A Method and System for Extracting Body Surface Respiratory Motion Features Based on Dynamic Region Update

Abstract

This invention belongs to the field of image processing technology and specifically discloses a method and system for extracting respiratory motion features of the body surface based on dynamic region updating. The method includes: collecting temporal point cloud data of the chest and abdomen of a subject during respiration; determining an initial region of interest based on the three-dimensional point cloud and dividing it into multiple pseudo-voxel blocks; calculating the geometric center of the midpoint of each pseudo-voxel block to obtain a region update score characterizing the activity level of the pseudo-voxel block; when the region update score of a pseudo-voxel block is higher than a preset threshold in multiple consecutive time windows, it is included in an active pseudo-voxel set; determining a cross-window stability score based on the salience of the active pseudo-voxel blocks in each time window, thereby selecting stable and salient regions from the active pseudo-voxels; extracting the phase propagation features and cross-window stability scores of the stable and salient regions to obtain the respiratory motion features of the body surface. This invention can achieve stable and accurate extraction of respiratory motion features based on temporal body surface point clouds.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and more specifically, relates to a method and system for extracting respiratory motion features of the body surface based on dynamic region updating. Background Technology

[0002] In scenarios such as surface-guided radiotherapy, respiratory monitoring, and motion management, structured light cameras, depth cameras, binocular vision devices, or other 3D acquisition devices are typically used to continuously sample the patient's chest and abdominal surface, thereby obtaining a time-varying surface point cloud sequence. Since the chest and abdominal surface fluctuates periodically with respiration, the changes in the surface point cloud over time can be analyzed to extract surface motion features that characterize the respiratory state, which can then be further used for respiratory gating, motion compensation, phase recognition, or other clinical auxiliary tasks.

[0003] However, raw body surface point clouds typically have the following characteristics: first, a large number of points and high data dimensionality; second, disordered distribution and uneven density; third, significant influence from acquisition perspective, occlusion, reflection, edge defects, and environmental noise; and fourth, differences in breathing patterns among individuals, and inconsistencies in the amplitude, phase, and stability of motion in different body surface regions. Therefore, how to stably extract key motion features closely related to respiration from disordered, noisy, and potentially locally missing temporal point clouds has always been a crucial technical challenge in this field.

[0004] In existing technologies, common methods for extracting body surface motion features mainly include the following categories.

[0005] The first category is based on global dimensionality reduction methods such as principal component analysis. While these methods can achieve dimensionality reduction to some extent, they typically focus more on overall statistical characteristics and struggle to fully preserve the spatial structural relationships and local motion differences between local areas of the body surface. Furthermore, when background points, local anomalies, or static regions exist in the point cloud, the principal component orientation is easily disturbed, thus affecting the accurate representation of the respiratory-dominant region.

[0006] The second category is motion tracking methods based on manually marked points. These methods require attaching reflective dots to the patient's body surface or manually specifying a small number of discrete feature points. Although this facilitates tracking at specific locations, it is cumbersome to operate, has poor clinical compliance, and can only obtain displacement information of sparse points, making it difficult to reflect the non-rigid deformation process of the continuous surface of the chest and abdomen.

[0007] The third category involves uniform sampling or overall input of the entire body surface area. Since thoracic breathing and abdominal breathing differ in their dominant motion areas, if high-motion, low-motion, and resting areas are not distinguished and all areas are used together for subsequent processing, irrelevant fluctuations can easily be introduced into low-relevance areas such as the vicinity of the clavicle, the vicinity of the sternum, and the vicinity of the bed background, thereby interfering with the extraction and expression of features in high-motion areas.

[0008] Furthermore, while some existing methods incorporate pseudo-voxel segmentation and saliency screening, they often only sort based on the displacement standard deviation or similar amplitude indicators within a single time window, then directly select the top few pseudo-voxel blocks and stitch their center coordinates together. Although this approach can preserve local spatial structure to some extent, it still has the following shortcomings: First, the region of interest is usually fixed, making it difficult to adapt to changes in the patient's breathing pattern, slight body position, or changes in locally visible areas; second, when the local point cloud becomes sparse or missing due to occlusion, changes in sampling angle, or surface reflection, the feature sequence is easily interrupted; third, high-amplitude anomalies in a single window may be misjudged as salient regions, making it difficult to distinguish between continuously stable breathing regions and occasional noise regions; fourth, using only geometric center stitching makes it difficult to fully characterize the respiratory phase propagation relationship between different regions. Summary of the Invention

[0009] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and system for extracting respiratory motion features of the body surface based on dynamic region update, the purpose of which is to achieve stable and accurate extraction of respiratory motion features based on body surface point clouds.

[0010] To achieve the above objectives, according to one aspect of the present invention, a method for extracting surface respiratory motion features based on dynamic region updating is proposed, comprising the following steps: Temporal point cloud data of the chest and abdomen of the subjects were collected during the respiratory process to obtain three-dimensional point clouds at multiple time points; Based on 3D point cloud, an initial region of interest is determined and divided into multiple pseudo-voxel blocks; Calculate the geometric center of each point in the pseudo voxel block as the representative position of that pseudo voxel block; By sliding forward through the time window, the active pseudo-voxel set and stable salient regions are periodically updated as follows: Based on the representative position of the pseudo-voxel block, a region update score representing its activity level is obtained; when the region update score of the pseudo-voxel block is higher than the preset threshold in multiple consecutive time windows, the pseudo-voxel block is included in the active pseudo-voxel set. For each pseudo-voxel block in the active pseudo-voxel set, a cross-window stability score characterizing the stability of the pseudo-voxel block in multiple time windows is determined based on the saliency of the pseudo-voxel block in each time window; based on the cross-window stability score, some pseudo-voxel blocks are selected from the active pseudo-voxel set as stable salient regions. Phase propagation features and cross-window stability scores of stable and significant regions at each time step are extracted to obtain surface respiratory motion features.

[0011] As a further preferred method, the method for determining the regional update score of the pseudo-voxel block within a certain time window is as follows: Select at least two of the following evaluation metrics for pseudo-voxels for calculation: Motion energy evaluation quantity: refers to the displacement amplitude or fluctuation intensity of the pseudo-voxel block within the current time window; Completeness evaluation metrics: refers to the point cloud effectiveness rate, empty block ratio, or repair confidence of pseudo-voxel blocks within the current time window; Neighborhood continuity evaluation: refers to the continuity of pseudo-voxel blocks with their neighboring pseudo-voxels in terms of spatial distribution and temporal variation; Respiratory consistency assessment: refers to the degree of consistency between the movement of the pseudo-voxel block and the overall respiratory rhythm; The calculated values ​​of the selected evaluation quantities are normalized and then weighted and combined to obtain the regional update score.

[0012] As a further preferred method, the method for determining the significance of pseudo-voxel blocks is as follows: Choose at least two of the following pseudo-voxel metrics for calculation: Amplitude index: refers to the displacement amplitude of the pseudo-voxel within the current time window; Periodic consistency index: refers to the degree of consistency between the movement of pseudo-voxel blocks and the main respiratory rhythm; the main respiratory rhythm is a reference respiratory curve obtained by aggregating the displacement curves of multiple pseudo-voxel blocks in the active pseudo-voxel set within the current time window; Completeness index: refers to the point cloud effectiveness and repair reliability of pseudo voxel blocks; Phase correlation index: refers to the temporal synchronization or phase stability of the pseudo-voxel block relative to the reference pseudo-voxel block; the reference pseudo-voxel block is the pseudo-voxel block with the highest joint significance score and whose integrity meets the preset conditions within the current time window. The calculated values ​​of the selected indicators are weighted and combined to obtain a joint significance score that characterizes the significance of pseudo-voxels.

[0013] As a further preferred method, the calculation method for the cross-window stability score is as follows:

[0014] in, For pseudo voxel blocks i Cross-window stability score, For pseudo voxel blocks i The average of the joint significance scores across multiple time windows, For pseudo voxel blocks i The degree of fluctuation in the joint significance score across multiple time windows. This is the balance coefficient.

[0015] As a further preferred approach, a pseudo-voxel block in an active pseudo-voxel set is considered a stable and significant region when it meets all of the following conditions: The cross-window stability score of the pseudo-voxel block is higher than the preset threshold; The cross-window stability score of the pseudo-voxel block fluctuated below the preset threshold; The frequency of pseudo-voxel blocks entering the high-score set in multiple time windows exceeds a preset threshold; the high-score set in a certain time window refers to pseudo-voxel blocks whose joint significance score exceeds the significance threshold in that time window; The integrity and neighborhood synergy of the pseudo-voxel block meet the requirements.

[0016] As a further preferred option, before calculating the geometric center of the pseudo-voxel block, the empty block is repaired, including: When the number of points in a pseudo-voxel block is less than a preset threshold, the pseudo-voxel block is determined to be an empty block; empty blocks are repaired using at least one of the following methods: Spatial neighborhood interpolation repair: weighted interpolation is performed using the geometric centers of the adjacent pseudo-voxels of the empty block; Time-based prediction repair: linear extrapolation or low-order prediction based on the geometric center trajectory of empty blocks at several historical moments; Spatiotemporal fusion restoration: weighted fusion of spatial neighborhood interpolation results and temporal prediction results; Confidence weighting: The repaired pseudo-voxel blocks are assigned a lower confidence weight to reduce their impact in subsequent screening.

[0017] As a further preferred method, the method for determining the phase propagation characteristics is as follows: Among all stable and salient regions, the stable and salient region with the highest cross-window stability score is selected as the reference region; the phase difference and time delay of the displacement curve of the stable and salient region relative to the displacement curve of the reference region are calculated as phase propagation characteristics.

[0018] As a further preferred method, time-series point cloud data of the chest and abdominal surface of the subjects during respiration were collected to obtain initial point clouds at multiple time points. These initial point clouds were then preprocessed to obtain three-dimensional point clouds at multiple time points. At least one of the following preprocessing methods was used: Background point clipping, removal of points from bed boards, clothing edges, or non-body surface areas; outlier removal; smoothing filtering; coordinate normalization or rigid body alignment; surface contour extraction or normal estimation.

[0019] As a further preferred approach, the initial region of interest is determined using one or more of the following methods in combination: Determined based on the bounding box, convex hull, or contour boundary of the chest and abdomen; Determined based on areas with significant undulations in the direction of body surface thickness; Defined based on historical respiratory energy distribution maps or prior anatomical regions; The determination is made after removing obvious background areas based on surface point density, connectivity, and surface continuity.

[0020] According to another aspect of the present invention, a system for extracting respiratory motion features of the body surface based on dynamic region updating is provided, including a processor, the processor being used to execute the above-described method for extracting respiratory motion features of the body surface based on dynamic region updating.

[0021] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention uses pseudo-voxometry expression, combined with active pseudo-voxotype set updates and stable significant region screening across time windows, to adaptively focus on the main respiratory drive sites of different subjects, reducing the reliance on manually defining regions of interest; thereby stably and accurately extracting highly reliable respiratory motion features from unstructured temporal point clouds of the body surface, which can be used for respiratory monitoring, respiratory gating, respiratory phase recognition, body surface motion modeling, and other scenarios requiring alternative respiratory signals from the body surface in surface-guided radiotherapy.

[0022] 2. Enhanced dynamic region update capability: Instead of fixing the region of interest, this invention updates the set of active pseudo-voxels based on a unified pseudo-voxel grid. Through the dynamic region of interest update mechanism, the regions involved in feature extraction can adaptively adjust with changes in breathing patterns, local visibility, and slight body movements, reducing errors caused by fixed region settings.

[0023] 3. Effectively suppresses spurious regions: This invention does not simply select several local regions based on a single time window and a single amplitude index. Instead, it uses a cross-time window stable significance screening mechanism, combined with cross-time window joint significance evaluation and stability analysis, to identify regions that continuously and reliably represent respiration from multiple time windows. This suppresses instantaneous abnormal fluctuations, local noise, and occasional high-amplitude interference, and can more effectively distinguish between the real respiration-dominant region and the occasional noise region.

[0024] 4. More robust to local occlusion and sparse point cloud: Through the empty block detection and repair mechanism, even if some local areas have insufficient point cloud due to occlusion, reflection or edge loss in a short period of time, the continuous output of temporal features of the area can still be maintained, thereby improving the stability of the overall features.

[0025] 5. Better able to express complex respiratory propagation relationships: Based on the cross-window stability score that reflects local displacement characteristics, the introduction of inter-regional phase propagation information can more fully reflect the non-rigid deformation and propagation laws of different regions of the chest and abdomen during the respiratory process, enhance the physical interpretability of the features, and provide input features with higher robustness and stronger physical interpretability for respiratory monitoring, respiratory gating, respiratory phase recognition, body surface motion modeling and other downstream analysis tasks.

[0026] 6. Clear local analysis objects and good temporal consistency: This invention first establishes a unified pseudo-voxel grid in the initial region of interest, and then performs subsequent empty block repair, dynamic region update and saliency evaluation on the grid, so that the entire feature extraction process has clear and stable local statistical objects, which is conducive to maintaining comparability between different time points. Attached Figure Description

[0027] Figure 1 The flowchart illustrates the surface respiratory motion feature extraction method based on dynamic region updating provided in this embodiment of the invention.

[0028] Figure 2 This is a schematic diagram illustrating the determination of the initial region of interest and the establishment of a unified pseudo-voxel mesh, as provided in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram illustrating the cross-time window joint significance evaluation and stable significant region screening provided in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of the final time series feature vector obtained in an embodiment of the present invention. 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 the accompanying drawings and 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] This invention provides a method for extracting surface respiratory motion features based on dynamic region updating, such as... Figure 1 As shown, it includes the following steps: S1. Acquisition and preprocessing of temporal point cloud of body surface.

[0033] Temporal point cloud data of the chest and abdomen surface of the subject during respiration were continuously acquired using a structured light camera, depth camera, binocular vision system, or other 3D acquisition devices to obtain initial point clouds at multiple time points. Each frame of the point cloud includes the spatial coordinate information of at least multiple 3D points. Specifically, a spatial rectangular coordinate system is constructed with the head-to-feet direction as the X direction, the direction between the arms as the Y direction, and the direction of chest rise and fall as the Z direction.

[0034] Furthermore, to improve the stability of subsequent feature extraction, preprocessing is performed on the initial point cloud. Preprocessing includes, but is not limited to, one or more of the following: 1) Background point cropping, used to remove points from bed boards, clothing edges, or non-body surface areas; 2) Outlier removal, used to remove isolated noise points; 3) Smoothing filtering, used to suppress high-frequency random noise; 4) Coordinate normalization or rigid body alignment are used to reduce the impact of device coordinate drift; 5) Surface contour extraction or normal estimation is used to assist in the identification of body surface regions.

[0035] After preprocessing, the main outline of the chest and abdomen can be extracted first by using the two-dimensional projection boundary, density distribution and continuity constraints of the body surface point cloud, and then the main outline can be used as the basis for subsequent dynamic region of interest initialization.

[0036] S2. Initial region of interest is determined.

[0037] In the preprocessed 3D point cloud, the initial region of interest related to respiratory motion is determined.

[0038] Furthermore, the initial region of interest is determined according to one or more of the following methods, with the intersection of the results preferred when multiple methods are combined: 1) Determined based on the bounding box, convex hull, or contour boundary of the chest and abdomen; 2) Determined based on areas with significant undulations in the thickness direction of the body surface; 3) Defined based on historical respiratory energy distribution maps or prior anatomical regions; 4) Determine the target area after removing obvious background areas based on the density of points on the body surface, connectivity, and surface continuity.

[0039] The initial region of interest preferably covers the surface of the chest and abdomen, rather than being limited to a fixed rectangular area, so as to provide an initial reference for subsequent dynamic updates.

[0040] S3. Pseudo-voxelation of the initial region of interest.

[0041] Within the initial region of interest, the surface projection plane is regularly divided to form multiple pseudo-voxel blocks. These pseudo-voxels are not strictly three-dimensional voxels, but rather analysis units formed by regularly meshing a local area of ​​the body surface.

[0042] In one embodiment, the initial region of interest can be divided into the XY plane as follows: Each pseudo-voxel block corresponds to a single pseudo-voxel. In other embodiments, pseudo-voxels can also be formed using fixed-size grids, fixed-number grids, equal-area grids, adaptive grids, or grids divided according to anatomical regions.

[0043] Specifically, a pseudo-voxel mesh is pre-built within the initial region of interest (ROI). A fixed-index pseudo-voxel block geometry is established for the initial ROI, and the mesh index (i.e., pseudo-voxel block partitioning) remains consistent throughout the temporal analysis. Subsequent dynamic ROI updates do not reconstruct the mesh, but rather update the set of active pseudo-voxels within the existing pseudo-voxel mesh. This helps maintain comparability in the temporal dimension and consistency in the feature dimension.

[0044] S4, calculation of pseudo-voxel geometric center and detection and repair of empty blocks.

[0045] For each pseudo-voxel at each time step, count the point cloud points falling within the pseudo-voxel and calculate the geometric center of all points in the pseudo-voxel, which serves as the representative position of the pseudo-voxel at the current time step.

[0046] Specifically, suppose the i-th pseudo-voxel block contains at time t There are a number of points, whose geometric center is denoted as . Then we have:

[0047] in , , Let x, y, and z represent the coordinates of the j-th point in the i-th pseudo-voxel block at time t, respectively, on the x, y, and z axes.

[0048] Furthermore, when a pseudo-voxel block has a point count of 0 at a certain moment, it is determined to be an empty block, with the point count falling below a preset threshold. When a block is identified as a sparse block, it is referred to as an empty block for ease of description. Since empty blocks may be caused by occlusion, changes in sampling perspective, edge loss, surface reflection, or local noise, this invention performs repair processing on empty blocks to avoid interruption of temporal features.

[0049] Empty block repair can be performed using at least one of the following methods: 1) Spatial neighborhood interpolation repair: weighted interpolation is performed using the effective geometric centers of adjacent pseudo-voxels; 2) Time prediction repair: Based on the geometric center trajectory of the pseudo voxel at several historical moments, perform linear extrapolation or low-order prediction; 3) Spatiotemporal fusion repair: The spatial neighborhood interpolation results are weighted and fused with the temporal prediction results; 4) Confidence labeling: Assign lower or adaptive confidence weights to the repaired pseudovoxels to reduce the impact of unreliable regions in subsequent screening.

[0050] Specifically, time-predictive repair can be represented as:

[0051] in, This represents the time interpolation result.

[0052] Spatiotemporal fusion repair can be represented as:

[0053] in, This represents the spatial interpolation result. For prediction coefficients, For weight fusion.

[0054] By calculating the geometric centers of pseudo-voxels and repairing empty blocks, the original disordered point cloud can be transformed into a temporally continuous and locally ordered sequence of representative points of the region, providing a stable input for subsequent dynamic region updates and saliency evaluation.

[0055] S5. Dynamic region of interest update based on pseudo-voxels.

[0056] After completing the pseudo-voxel partitioning and obtaining the temporal geometric center of each pseudo-voxel block, the dynamic region of interest is updated based on the pseudo-voxel level. Here, the dynamic region of interest does not refer to redefining the original geometric bounding range, but rather to dynamically determining the set of active pseudo-voxels currently participating in feature extraction within a given pseudo-voxel grid.

[0057] Specifically, within the sliding time window, at least two of the following evaluation metrics are calculated for each pseudo-voxel block: 1) Motion energy evaluation quantity, used to describe the displacement amplitude or fluctuation intensity of the pseudo-voxel block within the current time window; 2) Completeness evaluation metric, used to describe the point cloud effectiveness, empty block ratio, or repair confidence of the pseudo voxel block within the current time window; 3) Neighborhood continuity evaluation measure, used to describe the continuity of the pseudo-voxel block with its neighboring pseudo-voxels in terms of spatial distribution and temporal variation; 4) Respiratory consistency assessment, used to describe the degree of consistency between the movement of the pseudo-voxel block and the overall respiratory rhythm.

[0058] Based on the above evaluation metrics, a regional update score is constructed. :

[0059] in, This represents the motion energy evaluation value of the i-th pseudo-voxel block within the r-th time window. This represents the measure of completeness. Indicates a continuous evaluation quantity. This indicates the respiratory consistency assessment score. ~ These are the weighting coefficients.

[0060] When a pseudo-voxel block scores below a threshold in multiple consecutive time windows, it is removed from the active pseudo-voxel set of the dynamic region of interest; when a pseudo-voxel block scores above a threshold in multiple consecutive time windows and meets the requirements of completeness and neighborhood continuity, it can be included in the active pseudo-voxel set.

[0061] Dynamic region of interest updates are not performed every frame, but once every preset update step size U frames or after a preset number of sliding time windows, in order to balance real-time performance and stability.

[0062] The above approach allows subsequent significance evaluation and feature construction to focus on local areas that are truly relevant to respiration and have high credibility.

[0063] S6. Joint significance evaluation across time windows.

[0064] To avoid relying solely on displacement magnitudes within a single time window for significance ranking, this invention performs joint significance evaluation across time windows on the active pseudo-voxel set of the dynamically interested region, resulting in a cross-window stability score.

[0065] Specifically, let the sliding time window be... ,in Each time window has a length of L and a window step size of L. Within each time window, calculate at least two of the following metrics for the i-th active pseudo-voxel block: 1) Amplitude index : Used to characterize the displacement amplitude of a local region within the current time window, and can be expressed as longitudinal standard deviation, three-dimensional displacement modulus standard deviation, peak-to-peak value or root mean square; 2) Periodic Consistency Indicators : Used to characterize the consistency between the movement of the pseudo-voxel and the main respiratory rhythm; the main respiratory rhythm is obtained by aggregating the displacement sequences of highly active pseudo-voxel blocks, specifically including: selecting several pseudo-voxel blocks with high significance among the active pseudo-voxel blocks within the current time window; extracting the displacement curves of the above pseudo-voxel blocks; performing an aggregation operation on the above displacement curves to obtain the overall respiratory reference curve, which is the main respiratory rhythm; 3) Completeness index : Used to characterize the point cloud effectiveness and repair reliability of the pseudo voxel block; 4) Phase correlation index : Used to characterize the temporal synchronization or phase stability of the pseudo-voxel block relative to the reference pseudo-voxel block; the reference pseudo-voxel block is the pseudo-voxel block with the highest joint significance score and whose integrity meets the preset conditions within the current time window.

[0066] For example, when the standard deviation of longitudinal displacement is used as the amplitude index, it can be expressed as:

[0067] in, This represents the average z-axis coordinate of the i-th pseudo-voxel within the r-th time window.

[0068] Furthermore, a joint significance score is constructed:

[0069] in, ~ This represents the weighting coefficient.

[0070] To evaluate whether a particular pseudovoxel stably represents respiration across multiple time windows, a cross-window stability score was further calculated. :

[0071] in, for Average value over multiple time windows Its degree of fluctuation, The coefficient represents the balance; the higher the average score and the smaller the fluctuation, the more stable and reliable the pseudo-voxel is.

[0072] S7. Screening for stable and significant regions.

[0073] Based on the results of the joint significance evaluation across time windows, stable and significant regions were screened for active pseudovoxel sets.

[0074] Specifically, a pseudo-voxel block is considered a candidate for a stable and salient region when it simultaneously meets the following conditions: 1) The average score across windows is higher than the preset threshold; 2) The cross-window score fluctuation is lower than the preset threshold; 3) The frequency of entering the high-rating set exceeds a preset threshold across multiple time windows; the high-rating set is determined using a threshold method, formula: ,in This represents the joint significance score of pseudo-voxel block i within the r-th time window. This represents the preset significance threshold; 4) Completeness and neighborhood synergy meet the minimum requirements.

[0075] For spatially adjacent candidate pseudovoxels with similar scores, connectivity analysis can be performed to merge them into a salient region cluster. Preferably, the final output region can be adaptively determined based on the number of stable salient regions, regional connectivity, and spatial distribution; when a fixed-dimensional output is required, representative pseudovoxels can be further selected from the stable salient regions to form a fixed number of output sets.

[0076] This step reduces the risk of occasional high-amplitude areas and isolated noise areas in a single window being mistakenly selected as key respiratory feature areas.

[0077] S8, Phase propagation feature encoding.

[0078] To enhance the expression of respiratory propagation relationships between different body surface regions, this invention further encodes phase propagation features based on stable and significant regions.

[0079] Specifically, a reference region is selected from the stable and significant regions, preferably the region with the highest cross-window stability score, the highest integrity, or the most stable cross-window. For other stable and significant regions, the phase difference and time delay of their displacement curves relative to the reference region are calculated to characterize the propagation sequence of respiratory fluctuations between different regions.

[0080] Phase difference or time delay can be obtained through cross-correlation peak position, frequency domain phase difference analysis, or other time-series alignment methods. By introducing phase propagation characteristics, the asynchronous motion patterns between different body surface regions under thoracic breathing, abdominal breathing, and mixed breathing conditions can be more fully characterized.

[0081] S9. Construction of Time Series Feature Vectors and Time Window Update.

[0082] For the selected stable and significant regions, the phase propagation features relative to the reference region are extracted based on the displacement curve of the stable and significant regions within a preset time window. Combined with the cross-window stability score corresponding to the stable and significant regions, the time series feature vector of body surface respiratory motion is obtained.

[0083] Specifically, let the set of ultimately stable salient regions be... Then the eigenvector at time t can be expressed as:

[0084] in, This represents the phase difference between the nth pseudo-pixel block in the stable salient region and the reference region. This represents its current window stability score. This represents the time delay relative to the reference region, n=1,2… K .

[0085] As the sliding window moves forward, it periodically updates the active pseudo-voxel set, the joint significance score, and the stable salient region, ensuring that the final feature possesses both temporal continuity and stability while adapting to slow changes in breathing patterns. The resulting surface respiratory motion features can be used for respiratory monitoring, respiratory gating, respiratory phase recognition, surface motion state analysis, or other scenarios requiring robust surface respiratory representation.

[0086] The present invention provides a system for extracting respiratory motion features of the body surface based on dynamic region updating, characterized in that it includes a processor, which is used to execute the above-described method for extracting respiratory motion features of the body surface based on dynamic region updating.

[0087] The following are specific examples: A structured light camera was used to continuously acquire temporal point cloud data of the subject's chest and abdomen during respiration. The sampling frequency can be set according to the device performance; in this embodiment, it is 10Hz to 60Hz. Background cropping, outlier removal, and smoothing filtering were then performed on the original point cloud sequentially to obtain a preprocessed point cloud sequence.

[0088] In the preprocessed point cloud, an initial region of interest is determined based on the 2D projection boundary and the main connected components. A pseudo-voxel mesh with fixed indices is then established within this region; for example, the region is divided into... A pseudo-void block.

[0089] For each pseudo-voxel block at each time step, calculate the geometric center and the number of valid points. If the number of pseudo-voxel points falls below a threshold... When it is determined to be an empty block, it is spatiotemporally fused and repaired by combining the spatial interpolation results of adjacent pseudo-voxels with the historical trajectory prediction results of the pseudo-voxel.

[0090] Then, using a sliding time window approach, the motion energy, integrity, neighborhood continuity, and breathing consistency of each pseudo-voxel block are calculated to update the active pseudo-voxel set. For the pseudo-voxels in the active pseudo-voxel set, joint significance scores and cross-window stability evaluations are further calculated to screen stable and significant regions, and their displacement and phase propagation features are extracted to form the final time series feature vector.

[0091] The active pseudo-voxel set is updated in units of a preset-length sliding time window. For each pseudo-voxel, the displacement standard deviation, percentage of valid points, neighborhood correlation, and correlation coefficient with the reference respiration curve are calculated. If a pseudo-voxel consistently scores below a threshold across multiple time windows, it is removed from the active pseudo-voxel set; conversely, if a pseudo-voxel consistently scores high and has sufficient integrity across multiple time windows, it is added to the active pseudo-voxel set. In this implementation, the pseudo-voxel mesh is not re-divided; only the active pseudo-voxel set participating in subsequent analyses is updated.

[0092] For each active pseudo-voxel, a joint score is calculated across multiple time windows, and its average score, fluctuation level, and selection frequency are statistically analyzed. Only pseudo-voxels with high average scores, low fluctuations, high frequencies, and satisfactory completeness are retained as candidate regions. Then, connectivity analysis is performed on the candidate pseudo-voxels, merging spatially adjacent candidate pseudo-voxel blocks into region clusters. Finally, the output region is determined based on the region cluster scores and distribution.

[0093] In summary, this invention can extract robust respiratory motion features from temporal surface point clouds in unstructured surface point cloud scenarios through dynamic region of interest update, pseudo-voxelation expression, empty block repair, cross-time window stability saliency screening, and phase propagation encoding. It can be applied to respiratory monitoring, respiratory gating, respiratory phase recognition, surface motion modeling, and other scenarios requiring alternative respiratory signals from the surface in surface-guided radiotherapy.

[0094] 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 extracting surface respiratory motion features based on dynamic region updating, characterized in that, Includes the following steps: Temporal point cloud data of the chest and abdomen of the subjects were collected during the respiratory process to obtain three-dimensional point clouds at multiple time points; Based on 3D point cloud, an initial region of interest is determined and divided into multiple pseudo-voxel blocks; Calculate the geometric center of each point in the pseudo voxel block as the representative position of that pseudo voxel block; By sliding forward through the time window, the active pseudo-voxel set and stable salient regions are periodically updated as follows: Based on the representative position of the pseudo-voxel block, a region update score representing its activity level is obtained; when the region update score of the pseudo-voxel block is higher than the preset threshold in multiple consecutive time windows, the pseudo-voxel block is included in the active pseudo-voxel set. For each pseudo-voxel block in the active pseudo-voxel set, a cross-window stability score characterizing the stability of the pseudo-voxel block in multiple time windows is determined based on the salience of the pseudo-voxel block in each time window. Based on the cross-window stability score, some pseudo-voxel blocks are selected from the active pseudo-voxel set as stable and significant regions; Phase propagation features and cross-window stability scores of stable and significant regions at each time step are extracted to obtain surface respiratory motion features.

2. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, The method for determining the update score of a pseudo-voxel block within a certain time window is as follows: Select at least two of the following evaluation metrics for pseudo-voxels for calculation: Motion energy evaluation quantity: refers to the displacement amplitude or fluctuation intensity of the pseudo-voxel block within the current time window; Completeness evaluation metrics: refers to the point cloud effectiveness rate, empty block ratio, or repair confidence of pseudo-voxel blocks within the current time window; Neighborhood continuity evaluation: refers to the continuity of pseudo-voxel blocks with their neighboring pseudo-voxels in terms of spatial distribution and temporal variation; Respiratory consistency assessment: refers to the degree of consistency between the movement of the pseudo-voxel block and the overall respiratory rhythm; The calculated values ​​of the selected evaluation quantities are normalized and then weighted and combined to obtain the regional update score.

3. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, The method for determining the significance of pseudo-voxel blocks is as follows: Choose at least two of the following pseudo-voxel metrics for calculation: Amplitude index: refers to the displacement amplitude of the pseudo-voxel within the current time window; Periodic consistency index: refers to the degree of consistency between the movement of pseudo-voxel blocks and the main respiratory rhythm; the main respiratory rhythm is a reference respiratory curve obtained by aggregating the displacement curves of multiple pseudo-voxel blocks in the active pseudo-voxel set within the current time window; Completeness index: refers to the point cloud effectiveness and repair reliability of pseudo voxel blocks; Phase correlation index: refers to the temporal synchronization or phase stability of the pseudo-voxel block relative to the reference pseudo-voxel block; the reference pseudo-voxel block is the pseudo-voxel block with the highest joint significance score and whose integrity meets the preset conditions within the current time window. The calculated values ​​of the selected indicators are weighted and combined to obtain a joint significance score that characterizes the significance of pseudo-voxels.

4. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 3, characterized in that, The method for calculating the cross-window stability score is as follows: in, For pseudo voxel blocks i Cross-window stability score, For pseudo voxel blocks i The average of the joint significance scores across multiple time windows, For pseudo voxel blocks i The degree of fluctuation in the joint significance score across multiple time windows. This is the balance coefficient.

5. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, A pseudo-voxel block in an active pseudo-voxel set is considered a stable and significant region if it meets all of the following conditions: The cross-window stability score of the pseudo-voxel block is higher than the preset threshold; The cross-window stability score of the pseudo-voxel block fluctuated below the preset threshold; The frequency of pseudo-voxel blocks entering the high-scoring set in multiple time windows exceeds a preset threshold; The high-scoring set for a certain time window refers to the pseudo-voxel blocks whose joint significance score exceeds the significance threshold within that time window; The integrity and neighborhood synergy of the pseudo-voxel block meet the requirements.

6. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, Before calculating the geometric center of the pseudo-voxel block, the empty block is repaired, including: When the number of points in a pseudo-voxel block is less than a preset threshold, the pseudo-voxel block is determined to be an empty block; empty blocks are repaired using at least one of the following methods: Spatial neighborhood interpolation repair: weighted interpolation is performed using the geometric centers of the adjacent pseudo-voxels of the empty block; Time-based prediction repair: linear extrapolation or low-order prediction based on the geometric center trajectory of empty blocks at several historical moments; Spatiotemporal fusion restoration: weighted fusion of spatial neighborhood interpolation results and temporal prediction results; Confidence weighting: The repaired pseudo-voxel blocks are assigned a lower confidence weight to reduce their impact in subsequent screening.

7. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, The method for determining phase propagation characteristics is as follows: Among all stable and salient regions, the stable and salient region with the highest cross-window stability score is selected as the reference region; the phase difference and time delay of the displacement curve of the stable and salient region relative to the displacement curve of the reference region are calculated as phase propagation characteristics.

8. The method for extracting surface respiratory motion features based on dynamic region updating as described in claim 1, characterized in that, Temporal point cloud data of the chest and abdomen of the subjects during the respiratory process were collected to obtain initial point clouds at multiple time points. The initial point clouds were preprocessed to obtain three-dimensional point clouds at multiple time points. Preprocessing is performed using at least one of the following methods: Background point clipping, removal of points from bed boards, clothing edges, or non-body surface areas; outlier removal; smoothing filtering; coordinate normalization or rigid body alignment; surface contour extraction or normal estimation.

9. The method for extracting surface respiratory motion features based on dynamic region updating as described in any one of claims 1-8, characterized in that, The initial region of interest is determined using one or more of the following methods: Determined based on the bounding box, convex hull, or contour boundary of the chest and abdomen; Determined based on areas with significant undulations in the direction of body surface thickness; Defined based on historical respiratory energy distribution maps or prior anatomical regions; The determination is made after removing obvious background areas based on surface point density, connectivity, and surface continuity.

10. A system for extracting surface respiratory motion features based on dynamic region updating, characterized in that, Includes a processor for performing the body surface respiratory motion feature extraction method based on dynamic region updating as described in any one of claims 1-9.

Images