Construction method of tumor three-dimensional motion trajectory prediction model and prediction system

Abstract

The application belongs to the technical field of medical signal processing, and particularly relates to a construction method and a prediction system of a tumor three-dimensional motion trajectory prediction model. The body surface point cloud time series data and tumor time series position information of the human chest and abdomen are collected, and a unified space coordinate system is established. The body surface point cloud time series data is divided into n time periods with a total length of T1, and the tumor time series position information is divided into n time periods with a total length of T2, thereby obtaining n training samples. The body surface point cloud of the training sample is divided into a plurality of pseudo-voxel sub-blocks according to the human projection plane, and K pseudo-voxel sub-blocks are screened out according to the motion saliency, thereby obtaining a body surface motion time series feature vector. The body surface motion time series feature vector of the tumor three-dimensional motion trajectory prediction is taken as the input, the corresponding tumor time series position information of the tumor three-dimensional motion trajectory prediction is taken as the output, and the attention-enhanced dilated convolution prediction network is trained. The application can realize high-precision and non-invasive tumor three-dimensional motion trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing technology, specifically relating to a method for constructing a three-dimensional motion trajectory prediction model for tumors and a prediction system. Background Technology

[0002] During radiotherapy for thoracic and abdominal tumors, the patient's respiratory movements cause periodic or quasi-periodic displacement of the tumor target area, resulting in a deviation between the planned irradiation area and the actual tumor location. This affects dose coverage accuracy, increases the risk of radiation exposure to normal tissues, and reduces treatment efficacy. Especially for tumors such as lung cancer and liver cancer that move significantly with respiration, real-time acquisition or prediction of the tumor's three-dimensional trajectory is a crucial technical challenge for improving the precision of radiotherapy.

[0003] Existing tumor motion management techniques generally include direct tracking methods and indirect tracking methods. Direct tracking methods typically rely on X-ray fluoroscopy, cone-beam computed tomography (CBCT), or implanted gold nanoparticles to determine tumor location. While these methods can directly reflect the location of the target area in the body, they generally have problems such as additional radiation exposure, the need for invasive procedures, poor patient compliance, and complex clinical procedures.

[0004] Indirect tracking methods typically use surface motion signals to replace tumor movement information within the body, such as surface-guided radiotherapy. These methods offer advantages such as being non-invasive and requiring no additional radiation. However, due to the complex nonlinear mapping between surface respiratory motion and tumor displacement, and the inherent delays in clinical systems (image acquisition, data processing, transmission, and control execution), existing methods are prone to decreased prediction accuracy under highly dynamic and non-stationary respiratory conditions. Especially in multi-stage workflows, the total system delay further accumulates prediction errors, making it difficult to consistently meet clinical requirements for sub-centimeter or even 2mm accuracy.

[0005] Furthermore, existing prediction models based on recurrent neural networks, long short-term memory networks, or general temporal convolutional networks often focus on modeling one-dimensional or limited surface signals, failing to adequately utilize the spatial information of dense surface point clouds acquired by structured light. Alternatively, they may use single-frame information as input, failing to adequately consider the deformation differences, phase differences, and local correlations of different surface sub-regions during respiration. This makes it difficult for the models to simultaneously achieve prediction accuracy, robustness, and generalization ability in non-stationary breathing patterns, cross-individual scenarios, or when system delays exist.

[0006] Therefore, there is an urgent need to provide a non-invasive, radiation-free method for predicting the three-dimensional motion trajectory of tumors that can fully utilize the local spatial structure information of the body surface point cloud and is suitable for handling long-range dependence of respiratory time and the effects of system delay, in order to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for constructing a three-dimensional tumor motion trajectory prediction model, and to construct a mapping relationship model between the temporal features of the body surface point cloud and the temporal position of the tumor, so as to achieve high-precision real-time prediction of moving tumors in the chest and abdomen.

[0008] To achieve the above objectives, the first aspect of the present invention provides a method for constructing a three-dimensional motion trajectory prediction model for tumors, comprising the following steps: S1. Collect time-series point cloud data of the human chest and abdomen and the corresponding time-series tumor location information, and establish a unified spatial coordinate system; S2. Divide the body surface point cloud time series data into n time periods with a total time length of T1, and divide the tumor time series location information into n time periods with a total time length of T2 to obtain n training samples; wherein the time corresponding to the tumor time series location information of each training sample is the future time of the time corresponding to the body surface point cloud time series data. S3. Divide the surface point cloud of the training samples into several pseudo voxel sub-blocks according to the human projection plane, and select K pseudo voxel sub-blocks according to motion saliency to obtain the temporal feature vector of surface motion. S4. The attention-enhanced dilated convolutional prediction network is trained by taking the temporal feature vector of body surface motion as input and the corresponding temporal location information of the tumor as output.

[0009] Furthermore, in step S3, pseudo-sub-blocks are screened based on the standard deviation, variance, displacement amplitude, periodic energy, or a weighted combination of multiple indicators of the geometric center coordinates of the pseudo-sub-block.

[0010] Furthermore, the screening of pseudo-sub-blocks based on the aforementioned standard deviation includes the following steps: S31. Calculate the geometric center coordinates of the point cloud at time t within each pseudobody sub-block. and the average geometric center coordinates of the total time T1 ; S32. Calculate the standard deviation of the geometric center of each pseudo-sub-block in the time dimension. This is used to quantify the motion salience of the region where the pseudo-sub-block is located: Where i is the sampling point index, The sampling interval; S33, Standard deviation of several pseudo-sub-blocks Sort from largest to smallest, and select the highest significance level of the motion. KEach pseudo-voxel is assigned a sub-block, and its position in the temporal feature vector of body surface motion is determined according to the order of the standard deviations. The geometric center coordinates of the first K pseudo-voxel blocks in each frame of the total time T1 are concatenated, and the standard deviation and / or ranking of each pseudo-voxel block are used as saliency representation information and associated with the corresponding center coordinates to obtain the temporal feature vector of body surface motion. ,in, The temporal feature vector of body surface motion simultaneously carries the spatial motion information of each pseudo-body element and its motion saliency information.

[0011] Furthermore, the attention-enhanced dilated convolutional prediction network includes multiple cascaded attention-enhanced dilated convolutional modules and an output layer; each attention-enhanced dilated convolutional module includes a causal convolutional unit, a dilated convolutional unit, a self-attention mechanism unit, and a residual connection unit; The causal convolution unit is used to perform causal padding and one-dimensional convolution on the input body surface motion temporal feature vector, and output a causally constrained body surface motion temporal feature vector. The dilated convolutional unit is used to perform causal dilation padding and interval sampling convolution on the causal-constrained temporal feature vector of body surface motion to obtain the dilated convolutional features of each layer. The self-attention mechanism unit is used to calculate the attention weights between time steps based on the dilated convolution features of the current layer through the self-attention mechanism, and dynamically enhance the dilated convolution features of the current layer according to the attention weights to obtain the enhanced temporal features of the current layer; the enhanced temporal features are fused with the input residuals of the current layer dilated convolution unit through the residual connection unit and output to the next attention-enhanced dilated convolution module, thereby realizing the dynamic enhancement of the temporal features layer by layer, and finally outputting the attention-enhanced features; The output layer is used to output tumor temporal location information at one or more future time points.

[0012] Furthermore, each of the attention-enhanced dilated convolutional modules is stacked and includes a dilated convolutional unit, a self-attention dynamic enhancement unit, and a residual connection unit; the dilation rate of adjacent layers increases progressively; wherein, each dilated convolutional residual block dynamically enhances the temporal features of the current layer within the block, and the enhanced temporal features are propagated to deeper layers of the network through residual propagation, so as to strengthen the representation ability of key time steps, phase transition segments, and long-term dependent information of respiratory sequences; The training loss function can be either the mean squared error loss function or a combined loss function that combines position error and trajectory smoothing constraints.

[0013] Furthermore, the tumor temporal location information includes the three-dimensional coordinates of the tumor centroid at each time step and / or the three-dimensional coordinates of a representative location determined based on the contour segmentation results. Furthermore, the end time of T1 is adjacent to the start time of T2.

[0014] The second aspect of the present invention provides a method for predicting the three-dimensional motion trajectory of a tumor, comprising: real-time acquisition of point cloud time-series data of the human body surface of the chest and abdomen, and constructing a real-time body surface motion time-series feature vector based on the unified spatial coordinate system established in the method for constructing the tumor three-dimensional motion trajectory prediction model, and K pseudo voxel sub-blocks selected. The real-time temporal feature vector of body surface motion is input into the attention-enhanced dilated convolutional prediction network trained by the tumor three-dimensional motion trajectory prediction model construction method, and the three-dimensional motion trajectory of the tumor at future time is output.

[0015] A third aspect of the present invention provides a tumor three-dimensional motion trajectory prediction system, comprising: The model building module is used to build a tumor three-dimensional motion trajectory prediction model using the aforementioned tumor three-dimensional motion trajectory prediction model construction method; The data acquisition module is used to collect real-time point cloud data of the human chest and abdomen. The data processing module is used to construct a real-time temporal feature vector of body surface motion based on the unified spatial coordinate system established in the construction method of the tumor three-dimensional motion trajectory prediction model and the selected K pseudo voxel sub-blocks. The trajectory prediction module is used to input the real-time temporal feature vector of body surface motion into the trained attention-enhanced dilated convolutional prediction network and output the three-dimensional motion trajectory of the tumor at future moments.

[0016] A fourth aspect of the present invention provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the tumor three-dimensional motion trajectory prediction method.

[0017] A fifth aspect of the present invention provides a radiotherapy device, including a computer-readable storage medium and a processor; The computer-readable storage medium is used to store instructions; The processor is used to read instructions stored in the computer-readable storage medium and execute the tumor three-dimensional motion trajectory prediction method.

[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. The tumor three-dimensional motion trajectory prediction model provided by the present invention constructs a unified coordinate relationship between the surface point cloud and the tumor displacement in the body, uses a pseudo-voxelization strategy to extract local surface features with significant motion, and combines attention-enhanced dilated convolutional networks to establish a long-term dependency prediction relationship between the current surface motion and the future three-dimensional displacement of the tumor. It has better adaptability to complex and non-stationary breathing patterns, thereby achieving high-precision and non-invasive tumor three-dimensional motion trajectory prediction.

[0019] 2. This invention uses the standard deviation of the time dimension as the motion saliency index, which can better preserve local spatial construction information and effectively reflect the respiratory phase difference between different sub-regions. Therefore, the selected pseudo-voxel sub-blocks are more suitable for describing complex non-rigid respiratory deformation of the chest and abdomen. At the same time, this invention uses the standard deviation value and / or ranking position corresponding to each key pseudo-voxel as a saliency code and inputs it into the network along with the geometric center coordinates. This allows the feature vector to reflect not only the spatial motion state of each local region, but also the relative magnitude of the motion saliency of each local region. This solves the problem that existing deep learning models do not make sufficient use of the local spatial structure of the body surface point cloud, have limited ability to model long-term dependencies in complex and non-stationary breathing patterns, and are difficult to balance accuracy and generalization.

[0020] 3. Strong generalization ability: This invention employs a network structure combining causal convolution, dilated convolution, and self-attention mechanisms. This structure not only meets the causal requirements in online prediction scenarios but also effectively captures long-term dependency information in respiratory sequences and adaptively focuses on key phase transition moments, thereby improving the prediction accuracy of tumor motion under non-stationary respiratory patterns. Experiments demonstrate that the network performs stably in tests with unseen respiratory patterns (independent samples), indicating that the model does not overly rely on specific respiratory frequencies or amplitudes.

[0021] 4. High Precision: This invention achieved a low root mean square error (RMSE) in respiratory phantom experiments and maintained stable performance under different respiratory modes, demonstrating that it can ensure high precision while also considering model robustness and generalization performance. In respiratory phantom experiments, this method achieved a root mean square error (RMSE) of 1.272 ± 0.168 mm, consistently remaining within the clinically tolerable range of 2 mm. Compared to the TCN and GRU models, the RMSE was reduced by 13.2% and 15.3%, respectively.

[0022] 5. It can make full use of the local spatial structure information of the body surface: Through the pseudo-voxelization strategy, the present invention transforms the dense, disordered and uneven density of the body surface point cloud into temporal features with clear local regional semantics. This not only reduces the complexity of direct modeling of the original point cloud, but also preserves the local motion differences of different body surface sub-regions, which is conducive to more accurately representing the non-rigid body surface deformation caused by respiration.

[0023] 6. Non-invasive and radiation-free: This invention uses a structured light camera to acquire point clouds of the body surface, eliminating the need for implanting gold markers or continuous use of X-ray fluoroscopy equipment, thus reducing the risk of trauma and additional radiation exposure for patients and improving clinical safety and patient compliance.

[0024] 7. Strong Robustness (Delay Resistance): This invention addresses the unavoidable system delay problem in clinical surface-guided radiotherapy (SGRT) procedures. By constructing a future-time-oriented position prediction model, the system can still output relatively accurate tumor location estimates even with delays, thereby improving the real-time performance and stability of the treatment control chain. When dealing with a system delay of 700ms, the model performance degradation is minimal (RMSE increases by only about 0.012mm), demonstrating its suitability for high-latency clinical SGRT environments. Attached Figure Description

[0025] Figure 1 This is a diagram illustrating the coordinate system relationship (showing the transformation relationship between the CT and camera coordinate systems).

[0026] Figure 2 A flowchart of surface point cloud data acquisition and preprocessing (showing the process of generating pseudo-voxelated feature vectors from raw point clouds).

[0027] Figure 3 This is a diagram of a neural network architecture.

[0028] Figure 4 Structure diagram of the dilated convolutional module for attention enhancement.

[0029] Figure 5 This is a comparison diagram of the actual trajectory and the predicted trajectory under deep breathing state (left side) and shallow breathing state (right side) in an embodiment of the present invention.

[0030] Figure 6 This is a comparison chart of the predicted trajectory effects of the embodiments of the present invention and the prior art. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] Example 1 A method for constructing a three-dimensional motion trajectory prediction model for tumors, characterized by comprising the following steps: S1. Collect time-series point cloud data of the human chest and abdomen and the corresponding time-series tumor location information, and establish a unified spatial coordinate system; S2. Divide the body surface point cloud time series data into n time periods with a total time length of T1, and divide the tumor time series location information into n time periods with a total time length of T2 to obtain n training samples; wherein the time corresponding to the tumor time series location information of each training sample is the future time of the time corresponding to the body surface point cloud time series data. S3. Divide the surface point cloud of the training samples into several pseudo voxel sub-blocks according to the human projection plane, and select K pseudo voxel sub-blocks according to motion saliency to obtain the temporal feature vector of surface motion. S4. The attention-enhanced dilated convolutional prediction network is trained by taking the temporal feature vector of body surface motion as input and the corresponding temporal location information of the tumor as output.

[0033] Specifically, in step S1, a structured light camera is used to acquire continuous point cloud data of the patient's chest and abdominal surface during respiration, obtaining a time-varying surface point cloud sequence; simultaneously, 4D-CT is used to acquire the three-dimensional position information of the tumor target area at the corresponding time phase. The three-dimensional position of the tumor can be the coordinates of the tumor centroid, the coordinates of the tumor isocenter, and the coordinates of a representative position determined based on the contour segmentation results.

[0034] To ensure that surface point cloud data and tumor displacement data are under a unified spatial reference, a rigid body registration method based on surface markers is employed to transform the structured light camera coordinate system to the 4D-CT coordinate system. Optionally, a correspondence is established using at least three non-collinear markers, and the rotation matrix and translation vector are solved to complete the transformation between the two coordinate systems. For any given moment, the surface point cloud coordinates can be mapped to the CT coordinate system according to the rigid body transformation relationship. Figure 1 The specific steps are as follows: S11. Coordinate Transformation: Transform the camera coordinate system Transform to 4D-CT coordinate system This ensures that surface data and in vivo data are within the same spatial reference.

[0035] S12. Camera to CT coordinate system transformation: Rigid body transformation is performed using surface markers.

[0036] The final constructed 3D coordinate system has its X and Y axes corresponding to the length and width of the human body, and the XY plane is the projection plane of the human body. The Z axis corresponds to the thickness of the human body, i.e., the front-back direction.

[0037] In step S2, each time step corresponds to one frame of surface point cloud data. For example, the total time T1 for each sample is 1-20 seconds, with 1-60 frames per second, and the total time T2 is 1-5 seconds. The sampling step size for the temporal data of the surface point cloud of the chest and abdomen and the corresponding tumor temporal location information is the same. It is preferable to use a sliding window for sampling and prediction. The sliding step size for input and output is also the same. Adjacent prediction windows (i.e., outputs) contain overlapping time steps. The trajectory can be initially determined based on the prediction result of the previous prediction window, and then updated using subsequent prediction windows, such as weighted summation updates. This helps improve the real-time performance and trajectory continuity of the online prediction output. When used in radiotherapy equipment, the target coordinates can be initially planned based on the prediction result of the previous prediction window, and then updated based on subsequent prediction windows, improving planning efficiency.

[0038] In step S3, since the original body surface point cloud is usually disordered, unevenly dense, and has a large data volume, directly inputting it into a temporal network would lead to high computational cost, sensitivity to noise, and insufficient utilization of spatial structure. Therefore, this invention regularly divides the abdominal or thoracic abdominal surface region of interest (i.e., the XY plane region) into multiple pseudo-voxel sub-blocks. Each pseudo-voxel sub-block contains a set of point clouds located within a corresponding spatial range. With this division, when the body surface point cloud moves with respiration, the XY region to which each pseudo-voxel sub-block belongs remains unchanged, but the three-dimensional coordinates of the point clouds projected within that pseudo-voxel sub-block region will change, thus causing the center coordinates of all point cloud sets to change.

[0039] For each frame of surface point cloud data, the geometric center of all points within each pseudo-voxel sub-block is calculated to obtain the center coordinates of the pseudo-voxel at the current time. Subsequently, a motion significance index is calculated for the coordinate changes of the geometric center of each pseudo-voxel over the entire time series. The preferred motion significance index is the standard deviation, but it can also be variance, displacement amplitude, period energy, or a weighted combination of multiple indices. Based on the motion significance index, the pseudo-voxels are ranked, and the K pseudo-voxels with the most significant motion are selected as key feature regions.

[0040] Furthermore, when using standard deviation as a motion significance indicator, this invention not only utilizes standard deviation to screen pseudo-voxel sub-blocks, but also retains the standard deviation ranking order of the top K pseudo-voxel sub-blocks, and allows this ranking order to participate in subsequent feature construction. Specifically, pseudo-voxel sub-blocks with larger standard deviations are ranked higher in the body surface motion temporal feature vector, thus enabling the feature vector to reflect not only the spatial motion state of each local region, but also the relative magnitude of motion significance in each local region.

[0041] The geometric center coordinates of the K key pseudo-voxels in each time frame are concatenated in descending order of standard deviation to form a feature vector sequence for the neural network input. Preferably, the standard deviation and / or ranking of each key pseudo-voxel can also be used as a saliency code and input into the network along with the geometric center coordinates to further enhance the role of motion salient regions in feature extraction. If each pseudo-voxel uses three-dimensional center coordinates, the input feature dimension at each time step is 3×K. Preferably, K can be an integer between 10 and 100, for example, 50.

[0042] Compared with global dimensionality reduction methods such as principal component analysis, this invention preserves the spatial distribution and motion differences of local areas on the body surface through pseudo-voxelation, which is beneficial for characterizing the local phase difference and non-rigid deformation characteristics of different sub-regions on the body surface during respiration.

[0043] like Figure 2 The specific implementation steps are as follows: S31. Divide the abdominal surface point cloud region into... Given a set of regular pseudo-voxel sub-blocks, i.e., M rows and N columns of sub-blocks, calculate the geometric center coordinates of the point cloud at time t within the m-th and n-th pseudo-voxel sub-blocks. and the average geometric center coordinates of the total time T1 ; S32. Calculate the standard deviation of the geometric center of the m-th row and n-th column pseudo-voxel sub-block in the time dimension. This is used to quantify the motion salience of the region where the pseudo-sub-block is located: .

[0044] Where i is the sampling point index, The sampling interval; S33, Standard deviation of several pseudo-sub-blocks Arrange the data from largest to smallest, select the top K sub-blocks with the highest motion significance, and keep the standard deviation of the top K sub-blocks in the same sorting order; then concatenate the K center coordinates corresponding to each frame in the total time T1 according to the above sorting order to form a temporal feature vector of body surface motion. Preferably, the standard deviation and / or ranking of the corresponding sub-blocks are used as saliency codes and combined with the coordinates of each center, so that the temporal feature vector of body surface motion simultaneously represents the local motion state and its saliency. .

[0045] Compared to dimensionality reduction methods such as PCA, the above method better preserves local spatial structure and more effectively captures respiratory phase differences between different sub-regions. In particular, this invention does not simply perform a one-time screening based on standard deviation, but further retains the standard deviation ranking results and incorporates them into the feature vector construction process. This allows the feature vector to simultaneously contain geometric motion information and motion saliency hierarchy information of the local region, thus enhancing the contribution of highly saliency pseudo-voxel sub-blocks in subsequent network feature extraction and improving the representation ability of complex non-rigid respiratory deformations in the chest and abdomen. Compared to single-point markers (such as infrared markers in the University of Lübeck dataset), the pseudo-voxelization method of this invention utilizes dense surface point cloud information, which can more effectively capture complex non-rigid respiratory deformations.

[0046] like Figure 4 In step S4, the attention-enhanced dilated convolutional prediction network is RTA-Net, used to establish a nonlinear mapping from surface features to tumor locations in vivo. This network includes multiple cascaded attention-enhanced dilated convolutional modules and an output layer; each attention-enhanced dilated convolutional module includes a causal convolutional unit, a dilated convolutional unit, a self-attention mechanism unit, and a residual connection unit.

[0047] The causal convolution unit is used to perform causal padding and one-dimensional convolution on the input body surface motion temporal feature vector, and output a causally constrained body surface motion temporal feature vector. The causal convolution unit is used to ensure that the prediction result at any time depends only on the input at the current and historical times, avoid the leakage of information at future times, and meet the temporal causality requirements in actual online prediction scenarios.

[0048] The dilated convolutional unit is used to perform causal dilation padding and interval sampling convolution on the causally constrained temporal feature vector of body surface motion to obtain the dilated convolutional features of each layer. Within each dilated convolutional residual block, attention weights between time steps are calculated based on the dilated convolutional features of the current layer, and the temporal features of the current layer are dynamically enhanced according to the attention weights to obtain the enhanced temporal features of the current layer. Subsequently, the enhanced temporal features of the current layer are fused with the input of the layer through residual connections and passed to the next layer, thereby realizing the dynamic enhancement of temporal features layer by layer during the forward propagation of the network, and finally outputting the attention-enhanced features.

[0049] Dilated convolutions expand the temporal receptive field with low computational cost by introducing a gap between sampling points of the convolution kernel, enabling the capture of long-term dependent information in respiratory sequences. The self-attention dynamic enhancement unit maps the output of the current layer's dilated convolution to a query vector, key vector, and value vector, and obtains attention weights between different time steps through relevance calculation. Then, based on the attention weights, the temporal features of the current layer are adaptively weighted and enhanced, enabling the network to dynamically highlight effective information at key moments such as the inspiratory phase, expiratory phase, and phase transition segments at different layers. The residual connection unit is used to fuse the enhanced features of the current layer with the input of the dilated convolutional residual block of the current layer, thereby preserving the original temporal information while suppressing the gradient vanishing problem in deep network training.

[0050] Multiple attention-enhanced dilated convolutional modules can be stacked, with the dilation rate of adjacent layers increasing progressively, for example, by 2, 4, 8, or other exponentially increasing forms. Since each dilated convolutional residual block performs a dynamic enhancement of the temporal representation of the current layer within the block, as the network depth increases, the model can progressively enhance the representation of multi-timescale dependencies and key phase information in respiratory sequences, thereby improving the predictive ability for tumor motion trajectories under complex, non-stationary respiratory patterns.

[0051] The residual connection unit is used to make residual connections within and between each attention-enhanced dilated convolutional module, thereby mitigating the gradient vanishing problem and improving training stability during deep network training. This invention uses the standard deviation of the time dimension as a motion saliency indicator, which can better preserve local spatial construction information and effectively reflect the breathing phase difference between different sub-regions. The output layer is used to output tumor temporal location information at one or more future time points. .

[0052] A historical body surface feature sequence of length T1 is used as input, and the 3D location of the tumor at one or more future time points is used as a supervision label to train the temporal network. The loss function can be the mean squared error loss function, or a combined loss function combining position error and trajectory smoothing constraints. After training, in the online prediction stage, the real-time acquired body surface point cloud sequence is input into the network after undergoing coordinate transformation, pseudo-voxelization, and feature construction in the same manner, and the network can then output the 3D motion trajectory of the tumor at future time points.

[0053] Example 2 A method for predicting the three-dimensional motion trajectory of a tumor includes: real-time acquisition of temporal point cloud data of the human chest and abdomen (the same target as during model training), constructing a real-time temporal feature vector of the body surface motion based on the unified spatial coordinate system established in Example 1 and the selected K pseudo voxel sub-blocks. The real-time temporal feature vector of body surface motion is input into the attention-enhanced dilated convolutional prediction network trained in Example 1, and the three-dimensional motion trajectory of the tumor at future moments is output.

[0054] This invention constructs a unified coordinate relationship between surface point clouds and tumor displacement within the body. It employs a pseudo-voxelization strategy to extract motion-salient local surface features and combines this with an attention-enhanced dilated convolutional network to establish a mapping between surface motion and three-dimensional tumor displacement, thereby achieving high-precision, non-invasive prediction of tumor three-dimensional motion trajectories. This invention uses the standard deviation of the time dimension as a motion saliency indicator, which better preserves local spatial construction information and effectively reflects the respiratory phase difference between different sub-regions. Therefore, the selected pseudo-voxel sub-blocks are more suitable for describing complex non-rigid respiratory deformations in the chest and abdomen. This addresses the shortcomings of existing deep learning models in utilizing the local spatial structure of surface point clouds, their limited ability to model long-term dependencies under complex, non-stationary respiratory patterns, and their difficulty in balancing accuracy and generalization.

[0055] like Figure 5 The left side shows a comparison of prediction results under deep breathing conditions, while the right side shows a comparison of prediction results under shallow breathing conditions. This invention can accurately predict the three-dimensional coordinates of the tumor centroid changing over time under different breathing states, achieving a root mean square error (RMSE) of 1.272 ± 0.168 mm, consistently remaining within the clinically tolerable range of 2 mm. Figure 6 As can be seen, compared with the TCN and GRU models, RMSE was reduced by 13.2% and 15.3%, respectively.

[0056] Example 3 A three-dimensional motion trajectory prediction system for tumors, comprising: The model building module is used to build a tumor three-dimensional motion trajectory prediction model using the tumor three-dimensional motion trajectory prediction model construction method described in Example 1. The data acquisition module is used to collect real-time point cloud data of the human chest and abdomen. The data processing module is used to construct a real-time temporal feature vector of body surface motion based on the unified spatial coordinate system established in the construction method of the tumor three-dimensional motion trajectory prediction model described in Example 1, and the selected K pseudo voxel sub-blocks. The trajectory prediction module is used to input the real-time temporal feature vector of body surface motion into the trained attention-enhanced dilated convolutional prediction network and output the three-dimensional motion trajectory of the tumor at future moments.

[0057] Example 4 A storage medium storing a computer program that, when executed by a processor, implements the tumor three-dimensional motion trajectory prediction method as described in Example 2.

[0058] Example 5 A radiotherapy device, including a computer-readable storage medium and a processor; The computer-readable storage medium is used to store instructions; The processor is used to read instructions stored in the computer-readable storage medium and execute the tumor three-dimensional motion trajectory prediction method as described in Embodiment 2.

[0059] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a three-dimensional motion trajectory prediction model for tumors, characterized in that, Includes the following steps: S1. Collect time-series point cloud data of the human chest and abdomen and the corresponding time-series tumor location information, and establish a unified spatial coordinate system; S2. Divide the time-series data of the body surface point cloud into n time periods with a total time length of T1, and divide the tumor time-series location information into n time periods with a total time length of T2 to obtain n training samples. The time corresponding to the tumor temporal location information of each training sample is the future moment of the time corresponding to the time of the body surface point cloud temporal data; S3. Divide the surface point cloud of the training samples into several pseudo-voxel sub-blocks according to the human projection plane, and select K pseudo-voxel sub-blocks according to motion saliency to obtain the temporal feature vector of surface motion; select pseudo-voxel sub-blocks according to the standard deviation of the geometric center coordinates of the pseudo-voxel sub-blocks, including the following steps: S31. Calculate the geometric center coordinates of the point cloud at time t within each pseudobody sub-block. and the average geometric center coordinates of the total time T1 ; S32. Calculate the standard deviation of the geometric center of each pseudo-sub-block in the time dimension. This is used to quantify the motion salience of the region where the pseudo-sub-block is located: Where i is the sampling point index, The sampling interval; S33, Standard deviation of several pseudo-sub-blocks Sort from largest to smallest, and select the highest significance level of the motion. K Each pseudo-voxel is assigned a sub-block, and its position in the temporal feature vector of body surface motion is determined according to the order of the standard deviations. The geometric center coordinates of the first K pseudo-voxel blocks in each frame of the total time T1 are concatenated, and the standard deviation and / or ranking of each pseudo-voxel block are used as saliency representation information and associated with the corresponding geometric center coordinates to obtain the temporal feature vector of body surface motion. ,in, The temporal feature vector of body surface motion simultaneously carries the spatial motion information of each pseudo-voxel sub-block and its motion saliency information. S4. The attention-enhanced dilated convolutional prediction network is trained by taking the temporal feature vector of body surface movement as input and the corresponding temporal location information of the tumor as output.

2. The method for constructing a three-dimensional tumor motion trajectory prediction model according to claim 1, characterized in that, The attention-enhanced dilated convolutional prediction network includes multiple cascaded attention-enhanced dilated convolutional modules and an output layer; each attention-enhanced dilated convolutional module includes a causal convolutional unit, a dilated convolutional unit, a self-attention mechanism unit, and a residual connection unit. The causal convolution unit is used to perform causal padding and one-dimensional convolution on the input body surface motion temporal feature vector, and output a causally constrained body surface motion temporal feature vector. The dilated convolutional unit is used to perform causal dilation padding and interval sampling convolution on the causal-constrained temporal feature vector of body surface motion to obtain the dilated convolutional features of each layer. The self-attention mechanism unit is used to calculate the attention weights between time steps based on the dilated convolution features of the current layer through the self-attention mechanism, and dynamically enhance the dilated convolution features of the current layer according to the attention weights to obtain the enhanced temporal features of the current layer; the enhanced temporal features are fused with the input residuals of the current layer dilated convolution unit through the residual connection unit and then output to the next attention-enhanced dilated convolution module, thereby realizing the dynamic enhancement of the temporal features layer by layer, and finally outputting the attention-enhanced features; The output layer is used to output tumor temporal location information at one or more future time points.

3. The method for constructing a three-dimensional tumor motion trajectory prediction model according to claim 2, characterized in that, The tumor temporal location information is the three-dimensional coordinates of the tumor centroid at each time step and / or the three-dimensional coordinates of a representative location determined based on the contour segmentation results; Furthermore, the end time of T1 is adjacent to the start time of T2.

4. A method for predicting the three-dimensional motion trajectory of a tumor, characterized in that, include: Real-time point cloud time-series data of human chest and abdomen are collected. Based on the unified spatial coordinate system established in the construction method of tumor three-dimensional motion trajectory prediction model according to any one of claims 1-3, and the selected K pseudo voxel sub-blocks, a real-time body surface motion time-series feature vector is constructed. The real-time temporal feature vector of body surface motion is input into the attention-enhanced dilated convolutional prediction network trained by the method for constructing the tumor three-dimensional motion trajectory prediction model according to any one of claims 1-3, and the three-dimensional motion trajectory of the tumor at future moments is output.

5. A three-dimensional motion trajectory prediction system for tumors, characterized in that, include: The model building module is used to build a tumor three-dimensional motion trajectory prediction model using the method for building a tumor three-dimensional motion trajectory prediction model according to any one of claims 1-3; The data acquisition module is used to collect real-time point cloud data of the human chest and abdomen. The data processing module is used to construct a real-time body surface motion temporal feature vector based on the unified spatial coordinate system established in the construction method of the tumor three-dimensional motion trajectory prediction model according to any one of claims 1-3 and the selected K pseudo voxel sub-blocks. The trajectory prediction module is used to input the real-time temporal feature vector of body surface motion into the trained attention-enhanced dilated convolutional prediction network and output the three-dimensional motion trajectory of the tumor at future moments.

6. A storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the tumor three-dimensional motion trajectory prediction method as described in claim 4.

7. A radiotherapy device, characterized in that, Includes computer-readable storage media and processors; The computer-readable storage medium is used to store instructions; The processor is used to read instructions stored in the computer-readable storage medium and execute the tumor three-dimensional motion trajectory prediction method as described in claim 4.

Images