighbor predicted point, and optionally performs a nearest neighbor search for each point in the observation point set for the predicted point cloud to obtain the reverse nearest neighbor point. The matching error aggregation component calculates the forward matching error and the optional reverse matching error based on the distance metric between the query point and the corresponding nearest neighbor point, and weights the two according to a preset weight coefficient to obtain the partial matching error. The locally observable consistency matching loss module includes at least one nearest neighbor search algorithm, and calculates the matching error based on the matching error between the observation point and the predicted point cloud and participates in the optimization of the total loss.

[0086] Training optimization and parameter update module: used to optimize network parameters for the point cloud completion network based on the total loss.

[0087] The training optimization and parameter update module includes a loss aggregation component, a gradient calculation component, and a parameter update component. The loss aggregation component is used to weight and aggregate the multi-stage main loss and the partial matching error according to a preset weight coefficient to obtain the total loss. The gradient calculation component is used to calculate the gradient of the parameters of each layer of the network through the backpropagation algorithm based on the total loss. The parameter update component is used to update the network parameters using the optimizer based on the gradient value and the learning rate.

[0088] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments.

[0089] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A point cloud completion method for the medical field, characterized in that, include: Obtain the residual defect cloud after data preprocessing; wherein, the residual defect cloud is a three-dimensional coordinate matrix; The residual point cloud is input into a pre-trained point cloud completion network to obtain a multi-stage predicted point cloud. The point cloud completion network includes a feature encoder, a seed point generator, and a multi-stage point cloud decoder. The feature encoder extracts the global point cloud feature vector of the residual point cloud. The seed point generator generates an initial point cloud based on the global point cloud feature vector. The multi-stage point cloud decoder upsamples the initial point cloud stage by stage under the guidance of the global point cloud feature vector to obtain a multi-stage predicted point cloud, where each stage's predicted point cloud is a three-dimensional coordinate matrix. Extract the highest-stage predicted point cloud from the multi-stage predicted point cloud as the point cloud completion result; The training process of the point cloud completion network includes: Medical point cloud data is preprocessed to construct a training dataset of the incomplete point cloud to be completed and the ground truth point cloud corresponding to the incomplete point cloud. The residual point clouds in the training dataset are input into the point cloud completion network to obtain the predicted point clouds in four stages; wherein, the predicted point cloud is a three-dimensional coordinate matrix; Based on the predicted point cloud and the ground truth point cloud, the medical geometric perception quantity and the missing correlation quantity are calculated, and point-level weights are generated based on the medical geometric perception quantity and the missing correlation quantity. The reconstruction error between the predicted point cloud and the ground truth point cloud at each stage is weighted using the point-level weights to obtain the loss at each stage; dynamic stage weights are determined based on the stage loss to sum the weighted losses at each stage to obtain the multi-stage main loss; wherein, the dynamic stage weights at each stage are non-negative and sum to 1. A Partial Matching loss branch is introduced to perform one-way nearest neighbor matching between the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, and the matching error is obtained. The multi-stage main loss and the matching error are summed to obtain the total loss. The point cloud completion network is then trained and updated using backpropagation based on the total loss to complete the training of the point cloud completion network. The calculation of losses at each stage includes: Alignment sampling is performed on the ground truth point cloud based on the predicted point cloud at each stage to obtain the ground truth point set; wherein, the alignment sampling adopts the farthest point sampling strategy, and the number of points in the predicted point cloud and the ground truth point set is the same; The chamfer distance loss between the predicted point cloud and the ground truth point set is calculated based on the point-level weights and used as the corresponding stage loss; wherein, the distance metric in the chamfer distance loss adopts the L1 norm; The generation of the point-level weights includes: A local neighborhood of any point in the predicted point cloud is constructed based on K-nearest neighbors; where K=16 or K=32. Based on the local neighborhood, medical geometric perception quantities representing curvature and normal changes are calculated, and missing correlation quantities are calculated based on the distance from the point to the residual defect cloud. Based on the medical geometric perception quantity and the missing correlation quantity, a point-level original weight score is constructed, and the point-level original weight score is bounded normalized to obtain the point-level weight. The dynamic stage weights are determined based on the stage loss, including: A stage evaluation quantity is constructed based on the stage loss; wherein, the stage evaluation quantity is a monotonically decreasing function of the stage loss; The stage evaluation quantity is linearly transformed to obtain the stage score; the stage score is then substituted into a Softmax function with a temperature parameter to obtain the dynamic stage weight. The dynamic stage weights are iteratively updated according to a preset update cycle during the training of the point cloud completion network, and adaptively change with the training process based on weight smoothing constraints and temperature annealing strategies. The weight smoothing constraints are used to limit the change range of dynamic stage weights in adjacent update cycles, and the temperature annealing strategy is used to gradually reduce the temperature parameter with each training round to improve the distinguishability of weight allocation in each stage.

2. The point cloud completion method for the medical field according to claim 1, characterized in that, The feature encoder includes a sampling module, an input feature transformation module, a first graph-edge convolution module, a first farthest point downsampling module, a second graph-edge convolution module, a second farthest point downsampling module, a third graph-edge convolution module, a third farthest point downsampling module, a multi-scale feature splicing and fusion module, an output feature transformation module, and a global pooling module. The data processing flow of the feature encoder includes: The three-dimensional coordinate matrix of the residual point cloud is input to the sampling module, and the three-dimensional coordinate matrix of the sampled point cloud is obtained by sampling using the farthest point sampling strategy; wherein, the dimension of the three-dimensional coordinate matrix of the residual point cloud is... , The number of points in the residual point cloud is given, and the dimension of the three-dimensional coordinate matrix of the sampled point cloud is 5120×3. The three-dimensional coordinate matrix of the sampled point cloud is input to the input feature transformation module to obtain the initial feature matrix; wherein the dimension of the initial feature matrix is ​​5120×8; A local neighborhood of a point is constructed based on K-nearest neighbors; the local neighborhood and the initial feature matrix are input into the first graph edge convolution module to extract edge convolution features, resulting in a first-scale feature matrix; wherein the dimension of the first-scale feature matrix is ​​5120×64, and K in K-nearest neighbors is 16; The first scale feature matrix is ​​input into the first farthest point downsampling module for downsampling to obtain the first downsampled feature matrix, wherein the dimension of the first downsampled feature matrix is ​​512×64; The first downsampled feature matrix is ​​input into the second graph edge convolution module to obtain the second scale feature matrix; the second scale feature matrix is ​​downsampled by the second farthest point downsampled module to obtain the second downsampled feature matrix; wherein, the dimension of the second scale feature matrix is ​​512×128 and the dimension of the second downsampled feature matrix is ​​256×128. The second downsampled feature matrix is ​​input into the third graph edge convolution module to obtain the third scale feature matrix; the third scale feature matrix is ​​downsampled by the third farthest point downsampled module to obtain the high-level feature matrix; wherein, the dimension of the third scale feature matrix is ​​256×256, and the dimension of the high-level feature matrix is ​​128×512. The first-scale feature matrix, the second-scale feature matrix, the third-scale feature matrix, and the high-level feature matrix are input into the multi-scale feature splicing and fusion module for splicing and fusion to obtain the fused feature matrix; wherein, the dimension of the fused feature matrix is ​​128×960; The fused feature matrix is ​​input to the output feature transformation module to obtain the point feature matrix; wherein the dimension of the point feature matrix is ​​128×D, and D is the preset feature dimension; The point feature matrix is ​​input into the global pooling module for max pooling to obtain the global point cloud feature vector; wherein the dimension of the global point cloud feature vector is 1×D.

3. The point cloud completion method for the medical field according to claim 1, characterized in that, The seed point generator includes a point feature expansion module, a feature splicing module, a residual mapping module, and a coordinate generation module; The data processing flow of the seed point generator includes: The global feature vector is input to the point feature expansion module to obtain the point feature matrix; wherein the dimension of the point feature matrix is... , The number of seed points; The point feature matrix and the globally copied point cloud feature vectors are input to the feature stitching module, and stitched together along the feature dimension to obtain a first stitched feature matrix. The first stitched feature matrix is ​​then nonlinearly mapped by the residual mapping module to obtain the enhanced point feature matrix. The dimension of the first stitched feature matrix is... The dimension of the enhanced point feature matrix is ; The enhanced point feature matrix is ​​input into the coordinate generation module to obtain the three-dimensional coordinate matrix of the seed point cloud; wherein, the dimension of the three-dimensional coordinate matrix of the seed point cloud is... ; The 3D coordinate matrix of the seed point cloud and the 3D coordinate matrix of the residual point cloud are merged to obtain the 3D coordinate matrix of the merged point cloud; the farthest point is sampled from the 3D coordinate matrix of the merged point cloud to obtain the 3D coordinate matrix of the initial point cloud; wherein, the dimension of the 3D coordinate matrix of the merged point cloud is... The dimension of the initial point cloud's three-dimensional coordinate matrix is , To select the number of representative points from the merged point cloud.

4. The point cloud completion method for the medical field according to claim 1, characterized in that, The multi-stage point cloud decoder includes a first-stage point cloud deconvolution module, a second-stage point cloud deconvolution module, a third-stage point cloud deconvolution module, and a fourth-stage point cloud deconvolution module, and the module structure of the point cloud deconvolution module in each stage is the same. The data processing flow of the multi-stage point cloud decoder includes: The three-dimensional coordinate matrix of the initial point cloud is input into the first-stage point cloud deconvolution module to obtain the three-dimensional coordinate matrix of the first-stage predicted point cloud. The three-dimensional coordinate matrix of the predicted point cloud in the first stage is input into the deconvolution module of the point cloud in the second stage to obtain the three-dimensional coordinate matrix of the predicted point cloud in the second stage. The three-dimensional coordinate matrix of the predicted point cloud in the second stage is input into the deconvolution module of the point cloud in the third stage to obtain the three-dimensional coordinate matrix of the predicted point cloud in the third stage. The three-dimensional coordinate matrix of the predicted point cloud in the third stage is input into the deconvolution module of the third stage point cloud to obtain the three-dimensional coordinate matrix of the predicted point cloud in the fourth stage; wherein, the predicted point cloud in the fourth stage is used as the predicted point cloud in the highest stage.

5. A point cloud completion method for the medical field according to claim 4, characterized in that, The point cloud deconvolution module includes a first point feature extraction module, a feature fusion module, a feature interaction module, a point splitting module, a feature upsampling module, a displacement feature generation module, and a displacement prediction module. The data processing flow of the point cloud deconvolution module includes: Let the dimension of the three-dimensional coordinate matrix of the input point cloud at the current stage be... The three-dimensional coordinate matrix is ​​input into the first point feature extraction module to obtain the first point feature matrix; wherein, the dimension of the first point feature matrix is... , Determined based on the stage of the point cloud deconvolution module. For stage indexing; The first point feature matrix, the global aggregated features, and the point-replicated global point cloud feature vectors are concatenated along the feature dimension to obtain the second concatenated feature matrix; wherein, the global aggregated features are obtained by max pooling the first feature matrix along the point dimension, and the dimension of the second concatenated matrix is... ; The second concatenated matrix is ​​input into the feature fusion module to obtain the query feature matrix; wherein the dimension of the query feature matrix is... ; The query feature matrix is ​​input to the feature interaction module for feature interaction to obtain the interaction feature matrix; wherein, the dimension of the interaction feature matrix is... The feature interaction module adopts a self-attention mechanism or a feature interaction mechanism based on the Mamba architecture; The interaction feature matrix is ​​input into the point splitting module, and then analyzed according to the upsampling factor. Perform point-level splitting to obtain sub-point feature matrices; where the dimension of the sub-point feature matrices is... , The upsampling factor. The deconvolution module of the point cloud is set from low to high as 1, 1, 2, 4. The interaction feature matrix is ​​input into the feature upsampling module, and an upsampling interpolation strategy is used to obtain the upsampled interaction feature matrix; wherein, the dimension of the upsampled interaction feature matrix is... Upsampling interpolation strategies include nearest neighbor interpolation or bilinear interpolation; The sub-point feature matrix and the upsampled interactive feature matrix are concatenated along the feature dimension to obtain a third concatenated feature matrix; wherein the dimension of the third concatenated feature matrix is... ; The third concatenated feature matrix is ​​input into the displacement feature generation module to obtain the displacement feature matrix; wherein the dimension of the displacement feature matrix is... ; The displacement feature matrix is ​​input into the displacement prediction module to obtain a three-dimensional displacement matrix; wherein the dimension of the three-dimensional displacement matrix is... ; The three-dimensional displacement matrix is ​​added point by point to the three-dimensional coordinate matrix of the upsampled input point cloud to obtain the three-dimensional coordinate matrix of the output point cloud at the current stage.

6. The point cloud completion method for the medical field according to claim 1, characterized in that, Perform one-way nearest neighbor matching on the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, including: The set of observation points is determined from the residual point cloud as the query point set; in the highest stage prediction point cloud in the prediction point cloud, a nearest neighbor search is performed for each query point in the query point set to obtain the corresponding nearest neighbor prediction point; the positive matching error is calculated based on the distance metric between the query point and the corresponding nearest neighbor prediction point; wherein, the distance metric is the L1 norm. Using the predicted points in the predicted point cloud as query points, a nearest neighbor search is performed for each query point in the observation point set to obtain the corresponding nearest neighbor predicted points; the reverse matching error is calculated based on the distance metric between the query point and the corresponding nearest neighbor predicted point. The matching error is obtained by weighting the forward matching error and the reverse matching error according to a preset weighting coefficient; wherein the preset weighting coefficient is used to balance the contribution of the forward matching error and the reverse matching error.

7. A point cloud completion system for the medical field, used to execute the point cloud completion method for the medical field as described in claim 1, characterized in that, include: Data acquisition and processing module: used to acquire raw image data through the image acquisition interface, preprocess the raw image data, and obtain the incomplete cloud to be filled; Multi-stage point cloud completion module: Used to perform multi-stage point cloud prediction on the incomplete point cloud to be completed through the point cloud completion network to obtain multi-stage predicted point cloud; extract the highest stage predicted point cloud in the multi-stage predicted point cloud as the point cloud completion result. Medical geometry-defect coupling weight generation module: used to construct a local neighborhood of any point in the predicted point cloud based on K-nearest neighbors; calculate medical geometry perception quantities representing curvature and normal changes based on the local neighborhood; and calculate missing correlation quantities based on the distance from the point to the defective point cloud. as well as, Based on the medical geometric perception quantity and the missing correlation quantity, a point-level original weight score is constructed, and the point-level original weight score is bounded normalized to obtain the point-level weight. Multi-stage loss adaptive fusion evaluation module: used to weight the reconstruction error between the predicted point cloud and the ground truth point cloud at each stage using point-level weights to obtain the loss at each stage; as well as, Dynamic stage weights are determined based on stage losses, and the weighted summation of losses from each stage is used to obtain the multi-stage master loss. Locally observable consistency matching loss module: This module introduces the Partial Matching loss branch, performs one-way nearest neighbor matching between the residual point cloud and the highest-stage predicted point cloud in the predicted point cloud, and obtains the matching error. Training optimization and parameter update module: used to optimize network parameters of the point cloud completion network based on the total loss; where the total loss is obtained by summing the multi-stage main loss and matching error.

Images