A high-precision three-dimensional reconstruction method for burn areas in burn patients

Abstract

This invention relates to the field of medical image data processing technology, specifically to a high-precision three-dimensional reconstruction method for burn areas of burn patients. The invention acquires multi-view RGB images, aligned structured light point clouds, and patient vital sign data; based on the ratio of gray intensity to the median gray intensity of neighboring areas and the cross-view reprojection intensity residual, a pixel reliability metric coefficient is obtained by coupling with the depth confidence product; based on this coefficient and the three-dimensional point positions, a triangular mesh is constructed through iterative fusion to obtain vertex positions and geometric confidence; the vertex burn voting rate is obtained through multi-view burn color gamut classification, and the product of the mean voting rate and the minimum geometric confidence is used as the triangular face burn confidence. The TBSA% center estimate and uncertainty interval are output after area-weighted summation. This invention suppresses specular reflection artifact interference and provides quantitative burn area results with a confidence interval for clinical fluid resuscitation.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing technology, specifically to a high-precision three-dimensional reconstruction method for the burn area of ​​a burn patient. Background Technology

[0002] Accurate quantification of burn area percentage (TBSA%) is a fundamental parameter for fluid resuscitation and surgical planning. Current clinical practice primarily relies on the nine-point method for estimation, Lund-Browder atlas comparison, and two-dimensional radiographic planarization. These methods are highly subjective and lack three-dimensional information. In recent years, structured light and multi-view imaging technologies have been gradually introduced into the field of burn assessment, enabling the acquisition of dense three-dimensional surface data. However, how to perform high-fidelity three-dimensional reconstruction of multi-view images and point clouds and quantify their uncertainties remains a data processing problem that needs to be solved.

[0003] In reality, burn wounds commonly exhibit surface features such as an exudate layer, a reflective eschar surface, and a smooth blister surface. Under structured light and multi-view imaging conditions, these surface features produce significant specular reflection artifacts. The RGB intensity values ​​of the same anatomical location presented at different viewpoints differ due to the view-direction dependence of reflected specular highlights, directly undermining the photometric consistency assumption upon which multi-view 3D reconstruction relies. Simultaneously, depth measurements in exudate areas are truncated or scattered by specular reflections of structured light fringes, leading to systematic biases in the depth map; cross-view reprojection of high-frequency geometric undulations such as eschar elevations at wound edges and blister boundaries inherently exhibits inter-view inconsistencies. These three types of defects—spectral highlights, unreliable depth, and cross-view inconsistencies—are physically independent, and existing methods lack the ability to systematically perceive and adaptively handle these defects. Furthermore, current TBSA% estimation only outputs a single-point estimate without a confidence interval, making it impossible for relevant personnel to accurately determine the decision-making safety margin for this value. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a high-precision three-dimensional reconstruction method for the burn area of ​​a burn patient. The specific technical solution adopted is as follows: A high-precision three-dimensional reconstruction method for the burn area of ​​a burn patient, comprising: acquiring RGB images of the patient's body surface from multiple perspectives and structured light three-dimensional point clouds aligned with the images from each perspective; acquiring the camera intrinsic parameter matrix, the extrinsic poses of each perspective, the pixel coordinates of pre-annotated anatomical landmarks on the body surface, and the patient's height and weight, wherein each pixel corresponds to a grayscale intensity value, a three-dimensional point position in a common coordinate system, and a depth confidence value; calculating the local specular ratio based on the ratio of the grayscale intensity value of the current pixel to the median grayscale intensity within its local neighborhood window; and obtaining the three-dimensional point position corresponding to the current pixel through cross-view reprojection using the camera intrinsic parameter matrix and the extrinsic poses of each perspective. The grayscale intensity values ​​of the 3D point from other viewpoints are used to calculate the cross-viewpoint luminosity residual based on the median of the absolute difference between the current pixel's grayscale intensity value and the grayscale intensity values ​​from other viewpoints. The reliability metric coefficient of the pixel is obtained by multiplying and coupling the depth confidence value, local specular ratio, and cross-viewpoint luminosity residual. A triangular mesh is iteratively constructed based on the reliability metric coefficient and the 3D point position to obtain the spatial position and geometric confidence of each vertex. The burn binary classification is performed based on the color information sampled from each visible viewpoint based on the spatial position of each vertex to obtain the vertex burn voting rate. The burn confidence of the triangular face is calculated by multiplying the average burn voting rate of the three vertices of the triangular face with the minimum geometric confidence value. The expected value of the burn area is obtained by weighted summation of the triangular face geometric area and burn confidence, and the TBSA% center estimate and uncertainty interval are calculated in combination with the total surface area.

[0005] Furthermore, the method for obtaining the local mirror ratio includes: taking all pixels in the image with valid three-dimensional point positions as valid pixels; a valid adjacent pixel pair is defined as two pixels that are both valid pixels and are horizontally or vertically adjacent in the image; and constructing a square local neighborhood window centered on the current pixel, the side length of which is determined by the following formula: In the formula, This represents the side length of the square local neighborhood window centered on the current pixel; Indicates the first The median of the Euclidean distances between the corresponding 3D point positions of all effective adjacent pixel pairs in an RGB image of a view; Indicates the projection on the first The median of the Euclidean distance between all pairs of anatomical landmarks within the field of view of an image; the local mirror ratio is the ratio of the gray intensity value of the current pixel to the median gray intensity value of all pixels within the window.

[0006] Furthermore, the method for obtaining the cross-view luminous residual includes: randomly selecting a pixel in the current viewpoint RGB image as a reference pixel; based on the 3D point position corresponding to the reference pixel in the current viewpoint RGB image, the camera intrinsic parameter matrix, the extrinsic parameter pose of the current viewpoint, and the extrinsic parameter poses of other viewpoints, projecting the corresponding 3D point of the reference pixel onto the image planes of all other viewpoints to obtain the grayscale intensity value at the coordinates of the reprojected pixel; calculating the absolute difference between the grayscale intensity value of the reference pixel and the grayscale intensity value at the coordinates of the reprojected pixel under all other viewpoints; using the median of all absolute differences as the cross-view luminous residual of the reference pixel; and iterating through all pixels under the current viewpoint to obtain the cross-view luminous residual of each pixel in the current viewpoint.

[0007] Furthermore, the method for obtaining the reliability metric coefficient includes: obtaining the reliability metric coefficient according to the reliability metric coefficient calculation formula, as shown below: In the formula, Indicates the first View current pixel Reliability metric coefficient Indicates the first View current pixel The depth confidence value ranges from [0,1]. Indicates the current pixel. The local mirror ratio; Indicates the current pixel. The trans-spectral residual; Indicates the first The third quartile of the local specular ratio of all effective pixels in the view; Indicates the first The third quartile of the trans-view luminance residuals of all valid pixels in the RGB image under view.

[0008] Furthermore, a triangular mesh is iteratively constructed based on the reliability metric coefficients and the 3D point positions to obtain the spatial position and geometric confidence of each vertex. This includes: constructing an initial triangular mesh, assigning initial spatial positions to each vertex of the initial triangular mesh, and determining the mesh topology; iterating over the initial triangular mesh, and in each iteration, performing the following process: using vertices... The reliability metric coefficients of all valid pixels within the spherical neighborhood space are used as weights to calculate the weighted average of the 3D point positions of each pixel within the spherical neighborhood space, thus obtaining the data item position. The radius of the spherical neighborhood space is taken as half of the median Euclidean distance between all adjacent vertices of the current grid. The position of the data item is calculated using the distance between the vertices and the vertex. The arithmetic mean of the spatial positions of all directly connected adjacent vertices at the convergence of the previous iteration is used as the position of the neighborhood smoothing term; for each vertex... The cumulative reliability weight is obtained by summing the reliability metric coefficients of all valid pixels within the spherical neighborhood space; the median of the cumulative reliability weights of all vertices in the entire grid is used as the global threshold; the vertex is obtained according to the following formula. The current spatial position and geometric confidence level : ; In the formula, Represents vertices Cumulative reliability weight; Represents vertices The location of the data item; Indicates the global threshold; Represents vertices The neighborhood smoothing term position; after the full grid vertex synchronous update, when the maximum change in vertex spatial position between the current round and the previous round is less than the median of the weighted standard deviation of the effective pixels in the spherical neighborhood space of each vertex relative to its data term position, the iteration terminates; otherwise, the next round begins.

[0009] Furthermore, an initial triangular mesh is constructed, and each vertex of the initial triangular mesh is assigned an initial spatial position and the mesh topology is determined. This includes: calculating the three-dimensional convex hull using the three-dimensional coordinates of all markers in the anatomical landmark set; using the intercepted region of the three-dimensional convex hull on the body surface manifold as the tiling range; and tiling the triangular mesh with uniform density within the tiling range. The initial vertex spacing of the triangular mesh is set to three times the mean of the median of the Euclidean distances between the corresponding three-dimensional point positions of all effective adjacent pixel pairs in the RGB images under all viewpoints. The position of each initial vertex is taken as the arithmetic mean of the three-dimensional point positions corresponding to all viewpoint pixel points falling within its initial spherical neighborhood space. The radius of the initial spherical neighborhood space is half of the initial vertex spacing.

[0010] Furthermore, the method for obtaining binary burn classification includes: in the current session, using the burn wound color samples and healthy skin color samples marked by clinicians around anatomical landmarks as references, calculating the hue-saturation mean vector and covariance matrix of the two classes in the HSV color space respectively, and constructing a minimum Mahalanobis distance classifier; based on the spatial position of each vertex, the camera intrinsic parameter matrix, and the extrinsic pose of each viewpoint, projecting each vertex onto the image plane of each viewpoint to obtain the visible viewpoint; wherein, the visible viewpoint satisfies: the projected coordinates of each vertex fall within the effective pixel range of the image of that viewpoint, and the angle between the normal vector of that vertex and the optical center of the camera of that viewpoint does not exceed 90 degrees; sampling the color vector of each vertex in all visible viewpoints using bilinear interpolation and converting it to the HSV space, and outputting the binary discrimination result by the classifier; the burn voting rate of that vertex is set as the ratio of the number of visible viewpoints judged as burns to the total number of visible viewpoints.

[0011] Furthermore, the method for obtaining the confidence level of triangular facet burns includes: obtaining the confidence level of triangular facet burns according to the formula for calculating the confidence level of triangular facet burns, as shown below: In the formula, Indicates the confidence level for triangular facet burns; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices ,vertex With vertex The minimum geometric confidence level.

[0012] Furthermore, the method for obtaining the expected value of the burn area includes: obtaining the expected value of the burn area according to the formula for calculating the expected value of the burn area, as shown below: In the formula, This represents the expected burn area. Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; This indicates the preset number of triangular facet regions; further, the methods for obtaining the TBSA% center estimate and uncertainty interval include: calculating the standard deviation of the burn area, as shown in the following formula: In the formula, Indicates the standard deviation of burn area; Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; Represents the face set of a triangular mesh. Represent each triangle in the face set; calculate the total surface area of ​​the volume. The calculation formula is as follows: In the formula, This represents the total surface area of ​​the human body. Indicates the patient's height; This represents the patient's weight; the central estimate of TBSA% is determined by the following formula: In the formula, This represents the central estimate of TBSA%. This represents the expected burn area; the uncertainty interval for TBSA% is determined by the following formula: In the formula, Indicates the TBSA% uncertainty range; This represents the expected burn area. Indicates the standard deviation of burn area; It represents the total surface area of ​​the human body.

[0013] This invention offers the following advantages: By constructing a pixel reliability metric coefficient coupled with the product of a depth confidence factor, a local specular ratio factor, and a cross-view refractive index factor, this invention systematically reduces the impact of three independent defects—spectral highlights, unreliable depth, and cross-view inconsistency—on 3D reconstruction. This automatically reduces the weight of unreliable pixels, allowing reliable pixels to dominate the reconstruction process. During the iterative fusion phase, the fitting strategy is adaptively switched by comparing the accumulated reliability weights with a global threshold. Vertices in reliable data regions strictly adhere to multi-view observations, preserving true geometric undulations such as eschar ridges and blister boundaries. Vertices in unreliable data regions are dominated by neighborhood smoothing constraints, suppressing false fluctuations caused by specular artifacts. This switching is entirely driven by the data's inherent quality, requiring no manual setting of region boundaries. In the burn semantic annotation phase, grid vertices are projected back to each visible viewpoint; independent classification is performed based on a color gamut discriminator adaptively calibrated according to the current session, making burn color gamut determination independent of lighting conditions and patient skin color differences. The final output of the TBSA% center estimate and uncertainty interval allows clinicians to determine the decision safety margin for this assessment based on the interval width. This invention achieves high-fidelity 3D reconstruction of burn areas under specular reflection artifact interference and provides TBSA% quantization results with confidence intervals. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a high-precision three-dimensional reconstruction method for the burn area of ​​a burn patient, provided as an embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a high-precision three-dimensional reconstruction method for burn areas of burn patients proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] The following description, in conjunction with the accompanying drawings, details a specific scheme for a high-precision three-dimensional reconstruction method for burn areas of burn patients provided by the present invention.

[0019] Please see Figure 1 This invention illustrates a high-precision three-dimensional reconstruction method for the burn area of ​​a burn patient provided by an embodiment of the present invention. The method includes: step S1: acquiring RGB images of the patient's body surface from multiple perspectives and structured light three-dimensional point clouds aligned with the images from each perspective; acquiring the camera intrinsic parameter matrix, the extrinsic pose of each perspective, the pixel coordinates of pre-labeled anatomical landmarks on the body surface, and the patient's height and weight, wherein each pixel corresponds to a grayscale intensity value, a three-dimensional point position in a common coordinate system, and a depth confidence value.

[0020] The purpose of this step is to complete the acquisition of all the basic parameters required for subsequent processing. In the following text, "viewpoint" will be abbreviated as "view" to refer to a single frame observation corresponding to a certain camera or imaging position in a multi-view imaging system. The k-th view refers to the k-th viewpoint, where k is the viewpoint index, used to distinguish between images and point cloud data acquired from different directions on the patient's body surface.

[0021] The system retrieves the multi-view RGB image sequence for this session from the hospital's image archiving system interface. For the k-th view, the system obtains the RGB three-channel intensity values ​​at each pixel coordinate (u,v) under that view, records them as digital grayscale intensity values ​​after photometric calibration, and denoted as . , dimensionless.

[0022] Dense 3D point cloud data synchronized with the clock of each view image is pulled from the data stream interface of the structured light imaging system. In this data, each pixel (u,v) in each frame of the point cloud corresponds to a 3D point position in a common coordinate system obtained through structured light triangulation, denoted as . The unit is millimeters. It should be noted that the structured light imaging system projects known-coded structured light stripes onto the body surface. The camera captures the deformed image of the stripes on the body surface. For each pixel (u,v), the three-dimensional coordinates of the corresponding body surface position in a common coordinate system are calculated using the triangulation relationship between the projector and the camera. Therefore, Instead of directly converting 2D pixel coordinates into 3D spatial points, a lookup table mapping is used. That is, the 3D position vector corresponding to the pixel is read from the dense point cloud data output by the structured light system, using the 2D pixel coordinates as an index. Each view k has a 2D image and a depth map aligned with it at the pixel level. The superposition of the two provides the complete 3D observation for that view.

[0023] A common coordinate system is a fixed, global 3D reference frame used to uniformly describe the spatial positions of all cameras and all 3D points. In a structured light imaging system, each camera initially generates a 3D point cloud in a camera coordinate system with its own optical center as the origin. The origins and orientations of the camera coordinate systems from different views are different, and they are independent of each other. The purpose of the common coordinate system is to enable 3D points from different views to undergo vertex fusion and mesh reconstruction in subsequent steps within the same spatial frame.

[0024] The structured light imaging system outputs a 3D point cloud and simultaneously provides a self-reported depth confidence score for each pixel. This confidence score is generated pixel-by-pixel by the structured light system based on its internal phase resolution quality and signal-to-noise ratio metrics, and is denoted as . value range Dimensionless. The above three types of pixel-by-pixel parameters... , and They share the same pixel coordinate grid within a single view.

[0025] The camera intrinsic parameter matrix recorded during image acquisition is denoted as... It includes the equivalent focal length and principal point coordinates, where the focal length is in millimeters. The data comes from the image system's data header field and is stored along with each frame of the image.

[0026] The extrinsic pose data of the k-th view recorded during image acquisition is denoted as This parameter is a composite data object that contains a translation vector. With rotation matrix Translation is measured in millimeters, and rotation is measured in degrees. It should be noted that... and Both must be used simultaneously because knowing only the camera's orientation without knowing its position, or knowing only its position without knowing its orientation, is insufficient to transform a 3D point from the common coordinate system to the camera coordinate system. In subsequent cross-view reprojection, the 3D point of the k-th view needs to be transformed to the common coordinate system first, and then to the camera coordinate system of the k'-th view. Each step in this process requires both rotation and translation. This data also originates from the image system's header data field and is recorded synchronously with the intrinsic parameters.

[0027] The set of pixel coordinates of anatomical landmarks on the body surface pre-labeled by clinicians in the image is denoted as It covers bony landmarks on the body surface such as the acromion, umbilicus, and anterior superior iliac spine, which are manually marked on the images by clinicians before collection as a medical prior.

[0028] Patient-related basic parameters include patient height and patient weight All units are in centimeters, and the data comes from the patient basic information field of the electronic health record.

[0029] All the basic parameters collected above are used as a parameter set. Package: At this point, the collection of all required basic parameters is complete.

[0030] Step S2: Calculate the local specular ratio based on the ratio of the gray intensity value of the current pixel to the median gray intensity in its local neighborhood window; obtain the gray intensity value of the 3D point corresponding to the current pixel through cross-view projection using the camera intrinsic parameter matrix and the extrinsic pose of each viewpoint; calculate the cross-view residual based on the median of the absolute difference between the gray intensity value of the current pixel and the gray intensity values ​​of other viewpoints; and obtain the reliability metric coefficient of the pixel by multiplying and coupling the depth confidence value, the local specular ratio, and the cross-view residual.

[0031] Burn wounds commonly exhibit surface features such as an exudate layer, a reflective eschar surface, and a smooth blister layer. Under structured light and multi-view imaging conditions, these surface features produce significant specular reflection artifacts. The RGB intensity values ​​of the same anatomical location appearing under different views differ due to the view-direction dependence of reflected highlights, directly undermining the photometric consistency assumption upon which multi-view 3D reconstruction relies. Simultaneously, depth measurements in exudate areas suffer from systematic biases in the depth map due to specular reflection or scattering of structured light fringes, which contradicts the pixel-by-pixel confidence provided by the structured light system itself. This can, to some extent, identify such measurement anomalies. Furthermore, even excluding specular reflection, the cross-view reprojection of high-frequency geometric undulations such as eschar elevation at the wound edge and blister boundaries inherently exhibits interview inconsistency, reflected as an increase in cross-view photometric residuals. The physical sources of defects such as specular highlights, unreliable depth, and cross-view inconsistency are independent of each other; therefore, they need to be identified independently, coupled with a logic that rejects the contribution of a pixel if any defect reflects deterioration. The purpose of this step is to define the reliability metric coefficients for all valid pixels in images from different viewpoints and use them as the source of weighting coefficients for subsequent data fidelity candidate vertices.

[0032] Preferably, in one embodiment of the present invention, the method for obtaining the local mirror ratio includes: taking all pixels in the image with valid three-dimensional point positions as valid pixels; constructing a square local neighborhood window centered on the current pixel, the side length of which is determined by the following formula: In the formula, This represents the side length of the square local neighborhood window centered on the current pixel; Indicates the first The median of the Euclidean distances between the corresponding 3D point positions of all effective adjacent pixel pairs in an RGB image of a view; Indicates the projection on the first The median of the Euclidean distances between all pairs of anatomical landmarks within the field of view of an image; The calculation methods include: for the first The set of all valid pixels in a view image traverse each pixel one by one If its horizontally adjacent pixels Belonging to Then take the three-dimensional point positions of both. and And calculate the Euclidean distance between them; if their vertically adjacent pixels (u, v+1) also belong to the same category. Then, calculate the Euclidean distance in the same way; form a distance set from the distance values ​​of all the above valid adjacent pixel pairs, and take the median of this set as the distance. ; The calculation method is as follows: from the global set of anatomical landmarks Selecting the projections from the first A subset of anatomical landmarks within the field of view of a single image. ,calculate The pixel Euclidean distances between all pairwise markers in the dataset are used to form a distance set. The median of this set is the median distance. This value characterizes the typical scattering scale of the k-th suborbital anatomical landmark on the image plane; the closer the shooting distance or the higher the resolution, the larger the value.

[0033] The local specular ratio is the ratio of the gray intensity value of the current pixel to the median gray intensity value of all pixels within the window. Because specular highlights generated by exudate from burn wounds and the surface of blisters appear as abnormally bright pixels in a single-view image, with intensity values ​​significantly higher than the normal intensity levels of adjacent surface areas under the same view, a local specular ratio significantly greater than 1 indicates that the pixel's intensity far exceeds its surrounding background, suggesting specular highlight pollution. A local specular ratio close to 1 indicates that the pixel's intensity is close to its surrounding background, consistent with the photometric characteristics of diffuse reflection surfaces.

[0034] ratio Characterizes the number of adjacent pixel units within an anatomical scattering scale; the resulting value is in units of pixels, multiplied by [value]. Then, the number of pixels that the window radius should cover is determined. The outer layer rounding down ensures that the side length is an integer, and multiplying by 2 and adding 1 ensures that the window length is an odd number, thus making the window strictly symmetrical about the center pixel (u,v).

[0035] When the same anatomical location on the body surface is observed under different views, if the body surface conforms to the ideal diffuse reflection hypothesis (Lambertian surface), the observed intensity should be consistent across all views. However, the specular reflection of burn wounds is view-dependent, causing highlights to appear only at specific angles in a particular view. This results in the same location appearing as bright in one view but normal in another. Therefore, by measuring the degree of intensity inconsistency of a pixel across multiple views, unreliable pixels whose optical behavior deviates from the diffuse reflection hypothesis can be identified.

[0036] Preferably, in one embodiment of the present invention, the method for obtaining the cross-viewpoint luminance residual includes: arbitrarily selecting a pixel in the current viewpoint RGB image as a reference pixel; based on the 3D point position corresponding to the reference pixel in the current viewpoint RGB image, the camera intrinsic parameter matrix, the extrinsic parameter pose of the current viewpoint, and the extrinsic parameter poses of other viewpoints, projecting the corresponding 3D point of the reference pixel onto the image planes of all other viewpoints to obtain the grayscale intensity value at the coordinates of the reprojected pixel; calculating the absolute difference between the grayscale intensity value of the reference pixel and the grayscale intensity value at the coordinates of the reprojected pixel under all other viewpoints; using the median of all absolute differences as the cross-viewpoint luminance residual of the reference pixel; and iterating through all pixels under the current viewpoint to obtain the cross-viewpoint luminance residual for each pixel in the current viewpoint, as shown in the following calculation formula: In the formula, Indicates the residual across the optical range; and It is the position of the 3D point corresponding to the reference pixel. through , and After performing the reprojection transformation, at the 1st The pixel coordinates obtained in the image plane of each view. This is the grayscale intensity value at the coordinates of the reprojected pixel.

[0037] A small cross-viewpoint luminance residual indicates that the intensity values ​​measured under different views are close to each other, and the optical behavior of the body surface location conforms to the diffuse reflection hypothesis. A large cross-viewpoint luminance residual indicates that the intensity of the body surface location corresponding to the pixel point varies significantly under different views, and does not meet the premise of multi-viewpoint luminance consistency. Taking the median instead of the mean is to exclude the excessive influence of individual extreme high-brightness views on the residual estimation.

[0038] Then, determine the adaptive threshold. Mirror ratio threshold. Take the first Local mirror ratio of all effective pixels in a view The third quartile Cross-view residual threshold Take the first The trans-view luminous residual of all effective pixels in a view The third quartile The third quartile is used as the threshold, so that the judgment criterion adapts to the overall imaging quality of the view. When the imaging conditions are poor and the global noise is high, the threshold is automatically raised to avoid over-rejection; when the imaging conditions are good, the threshold is automatically tightened to improve the discrimination sensitivity.

[0039] Finally, the reliability metric coefficient is determined by the following formula: In the formula, Indicates the first View current pixel Reliability metric coefficient Indicates the first View current pixel The depth confidence value ranges from [0,1]. Indicates the current pixel. The local mirror ratio; Indicates the current pixel. The trans-spectral residual; Indicates the first The third quartile of the local specular ratio of all effective pixels in the view; Indicates the first The third quartile of the trans-view luminance residuals of all valid pixels in the RGB image under view.

[0040] When specular highlight, depth measurement confidence level, and cross-view diffraction difference occur alone, the credibility of that pixel as a multi-view observation source is negated. Therefore, any part approaching the... The overall weight also tends to be close to When the original error is much smaller than the adaptive threshold of the corresponding part of the view. Time factor close This indicates that the data quality in this dimension is good, when the error is much greater than... Time factor Attenuation is performed.

[0041] Step S3: Iteratively construct a triangular mesh based on the reliability metric coefficient and the three-dimensional point positions to obtain the spatial position and geometric confidence of each vertex.

[0042] The reliability metric coefficients of the effective pixels in each viewpoint obtained in step S2 are spatially dense but not uniform. In areas with good data quality (such as dry, healthy skin, diffuse reflective surfaces, etc.), the pixels are dense and highly reliable, while in areas with deteriorated data quality (such as areas with specular highlights from exudate, areas with abnormal depth measurements, etc.), the pixels may still be dense, but the reliability has been reduced to near zero. Furthermore, the geometric characteristics of the body surface also exhibit significant differences: at a macroscopic scale, the body surface is a locally smooth two-dimensional manifold, and adjacent vertices should have a certain geometric continuity; however, at the microscopic scale of the wound, eschar ridges, blister boundaries, necrotic tissue edges, etc., constitute real local high-frequency undulations, and these undulations are structural information that needs to be preserved. Therefore, in this embodiment, in areas with reliable data, the mesh vertices strictly fit the multi-view observations to preserve real details; in areas with unreliable data, the curved surfaces of the observation failure areas are allowed to transition naturally through the geometric extension of adjacent reliable areas, thereby suppressing the generation of false undulations.

[0043] Preferably, in this embodiment of the invention, the fitting strategy is adaptively switched in the unified fusion framework according to the reliability of the observation data. Specifically, the three-dimensional positions of multi-view pixels are fused into a unified triangular mesh. The iterative update process depends on the determination of the position of the neighboring vertex and the spherical belonging to the neighborhood space. Both require that the vertex has an initial position and topological connection. Therefore, it is necessary to construct the initial mesh first. The result of the initial construction is to give each vertex an initial position guess and establish the mesh topology, so that the iterative loop can obtain the cold start condition.

[0044] First, an initial triangular mesh is constructed. The mesh area covers the surface region affected by the current burn, and its spatial boundaries are defined by a set of anatomical landmarks. The three-dimensional position of the landmark points is determined using a common coordinate system. This is because burn assessment does not involve reconstructing the entire body, but only the surface area where the burn wound is located. This defines the objective area for burn assessment on the body surface. It needs to be clarified that... Scattered across the surface of the volume in a common coordinate system, these points do not directly form a cutting line, but rather generate a three-dimensional bounding volume through computational geometry methods, i.e., taking the entire volume. The three-dimensional convex hull of the marker points is calculated using the common coordinate system coordinates. This convex hull encloses all the marker points in a minimal convex polyhedron, and the mesh coverage area is the intercepted region of this convex hull on the surface manifold.

[0045] Within this area, a triangular mesh is laid out at a uniform density as the initial mesh. The initial vertex spacing is set to... ,in For RGB images from all viewpoints The average value of the fixed multiple is 3, which is suggested to balance computational efficiency and geometric resolution. The position of each initial vertex is taken as the position of the 3D point corresponding to all view pixels falling within its initial spherical neighborhood space. The arithmetic mean of the vertices. It should be noted that during the initial construction phase, vertices are not yet bound to any specific view; their task is to approximate the true position on the surface as closely as possible. Therefore, regardless of which view a pixel comes from, as long as its common coordinate system 3D point falls within the initial spherical neighborhood space of the current vertex, it will be included in the collection range and participate in the arithmetic mean. The radius of the initial spherical neighborhood space. Take half of the initial vertex spacing. The mesh topology is determined once during the initial construction, and subsequent iterations only update the vertex positions on this topology.

[0046] The mesh vertices are evenly spaced during the initial construction. The value only applies to the initial attribution determination. In subsequent iterations, vertex positions shift due to the combined effects of data pull and neighborhood constraints, resulting in uneven Euclidean spacing between adjacent vertices. If the initial radius is still used, vertices in reliable data regions tend to cluster due to repeated pull from numerous pixels, leading to excessive neighborhood overlap with a fixed radius. Conversely, the number of effective pixels within the fixed radius's coverage area in sparse data regions decreases significantly. Therefore, before each iteration, the median of the Euclidean spacing between all adjacent vertices in the current grid needs to be recalculated, and half of it should be assigned to the initial radius. This ensures that the scale of the sphere's neighborhood space always depends on the current resolution of the grid.

[0047] The initial triangular mesh is iterated, and in each iteration, the following process is performed: starting with the vertices... Using the reliability metric coefficients of all valid pixels within the spherically assigned neighborhood space as weights, a weighted average of the 3D point positions of each pixel within the spherically assigned neighborhood space is performed to obtain the data item positions. This is used to solve the problem of vertex positions when the data is fully trusted. The appropriate position; the radius of the sphere's neighborhood space is half the median Euclidean distance between all adjacent vertices of the current grid.

[0048] Calculation and Vertex The arithmetic mean of the spatial positions of all directly connected adjacent vertices at the convergence of the previous iteration is used as the position of the neighborhood smoothing term. This is used to solve for the vertex position when the neighborhood geometric relations are fully trusted. The appropriate location. Take the vertex. The set of all adjacent vertices The arithmetic mean of the current spatial coordinates of the midpoint is used to obtain the neighborhood relation position with smoothness as a prior condition. Its construction method expresses the smoothness prior of local surface smoothness as follows: when there is no observation information or the observation quality is unreliable, the most reasonable inferred position of a vertex is the geometric center of its topological neighborhood. An arithmetic mean is chosen instead of a weighted average based on neighborhood side lengths or areas because neighborhood vertices have equal geometric proximity at the topological level, and each connection... The edges with adjacent vertices do not exhibit differences similar to those caused by specular reflection or inconsistencies in luminance in pixel observations, and the uniform weights also satisfy the premise that topological neighborhood relations are the only reliable basis for inference when data is missing.

[0049] For vertices The cumulative reliability weight is obtained by summing the reliability metrics of all valid pixels within the spherical neighborhood space. This cumulative reliability weight is then used to aggregate the reliability metrics of all pixels scattered within the neighborhood into a single scalar, serving as a quantitative representation of the overall reliability of the local data at that vertex. This scalar represents the location where observations are dense and reliable. The tendency is towards larger areas, sparsely observed locations, or areas significantly contaminated by specular artifacts. It tends to be smaller.

[0050] The median of the cumulative reliability weights of all vertices in the entire grid is used as the global threshold; the vertex values ​​are obtained according to the following formula. The current spatial position and geometric confidence level : ; In the formula, Represents vertices Cumulative reliability weight; Represents vertices The location of the data item; Indicates the global threshold; Represents vertices The position of the neighborhood smoothing term.

[0051] when Higher than hour, As the vertex approaches 1, its new position approaches the position of the data item determined by all valid pixels within the local region. The observational evidence in this local area was deemed highly reliable, and the corresponding true geometric undulations of the wound surface were preserved; when Below hour, As the vertex approaches 0, its new position becomes closer to the position of the neighborhood smoothing term obtained from the mean of its neighboring vertices. The effect of all effective pixels within a local area is significantly weakened, and the new location is dominated by smoothness constraints, thus suppressing false fluctuations caused by mirror artifacts. Geometric confidence. The physical meaning is the intensity ratio of a vertex position driven by all effective pixels within a local region. When the geometric confidence approaches 1, it indicates that the vertex... The final position is mainly derived from the direct geometric evidence reflected by the effective pixels generated under multi-view observation. When the geometric confidence approaches 0, it means that the new position of the vertex is mainly inferred from the prior of neighborhood smoothness.

[0052] After the full grid vertex synchronization update, the iteration terminates when the maximum change in vertex spatial position between the current and previous rounds is less than the median of the weighted standard deviation of the effective pixels in the spherical neighborhood space of each vertex relative to their data item position; otherwise, it proceeds to the next round. The specific process includes: Find the maximum change in the spatial position of the vertices between the current and previous rounds. The maximum value. The convergence condition is: when the maximum change decreases to the adaptive convergence tolerance ratio. If the iteration ends, the iteration terminates; otherwise, the result of this round overwrites the result of the previous round and proceeds to the next round. The method for determining vertex v is as follows: for vertex v, in All valid pixels relative to their weighted average Calculate the weighted standard deviation and denot it as . Take all vertices the median of As The meaning of this convergence criterion is that when the maximum displacement of all grid vertices between two iterations has decreased to below the typical measurement noise level at various points on the current session surface, the remaining displacement is more of a random manifestation of noise than a substantial geometric improvement, and continuing iteration does not produce a discernible gain in accuracy, thus determining convergence.

[0053] The output at convergence after the iteration is completed That is, the vertex The final spatial location, corresponding to That is, the vertex The final geometric confidence level. The triangular mesh uses... It means that, among them The vertex set is the set of all nodes in the mesh, where each vertex... Carry its three-dimensional position With geometric confidence Two attributes; F is the face set, which is the collection of all triangular face regions in the mesh, each region It consists of three vertex indices.

[0054] Step S4: Perform binary burn classification based on the color information sampled from each visible viewpoint by projecting the spatial position of each vertex to obtain the vertex burn voting rate; calculate the burn confidence of the triangular face by multiplying the mean burn voting rate of the three vertices of the triangular face by the minimum geometric confidence value; obtain the expected value of the burn area by weighted summation of the geometric area of ​​the triangular face and the burn confidence value, and calculate the central estimate of TBSA% and the uncertainty interval by combining the total body surface area.

[0055] Since the triangular regions in the mesh only have spatial geometric properties and lack semantic labels for burns or health, this step first projects the mesh vertices back to the image plane of each visible view, performs burn determination independently in each view, aggregates the multi-view determination results, and finally performs data composition on each triangular region to obtain the burn confidence level corresponding to each triangular region.

[0056] First, a vertex burn voting process is performed. Preferably, in this embodiment of the invention, for each vertex... According to its spatial location Camera intrinsic parameter matrix and the Individual view external parameter pose Perform a perspective projection transformation to map the vertex from the common coordinate system to the image plane of each view. A visible view must satisfy the following two conditions: the projected coordinates fall within the effective pixel range of the view's image, and the angle between the vertex normal and the optical center of the view's camera does not exceed 90 degrees. Views that do not satisfy either condition are considered as views of the vertex. Invisible and not participating in subsequent voting processes. For views determined to be visible, bilinear interpolation is used at the projected coordinates from... The RGB color vector corresponding to the vertex is sampled and converted to the HSV color space.

[0057] The burn color gamut discrimination was performed using a minimum Mahalanobis distance classifier. The decision boundary of this classifier was not determined using a preset fixed threshold, but rather by the clinical physician's judgment of anatomical landmarks during the current session. The color distribution of manually labeled burn wound color reference samples and healthy skin color reference samples is adaptively determined. The hue-saturation two-dimensional mean vector and covariance matrix of the two types of reference samples are calculated separately in the HSV color space, and the discrimination boundary is taken as the spatial trajectory of the planes with equal Mahalanobis distances between the two types. This adaptive mechanism makes the burn color gamut determination independent of the lighting environment and patient skin color differences, and is independently calibrated each time a patient's burn is assessed.

[0058] For vertices Traverse all visible viewpoints and summarize the binary discrimination results to determine the vertex burn voting rate. The ratio of the number of visible views identified as burns to the total number of visible views is taken, with a value ranging from [0,1]. This value represents the proportion of views from which the vertex is identified as a burn area in multi-view independent observations.

[0059] Next, the confidence level of the triangular facet burn is calculated. Preferably, in this embodiment of the invention, the method for obtaining the confidence level of the triangular facet burn includes: obtaining the confidence level of the triangular facet burn according to the formula for calculating the confidence level of the triangular facet burn, as shown below: In the formula, Indicates the confidence level for triangular facet burns; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices ,vertex With vertex The minimum geometric confidence level.

[0060] In the formula, The arithmetic mean of the three vertices' burn voting rates reflects the overall burn tendency of the triangular face region under multi-view classification. A mean value close to 1 indicates a stronger tendency for the three vertices to be consistently judged as burns by multiple views. A mean value close to 0 indicates a stronger tendency to be consistently judged as healthy. A mean value in the middle indicates that the triangular face region crosses the burn boundary or that there is a discrepancy in the judgment of different views. Represents vertices ,vertex With vertex The minimum geometric confidence level indicates that if the spatial location of even one vertex is unreliable, the geometric location of the entire triangular facet region is questionable, and the burn determination cannot be accurately anchored to a specific location on the body surface. If the average burn voting rate is very low, the reflected area is likely not a burn area. Even if the geometric confidence level is high, the product result will still be low, and its contribution to subsequent facet accumulation will be weakened. If the geometric confidence level is low, it indicates that the region location is unreliable. Even if the burn voting rate is high, the product will still be reduced.

[0061] Finally, area quantification and TBSA% uncertainty interval output are performed. In burn clinical practice, TBSA% is one of the important influencing factors of fluid resuscitation dosage. Overestimation leads to overresuscitation, potentially causing pulmonary edema and abdominal compartment syndrome; underestimation leads to inadequate resuscitation, potentially causing acute kidney injury and multiple organ dysfunction. Therefore, outputting only a single-point TBSA% estimate without its confidence interval makes it impossible for clinicians to accurately determine the decision safety margin represented by that single-point value. This invention obtains the expected burn area by weighting and summing the triangular geometric areas using burn confidence levels, and then calculates the central estimate of TBSA% and its uncertainty interval by combining it with the total body surface area.

[0062] Preferably, in this embodiment of the invention, the triangular facet geometric area Based on Heron's formula and the spatial positions of the three vertices , and Calculate the expected burn area in square millimeters. The result is obtained by weighting and summing the areas of all triangular faces using their burn confidence scores: In the formula, This represents the expected burn area. Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; The preset number of triangular facet regions is indicated; the methods for obtaining the TBSA% center estimate and uncertainty interval include: calculating the standard deviation of the burn area, as shown in the following formula: In the formula, Indicates the standard deviation of burn area; Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; Represents the face set of a triangular mesh. Let represent each triangle in the face set; where the variance expression for a single triangle region is... The standard definitional derivation from the variance of a binary random variable models each triangular region as a binary random variable with the expectation of... The variance is the scale factor calculated by multiplying the expectation by (1 - expectation) and then by the square of the area. Under the assumption of independent triangular regions, the total variance equals the sum of all variances, which is a fundamental conclusion in probability theory regarding the additivity of variances of the sum of independent random variables. This variance structure exhibits adaptive behavior: exist The value is largest when it approaches 0.5, corresponding to the most ambiguous situation in determining the nature of burns. (Corresponding standard deviation) In When the value approaches 0 or 1, the standard deviation approaches 0, corresponding to a situation where the judgment is clearly certain. If the mirror artifact is suppressed... After geometric confidence Decline and After coupling, a large number of regions It tends to be around 0.5, leading to As a result, the range widens.

[0063] Calculate the total body surface area The calculation formula is as follows: In the formula, This represents the total surface area of ​​the human body. Indicates the patient's height; This indicates the patient's weight; the unit is centimeters, while BSA is in square meters.

[0064] The central estimate of TBSA% is determined by the following formula: In the formula, This represents the central estimate of TBSA%. This represents the expected burn area. This is the dimension conversion factor from square millimeters to square meters. Divide by Complete the dimensional conversion from square millimeters to dimensionless percentages to obtain the final percentage estimate.

[0065] The TBSA% uncertainty interval is determined by the following formula: In the formula, Indicates the TBSA% uncertainty range; This represents the expected burn area. Indicates the standard deviation of burn area; It represents the total surface area of ​​the human body.

[0066] In clinical applications, the central estimate of TBSA% can be used by relevant personnel to calculate the baseline dose using the standard formula for fluid resuscitation, while the TBSA% uncertainty interval indicates the confidence range of this baseline dose. When the TBSA% uncertainty interval is narrow, the safety margin of administration according to the standard formula is higher; when the TBSA% uncertainty interval is wide, relevant personnel can make detailed adjustments to the fluid resuscitation plan based on the upper limit of the interval and mark the uncertainty level of this assessment in the electronic medical record based on their personal experience.

[0067] In summary, RGB images of the patient's body surface from multiple viewpoints and structured light 3D point clouds aligned with the images from each viewpoint were acquired. The camera intrinsic matrix, extrinsic poses for each viewpoint, pixel coordinates of pre-annotated anatomical landmarks on the body surface, and the patient's height and weight were obtained. For each pixel, a grayscale intensity value, a 3D point position in a common coordinate system, and a depth confidence value were also acquired. The local specular ratio was calculated based on the ratio of the current pixel's grayscale intensity value to the median grayscale intensity within its local neighborhood window. The grayscale intensity value of the 3D point corresponding to the current pixel was obtained by cross-viewpoint reprojection using the camera intrinsic matrix and extrinsic poses for each viewpoint. The grayscale intensity values ​​of the 3D point in other viewpoints were then calculated based on the ratio of the current pixel's grayscale intensity value to the grayscale intensity values ​​of other viewpoints. The median of the absolute difference is used to calculate the cross-viewpoint residual. The reliability metric coefficient of the pixel is obtained by coupling the depth confidence value, local specular ratio, and cross-viewpoint residual through product. A triangular mesh is iteratively constructed based on the reliability metric coefficient and the 3D point position to obtain the spatial position and geometric confidence of each vertex. The vertex burn voting rate is obtained by performing binary burn classification based on the color information sampled from each visible viewpoint based on the spatial position of each vertex. The burn confidence of the triangular face is calculated by multiplying the mean burn voting rate of the three vertices of the triangular face with the minimum geometric confidence value. The expected value of the burn area is obtained by weighted summation of the triangular face geometric area and burn confidence. The center estimate of TBSA% and the uncertainty interval are calculated by combining the total surface area.

[0068] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0069] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A high-precision three-dimensional reconstruction method for burn areas of burn patients, characterized in that, The method includes: acquiring RGB images of the patient's body surface from multiple perspectives and a structured light 3D point cloud aligned with the images from each perspective; obtaining the camera intrinsic parameter matrix, extrinsic poses for each perspective, pixel coordinates of pre-labeled anatomical landmarks on the body surface, and the patient's height and weight, wherein each pixel corresponds to a grayscale intensity value, a 3D point position in a common coordinate system, and a depth confidence value; calculating the local specular ratio based on the ratio of the grayscale intensity value of the current pixel to the median grayscale intensity within its local neighborhood window; obtaining the grayscale intensity value of the 3D point corresponding to the current pixel from other perspectives through cross-view reprojection using the camera intrinsic parameter matrix and extrinsic poses for each perspective, and calculating the grayscale intensity value of the current pixel relative to the grayscale intensity value of the other perspectives. Calculate the cross-viewpoint residual based on the median of the difference; multiply and couple the depth confidence value, the local specular ratio, and the cross-viewpoint residual to obtain the reliability metric coefficient of the pixel; iteratively construct a triangular mesh based on the reliability metric coefficient and the three-dimensional point position to obtain the spatial position and geometric confidence of each vertex; perform binary burn classification based on the color information sampled from each visible viewpoint according to the spatial position of each vertex to obtain the vertex burn voting rate; calculate the triangular face burn confidence based on the product of the mean burn voting rate of the three vertices of the triangular face and the minimum geometric confidence; obtain the expected burn area by weighted summation of the triangular face geometric area and the burn confidence, and calculate the TBSA% center estimate and uncertainty interval in combination with the total body surface area.

2. The high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the local mirror ratio includes: taking all pixels in the image with valid three-dimensional point positions as valid pixels; a valid adjacent pixel pair is defined as two pixels that are both valid pixels and are horizontally or vertically adjacent in the image; constructing a square local neighborhood window centered on the current pixel, the side length of which is determined by the following formula: In the formula, This represents the side length of the square local neighborhood window centered on the current pixel; Indicates the first The median of the Euclidean distances between the corresponding 3D point positions of all effective adjacent pixel pairs in an RGB image of a view; Indicates the projection on the first The median of the Euclidean distances between all pairs of anatomical landmarks within the field of view of an image; the local mirror ratio is the ratio of the gray intensity value of the current pixel to the median gray intensity value of all pixels within the window.

3. The high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the cross-view luminance residual includes: randomly selecting a pixel in the current viewpoint RGB image as a reference pixel; based on the 3D point position corresponding to the reference pixel in the current viewpoint RGB image, the camera intrinsic parameter matrix, the extrinsic parameter pose of the current viewpoint, and the extrinsic parameter poses of other viewpoints, projecting the corresponding 3D point of the reference pixel onto the image plane of all other viewpoints to obtain the gray intensity value at the coordinates of the reprojected pixel; calculating the absolute difference between the gray intensity value of the reference pixel and the gray intensity value at the coordinates of the reprojected pixel under all other viewpoints; using the median of all absolute differences as the cross-view luminance residual of the reference pixel; and traversing all pixels under the current viewpoint to obtain the cross-view luminance residual of each pixel in the current viewpoint.

4. The high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the reliability metric coefficient includes: obtaining the reliability metric coefficient according to the reliability metric coefficient calculation formula, which is shown below: In the formula, Indicates the first View current pixel Reliability metric coefficient Indicates the first View current pixel The depth confidence value ranges from [0,1]. Indicates the current pixel. The local mirror ratio; Indicates the current pixel. The trans-spectral residual; Indicates the first The third quartile of the local specular ratio of all effective pixels in the view; Indicates the first The third quartile of the trans-view luminance residuals of all valid pixels in the RGB image under view.

5. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, Based on the reliability metric coefficient and the three-dimensional point positions, a triangular mesh is iteratively constructed to obtain the spatial position and geometric confidence of each vertex. This includes: constructing an initial triangular mesh, assigning initial spatial positions to each vertex of the initial triangular mesh, and determining the mesh topology; iterating over the initial triangular mesh, and in each iteration, performing the following process: using vertices... The reliability metric coefficients of all valid pixels within the spherical neighborhood space are used as weights to calculate the weighted average of the 3D point positions of each pixel within the spherical neighborhood space, thus obtaining the data item position. The radius of the spherical neighborhood space is taken as half of the median Euclidean distance between all adjacent vertices of the current grid. The position of the data item is calculated relative to the vertex. The arithmetic mean of the spatial positions of all directly connected adjacent vertices at the convergence of the previous iteration is used as the position of the neighborhood smoothing term; for each vertex... The cumulative reliability weight is obtained by summing the reliability metric coefficients of all valid pixels within the spherical neighborhood space; the median of the cumulative reliability weights of all vertices in the entire grid is used as the global threshold; the vertex is obtained according to the following formula. The current spatial position and geometric confidence level : ; In the formula, Represents vertices Cumulative reliability weight; Represents vertices The location of the data item; Indicates the global threshold; Represents vertices The neighborhood smoothing term position; after the full grid vertex synchronous update, when the maximum change in vertex spatial position between the current round and the previous round is less than the median of the weighted standard deviation of the effective pixels in the spherical neighborhood space of each vertex relative to its data term position, the iteration terminates; otherwise, the next round begins.

6. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 5, characterized in that, Constructing an initial triangular mesh, assigning initial spatial positions to each vertex of the initial triangular mesh, and determining the mesh topology includes: calculating a three-dimensional convex hull using the three-dimensional coordinates of all markers in the anatomical landmark set; using the area intercepted by the three-dimensional convex hull on the body surface manifold as the tiling range; and tiling the triangular mesh at a uniform density within the tiling range. The initial vertex spacing of the triangular mesh is set to three times the mean of the median of the Euclidean distances between the corresponding three-dimensional point positions of all effective adjacent pixels in the RGB images under all viewpoints. The position of each initial vertex is taken as the arithmetic mean of the three-dimensional point positions corresponding to all viewpoint pixels falling within its initial spherical neighborhood space, and the radius of the initial spherical neighborhood space is half of the initial vertex spacing.

7. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining binary burn classification includes: in the current session, using the burn wound color samples and healthy skin color samples marked by the clinician around the anatomical landmark as references, calculating the hue-saturation mean vector and covariance matrix of the two classes in the HSV color space respectively, and constructing a minimum Mahalanobis distance classifier; based on the spatial position of each vertex, the camera intrinsic parameter matrix, and the extrinsic pose of each viewpoint, projecting each vertex onto the image plane of each viewpoint to obtain the visible viewpoint; wherein, the visible viewpoint satisfies: the projected coordinates of each vertex fall within the effective pixel range of the image of that viewpoint, and the angle between the normal vector of the vertex and the optical center of the camera of that viewpoint does not exceed 90 degrees; sampling the color vector of each vertex in all visible viewpoints using bilinear interpolation and converting it to the HSV space, and the classifier outputs the binary discrimination result; the burn voting rate of the vertex is set as the ratio of the number of visible viewpoints judged as burns to the total number of visible viewpoints.

8. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the confidence level of the triangular facet burn includes: obtaining the confidence level of the triangular facet burn according to the calculation formula for the confidence level of the triangular facet burn, as shown below: In the formula, Indicates the confidence level for triangular facet burns; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices Burns voter turnout; Represents vertices ,vertex With vertex The minimum geometric confidence level.

9. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the expected value of the burn area includes: obtaining the expected value of the burn area according to the calculation formula for the expected value of the burn area, which is shown below: In the formula, This represents the expected burn area. Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; Indicates the preset number of triangular facet regions.

10. A high-precision three-dimensional reconstruction method for burn areas of burn patients according to claim 1, characterized in that, The method for obtaining the central estimate of TBSA% and the uncertainty interval includes: calculating the standard deviation of burn area, as shown in the following formula: In the formula, Indicates the standard deviation of burn area; Represents the triangular region The area; Represents the triangular region Confidence level of triangular burns; Represents the face set of a triangular mesh. Represent each triangle in the face set; calculate the total surface area of ​​the volume. The calculation formula is as follows: In the formula, This represents the total surface area of ​​the human body. Indicates the patient's height; This represents the patient's weight; the central estimate of the TBSA% is determined by the following formula: In the formula, This represents the central estimate of TBSA%. This represents the expected burn area; the uncertainty interval of TBSA% is determined by the following formula: In the formula, Indicates the TBSA% uncertainty range; This represents the expected burn area. Indicates the standard deviation of burn area; It represents the total surface area of ​​the human body.

Images