A three-dimensional key point-based automatic evaluation method and system for pet radio positioning

Abstract

The application discloses a pet radiation positioning automatic evaluation method and system based on three-dimensional key points, relates to the field of veterinary image technology, and comprises the following steps: three-dimensional pose geometric features of a pet are extracted from a three-dimensional key point sequence and a bed surface reference plane; the three-dimensional pose geometric features are compared and mapped with a clinical allowable threshold value to obtain a pose quantitative feature vector; the pose quantitative feature vector is quantitatively solved in combination with a severe deviation truncation and a segmented punishment mechanism to generate a positioning multi-dimensional feature parameter set; according to a score dimension in the positioning multi-dimensional feature parameter set and a corresponding deviation direction, a rectification rule library is matched to obtain a parameterized adjustment amount, and the parameterized adjustment amount is mapped and converted to obtain a real-time rectification instruction sequence; the pet positioning is closed-loop adjusted according to the real-time rectification instruction sequence, and when a plurality of continuous frames meet a target positioning template requirement, an automatic evaluation record is obtained. The application realizes automatic evaluation, deviation direction determination and parameterized rectification guidance of the pet positioning state.

Description

Technical Field

[0001] This invention relates to the field of veterinary imaging technology, and in particular to an automatic evaluation method and system for pet radiographic positioning based on three-dimensional key points. Background Technology

[0002] With the development of veterinary imaging, digital X-ray imaging technology, and intelligent sensing technology, pet radiological examinations have gradually evolved from relying on manual experience in positioning and image interpretation to a technical model that combines digital imaging equipment, standardized examination bed accessories, and image-assisted analysis. In clinical applications, to obtain a projection position that meets diagnostic requirements, it is usually necessary to adjust the positioning based on the pet's body surface landmarks, spinal alignment, and left-right symmetry, and to judge the quality of the positioning based on the imaging results. The continuous development of RGB cameras, depth cameras, image recognition, and key point detection algorithms, along with posture analysis methods based on visual perception, has provided the technical foundation for the digital acquisition, spatial pose analysis, and automated evaluation of pet radiological positioning.

[0003] Current pet radiology positioning techniques largely rely on human experience or two-dimensional image observation. There is a lack of unified, quantifiable, and real-time feedback methods for assessing three-dimensional posture deviations of the pet relative to the examination table, particularly trunk rotation, limb symmetry shifts, and the spatial relationships of key points. Therefore, it is difficult to provide continuous and objective parameter data for positioning adjustments. Accordingly, there is an urgent need for an automated pet radiology positioning assessment technology that combines RGB-D data, image recognition, three-dimensional key point reconstruction, and clinical threshold comparison to improve the objectivity and standardization of positioning assessments. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automatic evaluation method for pet radial positioning based on three-dimensional key points to solve the problems of difficulty in three-dimensional quantification and automatic determination of pet positioning status and difficulty in providing real-time parametric correction guidance during positioning.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an automatic assessment method for pet radial positioning based on three-dimensional key points, comprising: acquiring color and depth images of the pet; registering the depth images to the color image coordinate system and filtering out invalid depth pixels to obtain an RGB-D image frame sequence; identifying two-dimensional key points of the pet based on the RGB-D image frame sequence; performing pinhole back-projection to reconstruct the three-dimensional coordinates by combining camera intrinsic parameters and neighborhood robust depth values; determining the imaging angle based on the cross product sign of the spine and symmetrical key points; and performing spatial coordinate system transformation by fitting the examination bed reference plane using a random consistency algorithm to obtain a three-dimensional key point sequence; and extracting the three-dimensional key points from the data. The three-dimensional pose geometric features of the pet are extracted from the sequence and the bed reference plane. The three-dimensional pose geometric features are mapped and compared with the clinically acceptable threshold to obtain the pose quantification feature vector. The quantification is then performed by combining severe deviation truncation and segmented penalty mechanism to generate a multi-dimensional feature parameter set for positioning. The pet's positioning is matched with the correction rule library according to the scoring dimension and corresponding deviation direction in the multi-dimensional feature parameter set to obtain the parameterized adjustment amount. The parameterized adjustment amount is then mapped and transformed to obtain the real-time correction command sequence. The pet's positioning is adjusted in a closed loop according to the real-time correction command sequence. When the target positioning template requirements are met for multiple consecutive frames, an automatic evaluation record is obtained.

[0007] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the RGB-D image frame sequence are as follows: Collect color images, depth images, and time synchronization information of the pet; perform pixel-level registration based on the coordinate projection relationship from the depth image to the color image to obtain aligned dual-modal image data. Using time synchronization information, spatial consistency, temporal continuity, and geometric closed-loop constraint analysis are performed on the aligned bimodal image data, and invalid depth pixels are filtered out to obtain an RGB-D image frame sequence.

[0008] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the method involves identifying two-dimensional key points of the pet based on RGB-D image frame sequences, performing pinhole back projection to reconstruct three-dimensional coordinates by combining camera intrinsic parameters and neighborhood robust depth values, and determining the imaging angle based on the cross product sign of the spine and symmetrical key points. The specific steps are as follows: Pet keypoint candidate locations, local depth structure information, and cross-frame trajectory information are extracted from RGB-D image frame sequences and jointly organized to obtain a multimodal keypoint candidate field; By performing topological constraints, temporal consistency calibration, and local depth continuity screening on the multimodal keypoint candidate field, stable two-dimensional keypoint data is obtained. Based on stable two-dimensional keypoint data, neighborhood depth backtracking and pinhole backprojection calculations are performed to obtain three-dimensional keypoint skeleton data. The three-dimensional keypoint skeleton data is then jointly discriminated with the cross product signs corresponding to the spinal principal axis and symmetrical keypoints to obtain imaging viewpoint label data.

[0009] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the three-dimensional key point sequence are as follows: Animal occupancy and exclusion information and visible area information of the bed surface are extracted from 3D keypoint skeleton data and imaging viewpoint label data, and then cropped and filtered in combination with depth point cloud to obtain a set of candidate points on the bed surface; The candidate point set of the bed surface, the point-level planarity prior, and the historical bed surface memory are sampled and scored for quality ranking to obtain the bed surface master hypothesis data. Normal vector disambiguation and spatial coordinate system construction were performed on the master hypothesis data of the bed surface, and coordinate transformation was performed on the 3D key point skeleton data of the pet to obtain the 3D key point sequence.

[0010] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the pose quantization feature vector are as follows: The spinal principal axis, symmetrical bone segments, and key point height distribution are extracted from the three-dimensional key point sequence and the bed reference plane, and joint analysis is performed to obtain three-dimensional pose geometric data. The three-dimensional pose geometry data, imaging viewpoint labels and examination position templates are matched and combined with clinically acceptable thresholds for conditional mapping to obtain pose threshold constraint data. Based on the pose threshold constraint data, the body position deviation, trunk tilt, body symmetry and spinal curvature are quantitatively calculated to obtain a multi-dimensional pose quantification parameter set. The multi-dimensional pose quantification parameter set is then jointly judged with the bed surface reference relationship and deviation rules to obtain the pose quantification feature vector.

[0011] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for generating the multi-dimensional feature parameter set of positioning are as follows: The deviation of each dimension is extracted from the pose quantization feature vector and mapped with the examination position template and the clinically acceptable range to obtain the deviation input stream; The deviation from the input stream, the severe deviation truncation factor, and the segmentation penalty rule are quantized to obtain a single-dimensional quantization parameter sequence. Based on the single-dimensional quantization parameter sequence, the deviation components are separated and integrated to generate a multi-dimensional feature parameter set for the placement.

[0012] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the parameterized adjustment amount are as follows: Scoring dimensions, deviation direction, and severity information are extracted from the multidimensional feature parameters of the positioning, and hierarchical matching is performed in combination with the correction rule graph to obtain candidate correction rules; Candidate correction rules are analyzed and continuously solved, and then screened in conjunction with safety constraints to obtain parameterized adjustment amounts.

[0013] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the real-time correction instruction sequence are as follows: The target body part, direction of motion, and severity information are extracted from the parameterized adjustment values, and then mapped and transformed according to the action semantic map to obtain action semantic information. The semantic information of the actions is prioritized and instantiated in language, and then filtered in combination with historical instruction constraints to obtain a real-time correction instruction sequence.

[0014] As a preferred embodiment of the automatic evaluation method for pet radial positioning based on three-dimensional key points described in this invention, the specific steps for obtaining the automatic evaluation record are as follows: The pet's positioning status is retested in a closed loop based on the real-time correction instruction sequence, and the residual change, directional consistency and cross-frame stability are jointly tracked in conjunction with the target positioning template to obtain closed-loop execution information. The closed-loop execution information is converged and confirmed, and multiple consecutive frames are judged to meet the standards. The key frames and adjustment history are recorded and solidified to generate automatic evaluation records.

[0015] Secondly, the present invention provides an automatic evaluation system for pet radial positioning based on three-dimensional key points, including: an RGB-D acquisition module for acquiring color images and depth images of the pet, registering the depth images to the color image coordinate system and filtering out invalid depth pixels to obtain an RGB-D image frame sequence; The key point reconstruction module identifies two-dimensional key points of the pet based on RGB-D image frame sequences. It performs pinhole backprojection reconstruction of three-dimensional coordinates using camera intrinsic parameters and robust neighborhood depth values. The imaging perspective is determined based on the cross product sign of the spine and symmetrical key points. A random consistency algorithm is used to fit the examination bed reference plane for spatial coordinate system transformation, resulting in a three-dimensional key point sequence. The pose assessment module extracts the pet's three-dimensional pose geometric features from the three-dimensional key point sequence and the bed reference plane. These features are mapped and compared with clinically permissible thresholds to obtain a quantified pose feature vector. A severe deviation truncation and segmented penalty mechanism are used for quantification to generate a multi-dimensional feature parameter set for positioning. The deviation correction command module matches the scoring dimension and corresponding deviation direction in the multi-dimensional feature parameter set to a correction rule library, obtains parameterized adjustment amounts, and maps these amounts to obtain a real-time correction command sequence. The automatic evaluation and recording module performs closed-loop adjustments to the pet's positioning based on the real-time correction command sequence and obtains an automatic evaluation record when the target positioning template requirements are met for multiple consecutive frames.

[0016] The beneficial effects of this invention are as follows: Based on image recognition, combined with RGB-D data acquisition, two-dimensional key point detection, pinhole back projection three-dimensional reconstruction, and bed reference plane coordinate transformation, key anatomical parts in the pet radiographic positioning process are expressed in three-dimensional space; through pose geometric feature extraction, clinical threshold mapping, deviation quantification analysis, and correction rule matching, automatic assessment of pet positioning status, deviation direction determination, and parameterized correction guidance are achieved; further, by combining closed-loop retesting and continuous multi-frame standard determination, automatic assessment record generation is completed, thereby improving the objectivity, standardization, real-time performance, and intelligence level of pet radiographic positioning assessment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an automatic evaluation method for pet radial positioning based on three-dimensional key points.

[0019] Figure 2 This is a schematic diagram of an automatic evaluation system for pet radial positioning based on three-dimensional key points.

[0020] Figure 3 A flowchart for obtaining imaging viewpoint label data.

[0021] Figure 4The flowchart for obtaining pose quantization feature vectors. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an automatic evaluation method for pet radial positioning based on three-dimensional key points, including the following steps: S1. Acquire color and depth images of the pet, register the depth images to the color image coordinate system, and filter out invalid depth pixels to obtain an RGB-D image frame sequence.

[0026] S1.1 Acquire color images, depth images, and time synchronization information of the pet. Perform pixel-level registration based on the coordinate projection relationship from the depth image to the color image to obtain aligned dual-modal image data.

[0027] Furthermore, color images, depth images, and time synchronization information of the pet are collected. Based on the coordinate projection relationship from the depth image to the color image, the depth pixels are mapped to the color image coordinate system, and out-of-bounds pixels and invalid depth pixels are removed to obtain aligned bimodal image data.

[0028] It should be noted that the coordinate projection relationship is set by the intrinsic and extrinsic parameter calibration results of the depth camera and the color camera, and is used to transform the pixels in the depth image into the color image coordinate system according to the imaging model.

[0029] S1.2. Use time synchronization information to perform spatial consistency, temporal continuity and geometric closed-loop constraint analysis on the aligned bimodal image data, and filter out invalid depth pixels to obtain RGB-D image frame sequence.

[0030] It should be noted that time synchronization information is used to identify whether the color image and the depth image correspond to the same acquisition time. It usually includes timestamps, frame numbers or trigger markers to ensure that the bimodal data are accurately matched in time.

[0031] Furthermore, time synchronization information is used to perform temporal correspondence on the aligned bimodal image data, and combined with spatial consistency constraints, temporal continuity constraints, depth continuity constraints, and geometric closed-loop constraints of adjacent frames, pixels with temporal misalignment, spatial abnormalities, and invalid depth are filtered out to obtain an RGB-D image frame sequence. Among them, geometric closed-loop constraints refer to comparing the cumulative pose transformation result of the current frame through intermediate frames to historical frames with the direct pose transformation result between the current frame and historical frames, and filtering out abnormal pixels or abnormal frames based on the comparison results.

[0032] S2. Based on the RGB-D image frame sequence, identify the two-dimensional key points of the pet, combine the camera intrinsic parameters and neighborhood robust depth values ​​to perform pinhole back projection to reconstruct the three-dimensional coordinates, and determine the imaging angle based on the cross product sign of the spine and symmetrical key points. Use the random consistency algorithm to fit the examination bed reference plane to perform spatial coordinate system transformation and obtain the three-dimensional key point sequence.

[0033] S2.1 Extract pet keypoint candidate locations, local depth structure information, and cross-frame trajectory information from the RGB-D image frame sequence, and jointly organize them to obtain a multimodal keypoint candidate field.

[0034] Furthermore, key point candidate positions and corresponding local depth structure information of each key part of the pet are extracted from the RGB-D image frame sequence. Cross-frame trajectory information is extracted by combining the position change relationship in adjacent image frames. The key point candidate positions, local depth structure information and cross-frame trajectory information are jointly organized to obtain a multimodal key point candidate field.

[0035] S2.2. Perform topological constraints, temporal consistency calibration, and local depth continuity screening on the multimodal key point candidate field to obtain stable two-dimensional key point data.

[0036] Furthermore, the multimodal keypoint candidate field is subjected to anatomical topological constraints to screen out keypoint candidate positions that do not conform to the connection relationship of the pet's body structure. Temporal consistency calibration is performed by combining the positional change relationship in adjacent image frames. Then, based on the continuity of depth change in the neighborhood of the keypoint candidate position, keypoint candidate positions with abrupt depth changes or abnormalities are screened out to obtain stable two-dimensional keypoint data.

[0037] S2.3. Based on stable two-dimensional key point data, perform neighborhood depth backtracking and pinhole backprojection calculations to obtain three-dimensional key point skeleton data. Then, perform joint discrimination with the cross product symbols corresponding to the spinal principal axis and symmetrical key points to obtain imaging viewpoint label data.

[0038] It should be noted that stable 2D keypoint data refers to 2D keypoint data that, after undergoing anatomical topological constraints, temporal consistency calibration, and local depth continuity screening, exhibits stable positional changes, clear correspondences of body structures, and continuous changes in neighborhood depth within consecutive image frames.

[0039] Symmetrical key points refer to a pair of key points located on either side of the pet's central axis, corresponding to each other in anatomical structure, used to characterize the left-right directional features of the pet's body.

[0040] Furthermore, based on stable 2D keypoint data, the depth values ​​of each 2D keypoint location and its neighborhood range are found in the depth map corresponding to the RGB-D image frame sequence. Invalid and abnormal depth values ​​within the neighborhood range are removed, and a robust neighborhood depth value that can represent the actual spatial position of the corresponding 2D keypoint is determined based on the remaining valid depth values. Then, combined with camera intrinsic parameters, according to the pinhole imaging model, the image coordinates and robust neighborhood depth values ​​corresponding to each stable 2D keypoint are back-projected to restore the 2D keypoints in the image plane to 3D spatial coordinates in the camera coordinate system. Finally, 3D keypoint skeleton data is constructed based on the body structure connection relationship between the key parts of the pet.

[0041] The expression for calculating the robustness depth of the neighborhood is as follows: ; in, Indicates the first Stable two-dimensional keypoint data within the neighborhood range Inner pixel position Depth weight at location; Indicates the first Image coordinates of stable two-dimensional keypoint data; This represents the spatial distance attenuation parameter within the neighborhood range, which is set according to the size of the neighborhood window. This represents a decision function that takes a value of 1 if the condition within the parentheses is true, and a value of 0 otherwise. Indicates pixel position The depth value at that location; Indicates the first The neighborhood range corresponding to each stable two-dimensional key point data The median of the effective depth values ​​within the interior; This indicates the depth deviation screening threshold, which is set according to the sensor noise and ranging range, and can generally be taken as 5mm to 50mm. Indicates the first The neighborhood robust depth value corresponding to each stable two-dimensional keypoint data; This indicates a smoothing term to prevent the denominator from being zero.

[0042] Based on the 3D keypoint skeleton data, the spatial direction corresponding to the main axis of the spine is extracted, and the spatial connection relationship between symmetrical keypoints is determined. Then, the cross product sign of the spatial direction corresponding to the main axis of the spine and the spatial relationship corresponding to the symmetrical keypoints is determined, and the overall orientation presented by the 3D keypoint skeleton data is combined for joint determination, thereby distinguishing the viewpoint category corresponding to the imaging direction and obtaining the imaging viewpoint label data.

[0043] S2.4 Extract animal occupancy and exclusion information and visible area information of the bed surface from the 3D key point skeleton data and imaging viewpoint label data, and combine them with depth point cloud for cropping and filtering to obtain a set of candidate points on the bed surface.

[0044] It should be noted that 3D keypoint skeleton data refers to the 3D coordinates of each key part of the pet and the skeletal spatial distribution data formed by the connection relationship of the body structure.

[0045] Imaging viewpoint label data refers to the imaging direction, body orientation, and left-right relationship data corresponding to the current image frame, which are determined based on the three-dimensional keypoint skeleton data.

[0046] Animal occupancy exclusion information refers to the area occupied by a pet's body in three-dimensional space, used to exclude point cloud areas belonging to the pet's body.

[0047] The visible area information of the bed surface refers to the area of ​​the bed surface that is not obscured by the pet's body and is within the field of view from the current perspective, and is used to retain the point cloud area that can be used for bed surface fitting.

[0048] Furthermore, based on 3D keypoint skeleton data, the position range of key parts such as the pet's head, torso, limbs, and tail in 3D space is determined. The spatial occupancy area of ​​the pet's body is constructed by combining the skeleton connection relationship between keypoints, and this spatial occupancy area is corresponding to the animal occupancy exclusion information. At the same time, based on the imaging viewpoint label data, the occlusion relationship of the pet's body relative to the examination bed under the current imaging direction is determined, and the bed surface area that is not occluded by the pet's body and is within the current field of view is inferred as the visible area information of the bed surface. In the depth point cloud, point cloud data whose projection falls within the visible area information of the bed surface and whose 3D coordinates are outside the animal occupancy exclusion information are retained. Furthermore, based on the spatial continuity, local height variation, outlier distribution characteristics, and consistency with the surface morphology of the point cloud, isolated scattered points, suspended points, edge interference points, and non-bed surface points that obviously belong to the pet's body or the surrounding background are cropped and filtered to obtain a set of candidate points for bed surface fitting.

[0049] S2.5. Perform quality sorting sampling and consistency scoring on the candidate point set of the bed surface, the point-level planarity prior, and the historical bed surface memory to obtain the bed surface master hypothesis data.

[0050] Furthermore, a predetermined number of point cloud data are randomly sampled from the candidate point set of the bed surface as a minimum sampling subset. A candidate plane hypothesis is constructed based on this minimum sampling subset. The distance deviation, normal angle deviation, and neighborhood continuity index of the remaining point cloud data in the candidate point set to the candidate plane hypothesis are calculated. Point cloud data that satisfy at least one of the predetermined conditions (distance threshold, normal threshold, and continuity threshold) are identified as the consistency points of the candidate plane hypothesis. Based on the number of consistency points, the prior weight of point-level planeness, the continuity and dispersion of spatial distribution, each candidate plane hypothesis is ranked and scored for consistency. Random sampling, plane construction, and consistency point statistics are repeated to obtain multiple candidate plane results. Each candidate plane result is then matched and compared with the bed surface location, normal direction, and scale range in historical bed surface memory, and its historical consistency score in spatial location, normal consistency, and regional overlap is calculated. Based on the joint score of the current frame consistency score and the historical consistency score, the candidate plane result with the best score and satisfying the predetermined conditions is selected as the master hypothesis data for the bed surface.

[0051] S2.6 Perform normal vector disambiguation and spatial coordinate system construction on the master hypothesis data of the bed surface, and perform coordinate transformation on the pet's 3D key point skeleton data to obtain the 3D key point sequence.

[0052] Furthermore, the positional and normal vector information of the plane corresponding to the master hypothesis data of the bed surface is extracted. Addressing the ambiguity of the plane normal vector's direction, at least one constraint is considered, including the actual orientation of the examination bed, the camera's observation direction, the spatial relationship of the bed surface, and the continuity of the bed surface posture in adjacent frames, to determine the plane normal direction and identify the target normal vector consistent with the actual orientation of the examination bed. Based on the plane's spatial position, the target normal vector, and at least one reference direction (length or width) of the bed surface, a spatial coordinate system corresponding to the actual posture of the examination bed is established. Then, the 3D coordinates of each keypoint in the pet's 3D keypoint skeleton data in the original coordinate system are uniformly transformed to the examination bed spatial coordinate system through rotation and translation transformations. The transformed keypoint data of each frame is then organized according to the temporal order of the image frames to obtain a 3D keypoint sequence that can characterize the continuous posture changes of the pet.

[0053] S3. Extract the three-dimensional pose geometric features of the pet from the three-dimensional key point sequence and the bed reference plane, map and compare the three-dimensional pose geometric features with the clinically permissible threshold to obtain the pose quantification feature vector, and perform quantification calculation by combining severe deviation truncation and segmentation penalty mechanism to generate a multi-dimensional feature parameter set for positioning.

[0054] S3.1 Extract the spinal principal axis, symmetrical bone segments, and key point height distribution from the three-dimensional key point sequence and the bed reference plane, and perform joint analysis to obtain three-dimensional pose geometric data.

[0055] It should be noted that the spinal axis refers to the direction of extension of the pet's torso, which is obtained by fitting the three-dimensional spatial position of key points related to the spine. It is used to characterize the overall posture and deflection of the torso.

[0056] Symmetrical bone segments refer to spatial lines or corresponding bone segments formed by corresponding key points on the left and right sides of a pet's body, used to characterize the symmetry and spread-out state of the two sides of the body.

[0057] The height distribution of key points refers to the height relationship and changes of each key point relative to the reference plane of the bed surface, which is used to characterize the undulation state and the degree of contact with the bed of different parts of the pet's body.

[0058] Furthermore, spinal key points corresponding to the neck, chest, back, waist, and pelvis are extracted from the 3D key point sequence. Based on their spatial distribution, the spinal axis representing the extension direction of the pet's torso is fitted, and the projection direction, deflection angle, and straightness of the spinal axis relative to the bed reference plane are calculated. Simultaneously, the spatial correspondence between pairs of key points such as the left and right shoulders, left and right hips, left and right forelimbs, and left and right hindlimbs are extracted, and corresponding lines, angles, distances, and relative height relationships are constructed to form symmetrical bone segments representing body symmetry and limb extension state. Then, the vertical distance from each key point to the bed reference plane is calculated to obtain the relative height value, height difference, and height change relationship of each key part, forming the key point height distribution. The spinal axis, symmetrical bone segments, and key point height distribution are jointly analyzed to extract pose features such as axis direction, symmetry, extension, bed contact degree, and local undulations, obtaining 3D pose geometric data.

[0059] S3.2 Match the three-dimensional pose geometry data, imaging viewpoint labels and examination position templates, and perform condition mapping in combination with clinically acceptable thresholds to obtain pose threshold constraint data.

[0060] It should be noted that the examination position template refers to the standard positioning requirements that are pre-set for different radiological examination positions. These usually include the direction of the spine, the symmetry of the body, the extension of the limbs, and the spatial positional relationship of key points, which are used as the basis for positioning assessment.

[0061] Clinically permissible thresholds refer to the allowable range of various postural deviations set according to clinical examination requirements, used to limit the acceptable degree of spinal rotation, body tilt, symmetrical deviation, and height difference of key points.

[0062] Furthermore, the three-dimensional pose geometry data is associated with imaging viewpoint labels. Based on the imaging direction corresponding to the current image frame, the pet's body orientation, and the left-right relationship, the examination position category to which the current pet's position belongs is determined, and the standard pose constraint item corresponding to this examination position category is retrieved from the examination position template. Geometric feature parameters corresponding to the standard pose constraint item are extracted from the three-dimensional pose geometry data, including the direction of the spinal axis, the relationship between the head and pelvis, the symmetry deviation of the left and right shoulder and hip key points, the relative unfolding angle of the limb key points, the height difference of the chest and abdomen key points, and the offset of the body midline relative to the bed coordinate system. Each geometric feature parameter is then associated with the examination position. The target requirements and corresponding clinically permissible thresholds in the lookup template are compared item by item. When the deviation of a certain geometric feature parameter is less than 80% of the corresponding clinically permissible threshold, it is determined to be satisfied. When the deviation of a certain geometric feature parameter reaches 80% of the corresponding clinically permissible threshold but does not exceed the clinically permissible threshold, it is determined to be close to exceeding the limit. When the deviation of a certain geometric feature parameter exceeds the corresponding clinically permissible threshold, it is determined to be outside the permissible range. Based on this, a conditional mapping relationship is established for each geometric feature under the conditions of satisfying, being close to exceeding the limit, or exceeding the permissible range, forming pose threshold constraint data that includes constraint type, target range, deviation amount, and judgment conditions.

[0063] S3.3. Based on the posture threshold constraint data, the body position deviation, trunk tilt, body symmetry and spinal curvature are quantitatively calculated to obtain a multi-dimensional posture quantification parameter set. The multi-dimensional posture quantification parameter set is then jointly judged with the bed surface reference relationship and deviation rules to obtain the posture quantification feature vector.

[0064] It should be noted that the pose threshold constraint data refers to the data obtained from three-dimensional pose geometry data, imaging viewpoint labels, examination position templates, and clinically permissible thresholds, including the examination position requirements, permissible ranges, and deviation judgment conditions corresponding to body position deviation, trunk tilt, body symmetry, and spinal curvature.

[0065] The bed surface reference relationship refers to the positional and directional relationship between the multidimensional pose quantization parameter set and the bed surface reference plane, which is used to characterize the spatial posture of various parts of the pet's body relative to the examination bed.

[0066] Deviation rules refer to the rule conditions used to determine the type, direction and state of various pose deviations, and are used to convert a multidimensional pose quantization parameter set into a discriminable pose feature result.

[0067] Furthermore, based on the pose threshold constraint data, geometric features corresponding to body position deviation, trunk tilt, body symmetry, and spinal curvature are extracted from the three-dimensional pose geometric data. Each geometric feature is then compared and quantified with the allowable range in the pose threshold constraint data to determine the degree of deviation of each geometric feature from the corresponding inspection position requirements, thus obtaining a multi-dimensional pose quantification parameter set.

[0068] The formula for calculating spinal curvature is as follows: ; in, Indicates the curvature of the spine, used to characterize the overall degree of curvature of the spine; A sequential index representing the key points of the spine; Indicates the total number of key points in the spine; Indicates by the first The three-dimensional coordinates of the first spinal key point and the first A spinal segment vector composed of the three-dimensional coordinates of key points in the spine; Represents the vector of the next adjacent spinal segment; This represents the inverse cosine function, used to convert the angular relationship between two spinal segment vectors into an angle value.

[0069] After obtaining the multidimensional pose quantization parameter set, the multidimensional pose quantization parameter set is mapped to the bed surface reference relationship to determine the directional and spatial relationships of each deviation relative to the bed surface reference plane. Then, the multidimensional pose quantization parameter set is jointly judged with the deviation rules to identify the deviation attributes, deviation directions, and deviation states corresponding to body position deviation, trunk tilt, body symmetry, and spinal curvature, respectively, and organized in a unified order to obtain the pose quantization feature vector.

[0070] S3.4 Extract the deviation of each dimension from the pose quantization feature vector, and map it with the examination position template and the clinically acceptable range to obtain the deviation input stream.

[0071] Furthermore, the deviations of each dimension corresponding to body position offset, trunk tilt, body symmetry and spinal curvature are extracted from the pose quantification feature vector. Then, the deviations of each dimension are mapped to the target positioning requirements and clinically permissible intervals in the examination position template item by item to determine the deviation type, deviation direction and deviation interval position corresponding to each dimension deviation. The deviations are then organized in a uniform order to obtain the deviation input stream.

[0072] S3.5. Quantize the deviation from the input stream, the severe deviation cutoff factor, and the segmentation penalty rule to obtain a single-dimensional quantization parameter sequence. Based on the single-dimensional quantization parameter sequence, separate and integrate the deviation components to generate a multi-dimensional feature parameter set for the placement.

[0073] It should be noted that the deviation input stream refers to the data sequence formed by arranging the deviations of each dimension in a uniform order. The deviation input stream includes the deviation type, deviation direction, and deviation interval position corresponding to body position offset, trunk tilt, body symmetry, and spinal curvature.

[0074] The severe deviation cutoff ratio and segmentation penalty rules can be set based on clinical sample statistics, expert experience calibration, or template library presets. The severe deviation cutoff ratio is usually set to 1.5 to 3 times the corresponding clinically permissible threshold.

[0075] The segmented penalty rules set the penalty intensity according to the interval in which the deviation is located relative to the clinically permissible threshold.

[0076] Furthermore, the deviations of each dimension in the input stream are processed in correspondence with the severe deviation cutoff factor, and the deviations exceeding the severe deviation range are truncated. Then, the deviations of each dimension are solved in segments according to the segmented penalty rule to obtain the single-dimensional quantization parameters corresponding to the deviations of each dimension, and the single-dimensional quantization parameter sequence is organized in the original order.

[0077] The single-dimensional quantitative parameters corresponding to body position deviation, trunk tilt, body symmetry and spinal curvature in the single-dimensional quantitative parameter sequence are identified respectively. The deviation direction and deviation severity of each single-dimensional quantitative parameter are retained. Then, the single-dimensional quantitative parameters are separated according to the pose category to which the deviation component belongs. Finally, the single-dimensional quantitative parameters are integrated according to the positioning evaluation dimension corresponding to the examination position template to generate a multi-dimensional feature parameter set of positioning.

[0078] S4. Based on the scoring dimension and corresponding deviation direction in the multidimensional feature parameter set of the placement, match the correction rule library to obtain the parameterized adjustment amount, and perform mapping and transformation on the parameterized adjustment amount to obtain the real-time correction instruction sequence.

[0079] S4.1 Extract scoring dimensions, deviation direction and severity information from the multidimensional feature parameter set of the positioning, and perform hierarchical matching in combination with the correction rule map to obtain candidate correction rules.

[0080] Furthermore, the deviation direction and severity information corresponding to each scoring dimension are extracted from the multidimensional feature parameter set of the positioning, and hierarchical matching is performed with the applicable conditions and adjustment objects in the correction rule map to screen out candidate correction rules corresponding to the current positioning deviation state.

[0081] It should be noted that the candidate correction rules are used to provide a basis for determining the specific adjustment location, direction, and magnitude in the subsequent process.

[0082] S4.2 Perform directional analysis and continuous solution on the candidate correction rules, and filter them in combination with safety constraints to obtain parameterized adjustment amounts.

[0083] Furthermore, the adjustment target, the site of action, and the direction of adjustment corresponding to each candidate correction rule are analyzed to determine the corresponding correction action. Then, combined with the deviation direction, deviation amount, and severity in the multidimensional feature parameter set of the positioning, the adjustment range of each correction action is continuously calculated to form the parameter values ​​corresponding to the rotation angle, translation distance, or limb extension range. Then, based on safety constraints, the adjustment results that exceed the preset rotation angle range, translation distance range, or limb extension range, as well as those with action conflicts, adjustment risks, or clinically unexecutable situations, are screened out to obtain the parameterized adjustment amount.

[0084] It should be noted that safety constraints refer to the constraints that limit the adjustment range and execution boundaries of the corrective action. These typically include acceptable rotation angle range, translation distance range, limb extension range, and clinical risk boundaries to avoid causing discomfort to the pet, secondary displacement, or affecting subsequent radiological examinations. These constraints are used to screen the safety and feasibility of the corrective action.

[0085] S4.3 Extract the target body part, movement direction and severity information from the parameterized adjustment amount, and perform mapping transformation based on the action semantic map to obtain action semantic information.

[0086] Furthermore, the target body part, action direction, and severity information corresponding to each adjustment item are extracted from the parameterized adjustment amount, and mapped and transformed with the action category and semantic description in the action semantic graph to obtain action semantic information used to characterize the adjustment body part, adjustment direction, and adjustment degree.

[0087] S4.4 Prioritize and instantiate the action semantic information, and filter it in conjunction with historical instruction constraints to obtain a real-time correction instruction sequence.

[0088] Furthermore, the various adjustments in the action semantic information are prioritized according to the severity of the deviation and the order of execution, and the action semantic information is converted into explicit language instructions. Then, combined with historical instruction constraints, duplicate or conflicting language instructions are filtered out to obtain a real-time correction instruction sequence.

[0089] S5. Perform closed-loop adjustment of the pet's position according to the real-time correction instruction sequence, and obtain an automatic evaluation record when the target position template requirements are met for multiple consecutive frames.

[0090] S5.1. Perform closed-loop retesting of the pet's positioning status based on the real-time correction instruction sequence, and combine the target positioning template to jointly track residual changes, directional consistency, and cross-frame stability to obtain closed-loop execution information.

[0091] It should be noted that the target positioning template refers to the target positioning standard corresponding to the current radiological examination position. It usually includes the direction of the spine, the body symmetry, the position of the limbs, and the spatial position relationship of key points, and is used to determine whether the pet's positioning meets the expected requirements.

[0092] Furthermore, the current position of the pet is re-acquired and evaluated based on the real-time correction instruction sequence to obtain the corrected pose result. Then, combined with the target positioning template, the changes in the deviation residuals before and after correction, whether the adjustment direction is consistent, and the stability of the positioning in continuous image frames are jointly tracked to obtain closed-loop execution information.

[0093] S5.2. Converge and confirm the closed-loop execution information and determine the compliance of multiple consecutive frames, and record and solidify the key frames and adjustment history to generate automatic evaluation records.

[0094] Furthermore, convergence confirmation is performed based on the direction, magnitude, and continuous frame changes of various placement deviations in the closed-loop execution information. It is determined whether each placement deviation continues to decrease and stably approaches the target placement template's corresponding compliance range. Then, the current placement status is evaluated for compliance across multiple consecutive frames to confirm whether the compliance conditions are continuously met within a preset number of frames. Combining the key frame information before and after compliance with the corresponding adjustment history, the placement optimization process, compliance results, and time sequence are recorded and solidified to generate an automatic evaluation record.

[0095] It should be noted that the frame rate can be set according to the image sampling frame rate, the pet's micro-movement characteristics and clinical stability requirements. It is used to characterize the minimum number of frames that the current positioning state needs to be maintained continuously to meet the standard. In some implementations, it can be preset to 3 to 10 frames, preferably 5 consecutive frames.

[0096] This embodiment also provides an automatic evaluation system for pet radiographic positioning based on three-dimensional key points, including: an RGB-D acquisition module for acquiring color and depth images of the pet, registering the depth images to the color image coordinate system and filtering out invalid depth pixels to obtain an RGB-D image frame sequence; a key point reconstruction module for identifying two-dimensional key points of the pet based on the RGB-D image frame sequence, performing pinhole back-projection to reconstruct three-dimensional coordinates using camera intrinsic parameters and neighborhood robust depth values, determining the imaging angle based on the cross product sign of the spine and symmetrical key points, and using a random consistency algorithm to fit the examination bed reference plane for spatial coordinate system transformation to obtain a three-dimensional key point sequence; and a pose evaluation module for evaluating the position from three-dimensional... The system extracts the pet's three-dimensional pose geometric features from the keypoint sequence and the bed reference plane. These features are then mapped and compared with clinically acceptable thresholds to obtain a quantified pose feature vector. This vector is then quantified using a severe deviation truncation and segmented penalty mechanism to generate a multi-dimensional feature parameter set for positioning. A correction instruction module matches the scoring dimension and corresponding deviation direction from the multi-dimensional feature parameter set to a correction rule library, obtains parameterized adjustment amounts, and maps these adjustments to generate a real-time correction instruction sequence. An automatic evaluation and recording module performs closed-loop adjustments to the pet's positioning based on the real-time correction instruction sequence and generates an automatic evaluation record when the target positioning template requirements are met for multiple consecutive frames.

[0097] In summary, this invention, based on image recognition and combined with RGB-D data acquisition, two-dimensional key point detection, pinhole back-projection three-dimensional reconstruction, and bed reference plane coordinate transformation, enables three-dimensional spatial representation of key anatomical sites during pet radiographic positioning. Furthermore, through pose geometric feature extraction, clinical threshold mapping, deviation quantification analysis, and correction rule matching, it achieves automatic assessment of pet positioning status, deviation direction determination, and parameterized correction guidance. Finally, by combining closed-loop retesting and continuous multi-frame compliance determination, it automatically generates assessment records, thereby improving the objectivity, standardization, real-time performance, and intelligence of pet radiographic positioning assessment.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic evaluation method for pet radial positioning based on three-dimensional key points, characterized in that, include: Acquire color and depth images of the pet, register the depth images to the color image coordinate system, and filter out invalid depth pixels to obtain an RGB-D image frame sequence; Two-dimensional key points of pets are identified based on RGB-D image frame sequences. Three-dimensional coordinates are reconstructed by pinhole back projection combined with camera intrinsic parameters and neighborhood robust depth values. The imaging angle is determined based on the cross product sign of the spine and symmetrical key points. The spatial coordinate system is transformed by fitting the examination bed reference plane using the random consistency algorithm to obtain the three-dimensional key point sequence. The three-dimensional pose geometric features of the pet are extracted from the three-dimensional key point sequence and the bed reference plane. The three-dimensional pose geometric features are mapped and compared with the clinically permissible threshold to obtain the pose quantification feature vector. The quantification is then performed by combining severe deviation truncation and segmented penalty mechanism to generate a multi-dimensional feature parameter set for positioning. Based on the scoring dimension and corresponding deviation direction in the multidimensional feature parameter set of the placement, a correction rule library is matched to obtain the parameterized adjustment amount, and the parameterized adjustment amount is mapped and transformed to obtain the real-time correction instruction sequence. The pet's positioning is adjusted in a closed loop according to the real-time correction command sequence, and an automatic evaluation record is obtained when the target positioning template requirements are met for multiple consecutive frames.

2. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The specific steps for obtaining the RGB-D image frame sequence are as follows: Collect color images, depth images, and time synchronization information of the pet; perform pixel-level registration based on the coordinate projection relationship from the depth image to the color image to obtain aligned dual-modal image data. Using time synchronization information, spatial consistency, temporal continuity, and geometric closed-loop constraint analysis are performed on the aligned bimodal image data, and invalid depth pixels are filtered out to obtain an RGB-D image frame sequence.

3. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1 or 2, characterized in that, The method involves identifying two-dimensional key points of a pet based on RGB-D image frame sequences, performing pinhole backprojection to reconstruct three-dimensional coordinates using camera intrinsic parameters and robust neighborhood depth values, and determining the imaging perspective based on the cross product sign of the spine and symmetrical key points. The specific steps are as follows: Pet keypoint candidate locations, local depth structure information, and cross-frame trajectory information are extracted from RGB-D image frame sequences and jointly organized to obtain a multimodal keypoint candidate field; By performing topological constraints, temporal consistency calibration, and local depth continuity screening on the multimodal keypoint candidate field, stable two-dimensional keypoint data is obtained. Based on stable two-dimensional keypoint data, neighborhood depth backtracking and pinhole backprojection calculations are performed to obtain three-dimensional keypoint skeleton data. The three-dimensional keypoint skeleton data is then jointly discriminated with the cross product signs corresponding to the spinal principal axis and symmetrical keypoints to obtain imaging viewpoint label data.

4. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The specific steps for obtaining the three-dimensional keypoint sequence are as follows: Animal occupancy and exclusion information and visible area information of the bed surface are extracted from 3D keypoint skeleton data and imaging viewpoint label data, and then cropped and filtered in combination with depth point cloud to obtain a set of candidate points on the bed surface; The candidate point set of the bed surface, the point-level planarity prior, and the historical bed surface memory are sampled and scored for quality ranking to obtain the bed surface master hypothesis data. Normal vector disambiguation and spatial coordinate system construction were performed on the master hypothesis data of the bed surface, and coordinate transformation was performed on the 3D key point skeleton data of the pet to obtain the 3D key point sequence.

5. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The specific steps for obtaining the pose quantization feature vector are as follows: The spinal principal axis, symmetrical bone segments, and key point height distribution are extracted from the three-dimensional key point sequence and the bed reference plane, and joint analysis is performed to obtain three-dimensional pose geometric data. The three-dimensional pose geometry data, imaging viewpoint labels and examination position templates are matched and combined with clinically acceptable thresholds for conditional mapping to obtain pose threshold constraint data. Based on the pose threshold constraint data, the body position deviation, trunk tilt, body symmetry and spinal curvature are quantitatively calculated to obtain a multi-dimensional pose quantification parameter set. The multi-dimensional pose quantification parameter set is then jointly judged with the bed surface reference relationship and deviation rules to obtain the pose quantification feature vector.

6. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The specific steps for generating the multidimensional feature parameter set of the placement are as follows: The deviation of each dimension is extracted from the pose quantization feature vector and mapped with the examination position template and the clinically acceptable range to obtain the deviation input stream; The deviation from the input stream, the severe deviation truncation factor, and the segmentation penalty rule are quantized to obtain a single-dimensional quantization parameter sequence. Based on the single-dimensional quantization parameter sequence, the deviation components are separated and integrated to generate a multi-dimensional feature parameter set for the placement.

7. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1 or 6, characterized in that, The specific steps for obtaining the parameterized adjustment amount are as follows: Scoring dimensions, deviation direction, and severity information are extracted from the multidimensional feature parameters of the positioning, and hierarchical matching is performed in combination with the correction rule graph to obtain candidate correction rules; Candidate correction rules are analyzed and continuously solved, and then screened in conjunction with safety constraints to obtain parameterized adjustment amounts.

8. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The specific steps for obtaining the real-time correction instruction sequence are as follows: The target body part, direction of motion, and severity information are extracted from the parameterized adjustment values, and then mapped and transformed according to the action semantic map to obtain action semantic information. The semantic information of the actions is prioritized and instantiated in language, and then filtered in combination with historical instruction constraints to obtain a real-time correction instruction sequence.

9. The automatic evaluation method for pet radial positioning based on three-dimensional key points as described in claim 1, characterized in that, The automatic evaluation record is obtained through the following steps: The pet's positioning status is retested in a closed loop based on the real-time correction instruction sequence, and the residual change, directional consistency and cross-frame stability are jointly tracked in conjunction with the target positioning template to obtain closed-loop execution information. The closed-loop execution information is converged and confirmed, and multiple consecutive frames are judged to meet the standards. The key frames and adjustment history are recorded and solidified to generate automatic evaluation records.

10. An automatic evaluation system for pet radiation positioning based on three-dimensional key points, based on the automatic evaluation method for pet radiation positioning based on three-dimensional key points according to any one of claims 1 to 9, characterized in that, include: The RGB-D acquisition module is used to acquire color and depth images of the pet, register the depth image to the color image coordinate system, and filter out invalid depth pixels to obtain an RGB-D image frame sequence. The key point reconstruction module is used to identify the two-dimensional key points of pets based on RGB-D image frame sequences. It combines camera intrinsic parameters and neighborhood robust depth values ​​to perform pinhole back projection to reconstruct the three-dimensional coordinates. It also determines the imaging angle based on the cross product sign of the spine and symmetrical key points, and uses a random consistency algorithm to fit the examination bed reference plane to perform spatial coordinate system transformation to obtain the three-dimensional key point sequence. The pose assessment module is used to extract the three-dimensional pose geometric features of the pet from the three-dimensional key point sequence and the bed reference plane. The three-dimensional pose geometric features are mapped and compared with the clinically acceptable threshold to obtain the pose quantification feature vector. The quantification is then performed by combining severe deviation truncation and segmented penalty mechanism to generate a multi-dimensional feature parameter set for positioning. The deviation correction instruction module is used to match the deviation correction rule library according to the scoring dimension and corresponding deviation direction in the multidimensional feature parameter set of the placement, obtain the parameterized adjustment amount, and perform mapping and transformation on the parameterized adjustment amount to obtain the real-time deviation correction instruction sequence. The automatic evaluation and recording module is used to perform closed-loop adjustments to the pet's positioning based on the real-time correction instruction sequence, and to obtain an automatic evaluation record when the target positioning template requirements are met for multiple consecutive frames.

Images