Hand-held bedridden patient 3D optical scanning weighing method and device

Abstract

The invention discloses a handheld bedridden patient 3D optical scanning weighing method and device. The method comprises the steps that three-dimensional point cloud data of the body surface of a patient are collected; performing preprocessing on the three-dimensional point cloud data; extracting geometric features of the three-dimensional point cloud data; partitioning based on geometric features of the three-dimensional point cloud data, and constructing a three-dimensional human body model of the patient; and predicting the three-dimensional human body model based on a pre-trained deep learning model to obtain the weight of the patient. Through the treatment scheme disclosed by the invention, the body weight of the patient can be simply and quickly measured under the condition of not interfering the patient.

Description

Technical Field

[0001] The present invention relates to the technical field of body weight estimation, and particularly to a handheld 3D optical scanning weighing method and device for bedridden patients. Background Art

[0002] In medical institutions, body weight is one of the key vital signs, especially for critically ill patients. The body weight data plays an indispensable role in aspects such as drug dosage adjustment, formulation of nutritional support plans, and fluid management. However, due to severe illness, immobility, or critical condition, many bedridden critically ill patients are unable to measure their body weight in the traditional standing manner.

[0003] In the current existing technologies, there are many problems with some weighing devices for measuring the body weight of bedridden patients. On the one hand, the accuracy of some devices is poor, and the measurement results have large errors. For critically ill patients who need accurate body weight data to adjust drug dosages, plan nutritional support, and control fluid management, this is very likely to affect the accuracy of the treatment plan and thus have an adverse impact on the treatment effect of the patients. On the other hand, the structure of some devices is too complex, with large volume and heavy weight. Not only is it extremely inconvenient to carry and place, but it also occupies a large amount of space. Moreover, the operation process is cumbersome, and it requires professional personnel to spend a lot of time and energy to operate. In the time-critical medical scenario, this will undoubtedly reduce work efficiency and is difficult to meet the actual emergency measurement needs.

[0004] Therefore, it is extremely urgent to develop a portable, accurate, and easy-to-operate device to meet the body weight measurement needs of bedridden patients. The handheld 3D optical scanning weighing method and device of the present invention are precisely born to solve these deficiencies of the existing technologies. Summary of the Invention

[0005] In view of this, embodiments of the present disclosure provide a handheld 3D optical scanning weighing method for bedridden patients, which at least partially solves the problems existing in the prior art.

[0006] In a first aspect, embodiments of the present disclosure provide a handheld 3D optical scanning weighing method for bedridden patients, the method comprising the following steps: Collect three-dimensional point cloud data of the patient's body surface; Preprocess the three-dimensional point cloud data; Extract geometric features of the three-dimensional point cloud data; Partition based on the geometric features of the three-dimensional point cloud data and construct a three-dimensional human model of the patient; Predict the body weight of the patient based on the pre-trained deep learning model for the three-dimensional human model.

[0007] According to a specific implementation manner of an embodiment of the present disclosure, the preprocessing of the acquired three-dimensional point cloud data includes at least one of data filtering, noise reduction, feature extraction, point cloud downsampling, voxel downsampling, uniform downsampling, and upsampling.

[0008] According to a specific implementation manner of an embodiment of the present disclosure, the preprocessing of the acquired three-dimensional point cloud data further includes: point cloud registration; The point cloud registration includes: Finding matching point pairs through random sample consensus based on the geometric features of the three-dimensional point cloud data; Calculating the squared error of all matching point pairs and sorting them from smallest to largest; Using the preset number of smallest error values to calculate the optimal rotation and translation matrices and updating the source point cloud; Repeating the above steps until the convergence condition is met.

[0009] According to a specific implementation manner of an embodiment of the present disclosure, the extraction of the geometric features of the point cloud includes: Calculating the normal and curvature of the point cloud; Extracting the minimum bounding box or convex hull of the point cloud.

[0010] According to a specific implementation manner of an embodiment of the present disclosure, calculating the curvature of the point cloud includes: Selecting a point cloud set N(p) within a first radius centered on the first point cloud; Determining the fitting plane parameters by minimizing the error function: ; Where x i , y i , z i are points within the point cloud set N(p); , , , are plane parameters; Calculating the curvature based on the following formula: ; Where is the distance from the point within N(p) to the fitting plane; is the number of points within the neighborhood.

[0011] According to a specific implementation manner of an embodiment of the present disclosure, partitioning based on the geometric features of the three-dimensional point cloud data and constructing a three-dimensional human model of the patient includes: Partition the key parts of the human body based on the geometric features of 3D point cloud data; among them, the key parts of the human body include: head, neck, chest, abdomen, upper arm, forearm, thigh, calf. Perform semantic segmentation on the point cloud data based on a pre-trained deep learning model, and directly output the segmentation results of each part of the human body.

[0012] According to a specific implementation manner of the embodiments of the present disclosure, the method further includes: Transmit the prediction result to the visualization platform through a wireless transmission method.

[0013] In a second aspect, the embodiments of the present disclosure provide a handheld 3D optical scanning and weighing device for bedridden patients, and the device includes: A data acquisition module configured to acquire 3D point cloud data of the patient's body surface; A data processing module configured to preprocess the 3D point cloud data; A feature extraction module configured to extract the geometric features of the 3D point cloud data; A model construction module configured to partition based on the geometric features of the 3D point cloud data and construct a 3D human body model of the patient; A prediction module configured to predict the 3D human body model based on a pre-trained deep learning model to obtain the weight of the patient.

[0014] According to a specific implementation manner of the embodiments of the present disclosure, the device further includes: a visualization display module configured to display patient data.

[0015] The handheld 3D optical scanning and weighing method in the embodiments of the present disclosure scans the 3D physical data of bedridden patients, combines a deep learning model and a big data weight estimation model to partition and estimate the weight of the patients, accurately calculates the weight of the patients, and avoids the problem of difficult weight measurement caused by the patients being unable to stand or get out of bed. Description of the Drawings

[0016] The above is only an overview of the technical solutions of the present invention. In order to be able to understand the technical means of the present invention more clearly, the following further describes the present invention in detail with reference to the drawings and specific embodiments.

[0017] Figure 1 It is a schematic flowchart of a handheld 3D optical scanning and weighing method provided by the embodiments of the present disclosure; Figure 2 It is a schematic structural diagram of a handheld 3D optical scanning and weighing device provided by the embodiments of the present disclosure. Detailed Embodiments

[0018] The following describes the embodiments of the present disclosure in detail with reference to the accompanying drawings.

[0019] The following illustrates the implementation manners of the present disclosure through specific specific examples. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0020] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device and / or this method can be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0021] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0022] In the medical field, accurate measurement of body weight is crucial for patient care and treatment plan formulation, especially for critically ill bedridden patients. Since these patients cannot stand or get out of bed by themselves, traditional methods of body weight measurement are often not applicable. Therefore, the proposed handheld 3D optical scanning and weighing method for bedridden patients calculates the body weight of patients by partitioning based on the 3D body shape data of the patients to achieve high-precision body weight estimation.

[0023] Figure 1 It is a schematic diagram of the process of the handheld 3D optical scanning and weighing method for bedridden patients provided for the embodiments of the present disclosure.

[0024] As Figure 1 shown, at step S110, three-dimensional point cloud data of the patient's body surface is collected.

[0025] More specifically, next, proceed to step S120.

[0026] At step S120, preprocess the three-dimensional point cloud data.

[0027] In an embodiment of the present invention, the preprocessing of the collected three-dimensional point cloud data includes at least one of data filtering, noise reduction, feature extraction, point cloud downsampling, voxel downsampling, uniform downsampling, and upsampling.

[0028] In an embodiment of the present invention, the preprocessing of the collected three-dimensional point cloud data further includes: point cloud registration; the point cloud registration includes: finding matching point pairs through random sample consensus based on the geometric features of the three-dimensional point cloud data; calculating the squared error of all matching point pairs and sorting them from smallest to largest; using the pre-set number of smallest error values to calculate the optimal rotation and translation matrices and updating the source point cloud; repeating the above steps until the convergence condition is met.

[0029] The convergence condition of point cloud registration is defined as: the change in registration error for 3 consecutive iterations < 0.008 cm; the maximum number of iterations is limited to 150 times.

[0030] Next, proceed to step S130.

[0031] At step S130, extract the geometric features of the three-dimensional point cloud data.

[0032] In an embodiment of the present invention, the extraction of the geometric features of the point cloud includes: calculating the normal and curvature of the point cloud; extracting the minimum bounding box or convex hull of the point cloud.

[0033] In an embodiment of the present invention, calculating the curvature of the point cloud includes: Select a point cloud set N(p) within the first radius centered on the first point cloud; Determine the fitting plane parameters by minimizing the error function: ; where x i ,, y i ,, z i are points within the point cloud set N(p); , , , are plane parameters; Calculate the curvature based on the following formula: ; where is the distance from the point within N(p) to the fitting plane; is the number of points in the neighborhood.

[0034] Next, go to step S140.

[0035] At step S140, partition based on the geometric features of the three-dimensional point cloud data to construct a three-dimensional human model of the patient.

[0036] More specifically, In an embodiment of the present invention, the partitioning based on the geometric features of the three-dimensional point cloud data to construct a three-dimensional human model of the patient includes: dividing the key parts of the human body based on the geometric features of the three-dimensional point cloud data; wherein, the key parts of the human body include: head, neck, chest, abdomen, upper arm, forearm, thigh, calf; performing semantic segmentation on the point cloud data based on a pre-trained deep learning model to directly output the segmentation results of each part of the human body.

[0037] Next, go to step S150.

[0038] At step S150, perform prediction on the three-dimensional human model based on a pre-trained deep learning model to obtain the weight of the patient.

[0039] In an embodiment of the present invention, the method further includes: transmitting the prediction result to a visualization platform through a wireless transmission method.

[0040] More specifically, the present invention will be described below through embodiments.

[0041] I. Collect three-dimensional point cloud data of the patient's body surface Medical staff hold the 3D optical scanning weighing device proposed by the present invention and slowly and evenly move around the bedridden patient to ensure full coverage of the patient's body and collect data from different angles. The 3D optical scanning weighing device emits light signals of a specific wavelength. After the light signals contact the patient's body surface and are reflected back, the reflected light is received by the built-in sensor, and based on technologies such as the triangulation principle or the time-of-flight method, the three-dimensional coordinate information of each point in space is accurately calculated, thereby generating the original three-dimensional point cloud data. For example, when scanning an elderly patient who has been bedridden due to a fracture, the medical staff start from above the patient's head and scan the face, neck, chest, abdomen, limbs, and back along the body's central axis in sequence. The entire scanning process lasts about 3 - 5 minutes to obtain sufficiently fine and complete point cloud data. The collected point cloud data contains millions of discrete points, and each point carries its coordinate information (X, Y, Z) in the three-dimensional space.

[0042] II. Preprocess the collected three-dimensional point cloud data.

[0043] Data filtering: The Gaussian filtering algorithm is used to remove the noise points caused by factors such as ambient light interference and slight jitter of the 3D optical scanning weighing device. For example, for some points that deviate from the average position of the surrounding point cloud by more than a certain threshold (such as 0.5 cm) and have abnormal gray values, they are determined as noise points and removed. After filtering, the proportion of noise points in the original point cloud data is reduced from about 5% to less than 1%.

[0044] Noise reduction: The bilateral filter is used to further smooth the point cloud surface, reducing high-frequency noise while maintaining edge features. Taking the patient's arm part as an example, after bilateral filtering, the fine wrinkles and unevenness on the arm surface are moderately smoothed, making the data more regular during subsequent processing, and at the same time not affecting the recognition of the arm's contour features.

[0045] Feature extraction: Calculate the local feature descriptors of the point cloud, such as the Fast Point Feature Histogram (FPFH). For the point cloud of the patient's torso part, by calculating the FPFH, the geometric shape information of the local area of the point cloud can be effectively captured, providing a key basis for subsequent partitioning and model construction. These feature descriptors can better distinguish different-shaped body parts, such as distinguishing the rounded shoulder and the relatively flat back area.

[0046] Point cloud downsampling: The voxel grid downsampling method is used to divide the three-dimensional space into small voxel units (cubes with a side length of 0.2 cm), and only one representative point is retained in each voxel. For parts such as the legs with a large amount of data and relatively regular shapes, the number of points in the point cloud is reduced by about 70% after downsampling, greatly reducing the computational load of subsequent processing, while still retaining the main geometric features of the legs, such as the cylindrical shape of the thigh and the approximate prismatic shape of the calf.

[0047] Voxel downsampling: This is a more refined voxel-based downsampling method. According to the distribution of points in the voxel, the centroid method or the random method is used to select representative points. When processing the point cloud of the complex shape of the head, voxel downsampling can not only reduce data redundancy but also ensure that the key features of the head, such as the outlines of facial features and the shape of the skull, are retained. The downsampling ratio can be adjusted according to the actual situation, generally around 50%-60%.

[0048] Uniform downsampling: Points are selected on the point cloud at a fixed interval (every 0.3 cm) to ensure the uniformity of the point cloud distribution. When scanning the patient's back, uniform downsampling makes the back point cloud easier to analyze during subsequent processing. Especially when performing operations such as plane fitting, it can provide a more stable point cloud basis and reduce errors caused by uneven point cloud density.

[0049] Upsampling: When it is found that the point cloud in some key parts (such as finger joints) is too sparse, which may affect the accuracy of the subsequent model, an upsampling method based on radial basis function (RBF) is adopted. By using the information of known points around the sparse area, the points that need to be supplemented in the middle are deduced, so that the point cloud at the finger joints is denser, and the bending shape of the joints can be clearly presented, improving the restoration degree of the model to fine parts.

[0050] Point cloud registration: Based on the geometric features of the three-dimensional point cloud data, matching point pairs are found through random sample consensus: First, a certain number (such as 500 pairs) of point pairs are randomly selected from the source point cloud and the target point cloud (for example, the partial point clouds scanned from both sides of the patient's body), and the geometric feature similarities between them are calculated, such as the distance between two points, the normal angle, etc. Through multiple random samplings (usually more than 1000 times), enough inliers (i.e., matching point pairs that meet certain geometric consistency conditions) are found. These inliers are likely to come from the surface of the same real body part.

[0051] Calculate the squared error of all matching point pairs and sort them from smallest to largest: For the found matching point pairs, calculate the error according to the sum of the squares of the coordinate differences, and sort these error values. For example, when registering the point clouds of the left and right legs of a patient, it is found through calculation that most of the matching point pairs located on the straight part of the legs have smaller errors, while the errors at the joints may be slightly larger due to movement or scanning angle differences.

[0052] Use a preset number (such as 100 pairs) of the smallest error values to calculate the optimal rotation and translation matrices, and update the source point cloud: Using these 100 pairs of matching point pairs with the smallest errors, the rotation and translation parameters required to accurately align the source point cloud to the target point cloud are solved through the singular value decomposition (SVD) method, obtaining the rotation matrix R and the translation vector T. Then, the source point cloud is transformed according to the formula P_new = R * P_old + T to achieve accurate registration of the point cloud. Repeat the above steps until the convergence condition is met. The convergence condition can be set that the change in the point cloud registration error is less than a certain threshold (such as 0.01 cm) after two consecutive iterations, ensuring that the point cloud is accurately aligned globally and presenting the three-dimensional shape of the patient's body completely.

[0053] III. Extract the geometric features of the point cloud.

[0054] Calculate the normal and curvature of the point cloud: The normal vector is calculated using the local plane fitting method. For the patient's chest point cloud, taking each point as the center, points within a certain neighborhood (such as points within a radius of 0.5 cm) are selected for plane fitting. The plane equation is solved by the least squares method, and the normal vector of the plane is the normal of that point. The curvature is approximately obtained by calculating the second-order derivatives of the point cloud in different directions. In curved parts such as the shoulders, the normal direction can clearly indicate the surface orientation, and the curvature value accurately reflects the degree of curvature, providing important clues for subsequent distinguishing different body parts.

[0055] Curvature calculation and threshold setting: First, for the whole body point cloud data collected and preprocessed, taking each point as the center, a suitable neighborhood range (for example, a spherical neighborhood with a radius of 0.4 cm) is selected to calculate the curvature of that point. A method based on local quadratic surface fitting is used, and the parameters of the fitting surface are solved by the least squares method to obtain the curvature value of the point cloud.

[0056] Set the curvature threshold. Through the analysis and experiment of a large amount of human body point cloud data, it is found that the curvature in the head region (especially around the facial features) is usually relatively high. Generally, the curvature threshold is set to be greater than 0.35. This is because the contour of the head includes complex shapes such as the arc of the forehead, the depression of the eye sockets, the protrusion of the nose, and the curvature of the chin, and the point cloud curvature in these parts is significantly different from relatively flat areas such as the torso.

[0057] Normal direction analysis: At the same time, calculate the normal of the point cloud. Also based on the local neighborhood points (such as the above-mentioned 0.4 cm spherical neighborhood), the principal component analysis (PCA) method is used to fit the plane, and the normal vector of the plane is the normal of the point. For the head region, the change of the normal direction is relatively complex. In the forehead part, the normal is basically perpendicular upward; on both sides of the cheeks, the normal is inclined outward; and around the ears, the normal direction has an obvious circumferential change.

[0058] Through statistical analysis, set the normal direction change threshold. For example, if the average value of the normal angles of adjacent points (with a spacing less than 0.2 cm) is more than 30 degrees, and the curvature of this area meets the above threshold conditions, it is initially determined as a boundary point of the head region.

[0059] Region growing and refinement: Taking the boundary points that meet the curvature and normal conditions as seed points, the region growing algorithm is used to further expand the head region. During the region growing process, continuously check whether the curvature and normal of the newly added points meet the set standards. If the curvature of a certain point is less than the threshold or the change of the normal direction is abnormal, stop including it in the head region.

[0060] Perform morphological post - processing on the grown head region to remove isolated small noise point clusters (such as a small number of abnormal points caused by hair reflection, etc.). Through connected - component analysis, remove point sets with fewer than a certain number of points (such as 50) and not connected to the main head region, and finally accurately divide the head region.

[0061] The neck region division includes: Transition region search: Start searching downward from the bottom - boundary points of the determined head region. Since the neck is the transitional part connecting the head and the torso, its curvature and normal change relatively smoothly. Taking the bottom - boundary points of the head as the center, select a slightly larger neighborhood (such as a radius of 0.6 cm), calculate the curvature and normal of the points in the neighborhood. Set the curvature - change - rate threshold and the normal - change threshold. The curvature - change rate in the neck region is much smaller than that in the head, generally set to less than 0.05 curvature change per centimeter of length. The normal direction remains relatively stable in the neck region, and the average value of the included angle between adjacent - point normals is controlled within 10 degrees. By traversing the point cloud below the head, find the continuous region that meets these conditions as the candidate region for the neck.

[0062] Length and width constraints: Further screen the candidate neck regions. According to human anatomy knowledge, the neck length is usually between 6 - 12 cm (there are differences among different genders, ages, and body types). By measuring the length of the candidate region in the vertical direction, exclude regions with significantly inconsistent lengths.

[0063] The width of the neck is relatively narrow and more uniform, generally with a width of 4 - 8 cm. Calculate the minimum - bounding box of the candidate region and check whether the width is within a reasonable range. If the width is too large (such as greater than 10 cm, which may be misjudged as part of the shoulder) or too small (such as less than 3 cm, which may be an abnormal point set caused by scanning errors), make adjustments or re - search.

[0064] Final determination: Comprehensively consider factors such as curvature, normal, length, and width to determine the neck region that best meets the conditions. Smooth the determined neck region to remove a small number of jagged edge points that may be caused by scanning jitter, etc., and ensure the smooth transition of the neck region with the head and torso regions.

[0065] Torso region division: Plane feature extraction: For the remaining unpartitioned point cloud (excluding the determined areas of the head and neck), focus on large relatively flat areas, which usually correspond to the torso part. Use the plane fitting algorithm based on Random Sample Consensus (RANSAC) to randomly extract point sets from the point cloud multiple times to attempt to fit a plane. Set the inlier ratio threshold. For example, if the inlier ratio of the plane obtained in one fitting (i.e., the ratio of the points satisfying the plane equation to the extracted point set) is greater than 60%, and the area of this plane is large (judged by calculating the number of points covered by the plane, such as more than 5000 points), then it is preliminarily determined as part of the torso.

[0066] Analyze the normal direction of the fitted plane. The normal directions of the front, back, left, and right planes of the torso are relatively fixed. For example, the normal of the chest plane is basically forward, the normal of the back plane is basically backward, and the normals of the left and right side planes are respectively to the left and right sides. By setting the tolerance range of the normal direction (such as the angle with the standard direction within 20 degrees), further confirm whether the plane belongs to the torso area.

[0067] Curvature change analysis: Based on the preliminarily determined torso plane area, check the curvature change of the point cloud. Although the torso is relatively flat, there are still certain curvature changes at the edges of the chest, on both sides of the spine, etc. Set the local curvature threshold. For example, at the edge of the chest, the curvature value is between 0.1 - 0.2, for the head it is 0.35 - 0.6, and for the limbs it is 0.1 - 0.25. By traversing the points around the plane area, ensure that the curvature of these parts meets the expectations to accurately define the boundary of the torso.

[0068] Subdivide the torso area (chest, abdomen, etc.): According to the human physiological structure, use the existing curvature and normal information to further subdivide the torso. There are collarbones above the chest area and ribs below, and the point cloud curvature and normals have obvious characteristics at these parts. By detecting the arc contour of the chest (based on curvature change) and the normal distribution corresponding to the rib orientation, accurately divide the chest area from the torso.

[0069] The abdomen area is relatively flatter than the chest, with lower point cloud curvature, generally less than 0.05, and the normal direction is relatively single (basically vertically downward). Through these characteristics, combined with the relative position relationship with the chest area, accurately divide the abdomen area to complete the subdivision of the torso.

[0070] Limb area division: Slender shape detection: For the remaining unpartitioned point cloud, find the regions with slender shape features, which usually correspond to the limbs. Adopt a method based on principal component analysis (PCA) to analyze the neighborhood of each point (such as a radius of 0.5 cm). If the ratio of the length of the maximum principal component direction to the length of the second principal component direction in the neighborhood of a certain point is greater than 5 (indicating obvious elongation in one direction), and the number of consecutive points satisfying this condition in this region exceeds a certain number (such as 200), then it is regarded as a candidate region for the limbs.

[0071] Axial normal feature: The normals of the limbs in their axial directions are basically consistent. Based on the candidate limb regions, check the stability of the normal direction of the point cloud along the axis. For example, for the arm, from the shoulder to the wrist direction, the normal is basically perpendicular to the long axis direction of the arm, and the average angle between the normals of adjacent points is less than 15 degrees. The same applies to the legs. From the thigh root to the ankle, the normal direction conforms to the anatomical leg orientation of the human body. By setting the normal direction stability threshold, exclude the regions that do not meet the conditions.

[0072] Judgment of the connection part: The connection parts between the limbs and the torso are the key to the partition. At one end of the candidate limb region close to the torso, check the connection relationship with the torso region. Usually, there will be a certain change in the curvature of the point cloud at the connection part, and the normal direction will also have a transition. By setting the curvature change threshold (such as 0.08 per centimeter) and the normal transition threshold (such as the average angle between the normals of adjacent points of 20 degrees) at the connection part, accurately judge the connection points between the limbs and the torso, and finally completely partition the limb regions, including specific parts such as the upper arm, forearm, thigh, and calf.

[0073] Extract the minimum bounding box or convex hull of the point cloud: Use the library function based on the geometric algorithm to calculate the minimum bounding box. For the point cloud of the entire patient's body, find a cuboid (minimum bounding box) that can just contain all the points. The length, width, height, and orientation information of the cuboid reflect the approximate range and posture of the body in three-dimensional space. At the same time, extract the convex hull. The convex hull is the smallest set that contains all the point cloud and whose surface is a convex polygon. When processing the limb point cloud, the convex hull can concisely outline the approximate contour of the limbs, remove the internal redundant information, and facilitate subsequent partitioning operations based on geometric features.

[0074] IV. Partition based on the geometric features of the three-dimensional point cloud data to construct a three-dimensional human model of the patient.

[0075] Key parts of the human body partitioned based on the geometric features of the three-dimensional point cloud data: First, based on the curvature of the point cloud, the change in the normal vector, and geometric features such as the previously extracted minimum bounding box and convex hull, the head region is identified. The head usually has a relatively high curvature change, especially around the facial features, and the orientation and size of its minimum bounding box in space have obvious characteristics. For example, the head region is determined by setting the curvature threshold to be greater than 0.3 and the ratio of the height to the width of the minimum bounding box within a specific range (such as 0.8 - 1.2).

[0076] The neck region is divided according to its connection between the head and the torso, being relatively thin and having a gentle curvature transition. Using the connectivity analysis of the point cloud, starting from below the head, the part where the width gradually changes until it connects to the torso is found as the neck, and its length is generally about 5 - 10 centimeters (adjusted according to the body types of different patients).

[0077] The torso part is identified by a relatively large planar area, a relatively stable normal vector direction, and a relatively large minimum bounding box size. The chest and abdomen regions can be further subdivided. The chest is based on the arc characteristics of the chest cavity and the point cloud normal distribution corresponding to the rib orientation, and the abdomen is based on the relatively flat and large area, combined with the previously calculated geometric features for precise division.

[0078] The limbs are distinguished by their slender shapes, relatively stable normal vector directions along the axis, and connection relationships with the torso. For example, starting from the shoulder, along the extending direction of the arm, the point cloud presents a slender columnar shape. The range of the arm is determined by detecting features such as the ratio of the length to the width of the point cloud in the axial direction being greater than 5. Similarly, the range of the leg is determined.

[0079] Based on a pre - trained deep learning model, semantic segmentation is performed on the point cloud data, directly outputting the segmentation results of each part of the human body: A deep learning model based on a convolutional neural network (CNN) is used. It is pre - trained on a large number of 3D point cloud data of human bodies with different body types and postures, learning the point cloud feature patterns of different body parts. The pre - processed and feature - extracted point cloud data of the patient is input into this model, and the model can directly output the category label of the body part to which each point belongs, such as head, neck, chest, abdomen, upper arm, forearm, thigh, calf, etc. In practical applications, for a new bedridden patient, the model can complete the semantic segmentation of the point cloud within seconds, obtaining accurate segmentation results of each part of the human body, providing a structured model basis for subsequent weight prediction.

[0080] V. Based on a pre - trained deep learning model, predictions are made on the three - dimensional human body model to obtain the patient's weight.

[0081] Building a large database of patients' 3D body shapes and actual weights is the primary step in model development. The establishment of this database includes the following aspects: Data collection: Through cooperation with hospitals and medical institutions, collect patient data of different genders, ages, body shapes, and disease states. The data collection process needs to ensure patient privacy and data security, using anonymization processing methods. Through optical scanning devices, obtain the 3D body shape data of patients, and record important parameters such as height, body width, body length, and weight. Data annotation: Annotate the collected 3D body shape data to ensure that each data sample corresponds to its actual weight. This process requires the participation of a professional medical team to ensure the accuracy of the data.

[0082] Construct the PointNet++ model: Set Abstraction (SA) module: SA1: Set the number of sampling points (npoint) to 512, the search radius (radius) to 0.2 meters, and the number of neighborhood points (nsample) to 32. The number of input channels (in_channel) is 3 (the three-dimensional coordinates of the points), and the number of channels of the multi-layer perceptron (MLP) is set to [64, 64, 128].

[0083] SA2: The number of sampling points (npoint) is 128, the search radius (radius) is 0.4 meters, and the number of neighborhood points (nsample) is 64. The number of input channels (in_channel) is 128 + 3 (the output features of the previous layer and the three-dimensional coordinates of the points), and the number of channels of the MLP is set to [128, 128, 256].

[0084] SA3: Do not perform sampling (npoint is None), and the search radius and the number of neighborhood points are also None. The number of input channels (in_channel) is 256 + 3, and the number of channels of the MLP is set to [256, 512, 1024].

[0085] Fully connected layer: The input dimension of the first fully connected layer (fc1) is 1024, the output dimension is 512, and batch normalization (BatchNorm1d) and Dropout (dropout rate of 0.4) are used to prevent overfitting.

[0086] The input dimension of the second fully connected layer (fc2) is 512, the output dimension is 256, and batch normalization and Dropout (dropout rate of 0.4) are also used.

[0087] The input dimension of the last fully connected layer (fc3) is 256, the output dimension is 1, and it is used to predict the weight.

[0088] Define the loss function and optimizer: Loss function: The mean squared error loss (MSE Loss) is used to measure the error between the predicted weight of the model and the actual weight.

[0089] Optimizer: The Adam optimizer is selected, with the learning rate set to 0.001 and the weight decay set to 0.0001.

[0090] The number of training epochs (num_epochs) is set to 200.

[0091] In each training epoch, the data is input into the model in batches of size 16 for training. The forward pass calculates the predicted values of the model, and then the loss function is calculated; the backward pass updates the parameters of the model.

[0092] After each training epoch, the model is evaluated using the validation set. The mean squared error (MSE) and mean absolute error (MAE) on the validation set are calculated, and the hyperparameters of the model are adjusted according to the performance of the validation set to prevent overfitting.

[0093] After training is completed, the final model is evaluated using the test set. The mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) on the test set are calculated to evaluate the generalization ability and prediction accuracy of the model.

[0094] The constructed 3D human body model of the patient with precise partitioning and semantic segmentation is input into a pre-trained deep learning model, which is constructed by learning the mapping relationship between a large number of known 3D human body models with weights and the actual weights. The model architecture uses the PointNet++ model, with the input being the geometric feature parameters of each part of the 3D human body model (such as volume, surface area, centroid coordinates of each part, etc.) and the point cloud density information, and the output being the predicted weight value. For example, for a bedridden patient of medium build, the model predicts a weight of about 65 kg based on the characteristics of the volume of their torso, the length and thickness of their limbs, etc., combined with the previously learned association rules between human body shape and weight. The prediction error is controlled within the range of ±3 kg, meeting the accuracy requirements for weight monitoring of bedridden patients in clinical practice, and providing an important reference basis for medical staff to formulate treatment plans and adjust medication doses.

[0095] Through the above detailed embodiments, the 3D optical scanning and weighing method for handheld bedridden patients can be systematically and accurately implemented, bringing convenience and accurate data support to medical care work.

[0096] Figure 2 Fig. shows the 3D optical scanning and weighing device 200 for handheld bedridden patients provided by the present invention, including a data acquisition module 210, a data processing module 220, a feature extraction module 230, a model construction module 240, and a prediction module 250.

[0097] The data acquisition module 210 is used to acquire three-dimensional point cloud data on the patient's body surface; The data processing module 220 is used to preprocess the three-dimensional point cloud data; The feature extraction module 230 is used to extract geometric features of the three-dimensional point cloud data; The model construction module 240 is used to partition based on the geometric features of the three-dimensional point cloud data and construct a three-dimensional human body model of the patient; The prediction module 250 is used to predict the three-dimensional human body model based on a pre-trained deep learning model to obtain the patient's weight.

[0098] In the embodiment of the present invention, the device further includes: a visualization display module configured to display patient data.

[0099] As described above, the above is only the specific implementation manner of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present disclosure should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A handheld 3D optical scanning weighing method for bedridden patients, characterized in that: The method comprises the following steps: Collect three-dimensional point cloud data of the patient's body surface; Preprocessing the three-dimensional point cloud data; Extract geometric features of 3D point cloud data; Partitioning is performed based on the geometric features of the 3D point cloud data to construct a 3D human body model of the patient; The three-dimensional human body model is predicted based on the pre-trained deep learning model to obtain the patient's weight.

2. The handheld 3D optical scanning weighing method for bedridden patients according to claim 1, characterized in that: The preprocessing of the three-dimensional point cloud data includes at least one of data filtering, noise reduction, feature extraction, point cloud downsampling, voxel downsampling, uniform downsampling and upsampling.

3. The handheld 3D optical scanning weighing method for bedridden patients according to claim 1, characterized in that: The pre-processing of the three-dimensional point cloud data further includes: point cloud registration; The point cloud registration includes: Find matching point pairs through random sampling consistency based on the geometric features of 3D point cloud data; Calculate the square error of all matching point pairs and sort them from small to large; Use the preset number of minimum error values to calculate the optimal rotation and translation matrix and update the source point cloud; Repeat the above steps until the convergence condition is met.

4. The handheld 3D optical scanning and weighing method for bedridden patients according to claim 1, characterized in that The extraction of geometric features of the point cloud includes: Calculate the normal and curvature of the point cloud; Extract the minimum bounding box or convex hull of a point cloud.

5. The handheld 3D optical scanning and weighing method for bedridden patients according to claim 4, wherein Calculate the curvature of a point cloud, including: Select the point cloud set N(p) with the first point cloud as the center and within the first radius; The fitting plane parameters are determined by minimizing the error function: ； Among them, x i , y i , z i are points within the point cloud set N(p); , , , are plane parameters; The curvature is calculated based on the following formula: ； in, is the distance from the point inside N(p) to the fitting plane; is the number of points in the neighborhood.

6. The handheld 3D optical scanning and weighing method for bedridden patients according to claim 1, wherein Partitioning based on the geometric features of the three-dimensional point cloud data to construct a three-dimensional human body model of the patient includes: Divide the key parts of the human body based on the geometric features of the three-dimensional point cloud data; wherein the key parts of the human body include: head, neck, chest, abdomen, upper arm, forearm, thigh, and calf; Perform semantic segmentation on point cloud data based on the pre-trained deep learning model and directly output the segmentation results of various parts of the human body.

7. The handheld 3D optical scanning and weighing method for bedridden patients according to any one of claims 1-6, characterized in that, The method further comprises: The prediction results are transmitted to the visualization platform via wireless transmission.

8. A handheld 3D optical scanning weighing device for bedridden patients, characterized in that: The device comprises: a data acquisition module configured to acquire three-dimensional point cloud data of a patient's body surface; A data processing module, configured to pre-process the three-dimensional point cloud data; a feature extraction module configured to extract geometric features of the three-dimensional point cloud data; a model building module configured to perform partitioning based on geometric features of the three-dimensional point cloud data to build a three-dimensional human body model of the patient; The prediction module is configured to predict the three-dimensional human body model based on a pre-trained deep learning model to obtain the patient's weight.

9. The handheld 3D optical scanning weighing device for bedridden patients according to claim 8, characterized in that: The device further includes a visual display module configured to display patient data.

Images