Intelligent measurement and recording system for emergency trauma wound area based on deep learning

Abstract

The application discloses an emergency trauma wound area intelligent measurement and recording system based on deep learning, relates to the field of computer vision, synchronously acquires a two-dimensional image sequence, a six-axis attitude vector and a time stamp, extracts a mask by using a segmentation network, and constructs a local curved surface geometric model in combination with prior parameters; a non-homogeneous weight compensation matrix is generated by calculating the included angle distribution of a space vector and an optical axis vector; target wound physical surface areas are acquired by performing pixel-by-pixel weighted integration on mask pixels based on the matrix, and the shrinkage caused by projection is compensated. According to the weight matrix gradient, confidence is evaluated, and dynamic acquisition guidance or structured data encapsulation is realized. The application effectively corrects non-homogeneous curved surface projection distortion, and improves measurement accuracy and the reliability of clinical records.

Description

Technical Field

[0001] This invention relates to the field of computer vision, specifically to an intelligent measurement and recording system for emergency trauma wound area based on deep learning. Background Technology

[0002] In the fields of clinical medicine and emergency care, quantitative analysis of tissue damage areas is a core component in assessing the severity of a patient's injury and monitoring healing progress. With the widespread adoption of digital medical technologies, automatically extracting damaged features from clinical images using computer vision algorithms and deep learning, and automating area calculation based on the statistical distribution of image units, has become a mainstream technological approach to improve clinical efficiency and ensure the objectivity of records.

[0003] In standard measurement logic, the geometric features presented in an image are usually considered to have a constant projection ratio with the actual physical scale of the target object. However, due to the complex spatial morphology of the human anatomy, the wound area to be measured is often distributed on non-flat surfaces with different curvatures. This intervention of three-dimensional geometric features alters the linear relationship in the imaging process. When mobile image acquisition is performed using handheld devices, a multi-dimensional pose deviation inevitably occurs between the optical axis of the imaging sensor and the undulating damaged surface. This spatial pose uncertainty is coupled with the surface undulation features, resulting in a severe geometric projection shrinkage effect during imaging. This causes the distribution of units recorded on the image plane to lose its correspondence with the actual physical surface area. As the local slope of the surface increases, the feature regions extracted from the image undergo non-homogeneous compression, leading to deviations in the final calculation results and an inability to accurately reproduce the physical scale of the damaged area. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent measurement and recording system for emergency trauma wound area based on deep learning. This system has the advantages of reducing the interference of mobile acquisition terminal pose deviation and surface curvature of complex anatomical sites on the consistency of wound projection imaging, improving the accuracy of wound edge topological positioning and physical surface area measurement, thereby enhancing the reliability of clinical quantitative assessment results of emergency trauma.

[0005] The objective of this application can be achieved through the following technical solution: a deep learning-based intelligent measurement and recording system for emergency trauma wound area, comprising the following modules:

[0006] The data synchronization acquisition module is used to synchronously acquire two-dimensional color image sequence data, six-axis spatial attitude vector data and measurement timestamp data of the wound area to be measured.

[0007] The feature modeling module is used to input the two-dimensional color image sequence data into a preset segmentation network to extract the wound mask pixel data of the wound area to be tested; and to construct a local surface geometric model of the corresponding wound area to be tested based on the six-axis spatial pose vector data and preset prior parameters of human body parts.

[0008] The projection analysis module is used to calculate the angle distribution between the spatial normal vector and the camera optical axis vector of each sampling point on the local curved surface geometric model, and to map the angle distribution to the image coordinate system to generate a non-homogeneous weight compensation matrix corresponding to the wound mask pixel data.

[0009] The area fitting correction module is used to perform a pixel-by-pixel weighted integral operation on the wound mask pixel data according to the non-homogeneous weight compensation matrix, and calculate the physical surface area of ​​the target wound by compensating for the pixel shrinkage caused by the curved surface projection.

[0010] The evaluation feedback module is used to calculate the measurement confidence score of the current wound mask pixel data based on the gradient change rate of the non-homogeneous weight compensation matrix; if the measurement confidence score is less than a preset measurement threshold, then acquisition guidance instruction data is generated; if the measurement confidence score is greater than or equal to the preset measurement threshold, then the target wound physical surface area value, the two-dimensional color image sequence data and the measurement timestamp data are associated and encapsulated to generate structured measurement record data.

[0011] Compared with the prior art, the beneficial effects of this application are:

[0012] 1. This invention constructs a local curved surface geometric model by fusing a six-axis spatial attitude vector with prior parameters of human body parts, and generates a non-homogeneous weight compensation matrix using the angle distribution between the spatial normal vector and the camera optical axis vector, thus achieving pixel-by-pixel compensation for pixel shrinkage. This corrects the perspective shrinkage effect, resulting in improved area measurement accuracy for high-curvature areas such as limbs and the head compared to traditional two-dimensional planar measurement methods.

[0013] 2. This invention extracts the minimum distortion pixel features in the temporal dimension through a multi-frame orthophoto fusion module, and combines it with a topological constraint repair module to perform pixel trajectory interpolation to complete the feature fracture region using local curvature features. This reduces the impact of drone or handheld terminal attitude jitter and foreign object occlusion on image quality, ensuring the continuity of edge localization and the reliability of area measurement in complex clinical environments.

[0014] 3. This invention analyzes the gradient change rate of the weight compensation matrix through an evaluation feedback module, calculates the measurement confidence level in real time, and generates acquisition guidance instructions, thereby achieving dynamic monitoring of image acquisition quality. This real-time feedback mechanism, combined with the association and encapsulation of measurement timestamps and structured data, not only reduces the statistical bias and subjectivity caused by manual measurement but also improves the efficiency of automated recording of clinical wound data and the traceability of the medical process, providing scientific data support for the precise diagnosis and treatment of emergency trauma. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the module of the deep learning-based intelligent measurement and recording system for emergency trauma wound area in this application. Detailed Implementation

[0016] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0017] Example 1,

[0018] Application Overview:

[0019] In the fields of clinical medicine and emergency care, quantitative analysis of tissue damage areas is a core component in assessing the severity of a patient's injury and monitoring healing progress. With the widespread adoption of digital medical technologies, using computer vision algorithms to automatically extract damaged features from clinical images and to automatically calculate the area based on the statistical distribution of image units has become a mainstream technological approach to improve clinical efficiency and ensure the objectivity of records.

[0020] However, due to the complex spatial morphology of the human anatomy, the wound area to be measured is often distributed on non-flat surfaces with different curvatures. This intervention of three-dimensional geometric features alters the linear relationship in the imaging process. When mobile image acquisition is performed using handheld devices, a multi-dimensional pose deviation inevitably occurs between the optical axis of the imaging sensor and the undulating damaged surface. This spatial pose uncertainty, coupled with the surface undulations, leads to a severe geometric projection shrinkage effect during imaging, causing a loss of direct correspondence between the unit distribution recorded on the image plane and the actual physical surface area. As the local slope of the surface increases, the feature regions extracted from the image undergo non-homogeneous compression, resulting in deviations in the final calculation results and an inability to accurately reconstruct the physical scale of the damaged area. How to solve the mapping distortion problem between the image unit distribution and the actual physical surface area of ​​the target object caused by spatial projection deviation in non-homogeneous curved surface environments has become a key bottleneck restricting the achievement of high-precision measurements in related fields.

[0021] If the aforementioned problems are not addressed, traditional area measurement methods based on the assumption of planar projection will continue to output results with systematic biases, leading to inaccurate emergency injury assessment and treatment decisions. For patients with large-area burns, if the wound area is underestimated by more than 10%, the fluid resuscitation plan, skin grafting plan, and nutritional support dosage will all be inaccurate, threatening the patient's life. Fixed geometric projection models cannot capture real-time changes in posture and angle during handheld data acquisition, and cannot establish a dynamic correlation between the compensation matrix and measurement confidence, thus creating a negative feedback loop that cannot self-correct, affecting the objectivity and traceability of clinical records.

[0022] To address the aforementioned challenges, this application first considers establishing a dynamic compensation mechanism between pixel distribution in the image coordinate system and the actual physical surface area. Traditional methods calculate area using a fixed projection ratio, which cannot adapt to non-homogeneous geometric deformations in curved environments. To resolve this, this application attempts to introduce six-axis spatial attitude vector data and prior parameters of human body parts. By constructing a local curved surface geometric model, it captures the three-dimensional undulations of the wound area to be measured. Then, for each pixel, it calculates the angle between its corresponding normal vector and the camera's optical axis vector, thereby generating a non-homogeneous weight compensation matrix that quantifies the degree of projection shrinkage. Further analysis reveals that relying solely on compensation calculations for a single frame image is insufficient to guarantee measurement reliability in highly distorted regions. Therefore, a confidence assessment mechanism and acquisition guidance feedback are needed to achieve closed-loop self-correction in the measurement process. Through this systematic design, the area measurement results can adaptively adjust in real-time according to the acquisition attitude, thus overcoming the technical shortcomings of traditional methods where projection deviations and geometric deformations cannot be compensated.

[0023] like Figure 1As shown, the intelligent measurement and recording system for emergency trauma wound area based on deep learning includes the following modules:

[0024] The data synchronization acquisition module is used to synchronously acquire two-dimensional color image sequence data, six-axis spatial attitude vector data, and measurement timestamp data of the wound area to be measured; it is a component used to synchronously acquire multimodal sensing data from the imaging sensor and the inertial measurement unit. Specifically, it can be implemented by combining the image acquisition unit and the six-axis IMU sensor with a hardware-triggered synchronization mechanism. Its function is to ensure the temporal consistency between the image frame and the corresponding spatial pose data, so as to provide accurate input for subsequent geometric modeling and compensation calculations.

[0025] The feature modeling module is used to input the two-dimensional color image sequence data into a preset segmentation network to extract the wound mask pixel data of the wound area to be tested; based on the six-axis spatial pose vector data and preset prior parameters of human body parts, it constructs a local surface geometric model of the corresponding wound area to be tested; the feature modeling module is the core processing component for joint analysis of acquired images and pose data, which can be implemented by combining semantic segmentation neural networks and pose-driven surface parameterization methods. Its role is to simultaneously complete the pixel-level localization of the wound and the reconstruction of the three-dimensional geometry of the wound area to be tested, laying the foundation for the accurate quantification of projection error.

[0026] The projection analysis module is used to calculate the angle distribution between the spatial normal vector and the camera optical axis vector of each sampling point on the local curved surface geometric model, and to map the angle distribution to the image coordinate system to generate a non-homogeneous weight compensation matrix corresponding to the pixel data of the wound mask. The projection analysis module is a key calculation component for quantifying the geometric error of the surface projection. Specifically, it can be implemented by discretization normal vector field calculation and angle cosine mapping method. Its function is to transform the geometric tilt information of the three-dimensional surface into pixel-by-pixel area compensation weights in the image domain, driving the accurate execution of subsequent area correction calculations.

[0027] The area fitting correction module is used to perform a pixel-by-pixel weighted integral operation on the wound mask pixel data according to the non-homogeneous weight compensation matrix. By compensating for the pixel shrinkage caused by the curved surface projection, the physical surface area of ​​the target wound is calculated. The area fitting correction module is the final calculation unit that integrates the compensation weights and pixel statistics into the physical area value. Specifically, it can be implemented by combining a pixel-by-pixel multiplication and weighted summation algorithm with a pre-calibrated pixel unit step size. Its function is to reduce the systematic area underestimation bias caused by the curved surface projection and output the real wound area with physical meaning.

[0028] The evaluation feedback module calculates the measurement confidence score of the current wound mask pixel data based on the gradient change rate of the non-homogeneous weight compensation matrix. If the measurement confidence score is less than a preset measurement threshold, acquisition guidance instruction data is generated. If the measurement confidence score is greater than or equal to the preset measurement threshold, the target wound physical surface area value, the two-dimensional color image sequence data, and the measurement timestamp data are associated and encapsulated to generate structured measurement record data. The evaluation feedback module is an adaptive control component for achieving closed-loop management of the measurement process. Specifically, it can be implemented using compensation matrix difference analysis and a confidence inverse proportional mapping method. Its function is to quantitatively evaluate the reliability of the current measurement results and drive acquisition guidance feedback when the confidence is insufficient, thereby ensuring the quality and integrity of the final archived data.

[0029] The core innovation of this application lies in the surface projection error correction process driven by a non-homogeneous weight compensation matrix, which integrates spatial pose information with prior human geometric knowledge into pixel-by-pixel area compensation weights, and then dynamically controls the correction coefficients in the area integration process, so that the measurement results can be adaptively adjusted according to the actual acquisition posture and surface shape, thus solving the technical defect that the traditional planar projection assumption method cannot adapt to geometric deformation in curved surface environments.

[0030] This application further proposes that the specific steps for simultaneously acquiring two-dimensional color image sequence data, six-axis spatial attitude vector data, and measurement timestamp data of the wound area to be measured include:

[0031] Two-dimensional color image sequence data refers to a collection of RGB format digital images acquired frame by frame at a fixed frame rate within a continuous time step by an image acquisition unit. Specifically, this can be achieved using a camera module combining a CMOS image sensor and an optical distortion correction lens. The resolution of each frame in the two-dimensional color image sequence data is W×H (width×height, unit: pixels). Each pixel consists of quantized grayscale values ​​from three channels: red (R), green (G), and blue (B), with a quantization bit depth of 8 bits (i.e., each channel's value range is 0–255). The imaging focal length (unit: millimeters), pixel size (unit: millimeters / pixel), and principal point coordinates of the image acquisition unit together constitute the camera's intrinsic parameter matrix. This matrix describes the projection mapping relationship between points in three-dimensional space and pixels in the two-dimensional image.

[0032] The camera intrinsic parameter matrix is ​​predetermined by a checkerboard calibration experiment when the device leaves the factory and is stored in the system firmware as a preset calibration parameter for subsequent projection analysis module to perform mutual conversion between pixel coordinates and spatial direction vectors.

[0033] Six-axis spatial attitude vector data refers to the joint vector of three-axis acceleration components and three-axis angular velocity components synchronously output by a six-axis inertial measurement unit (IMU) at each acquisition moment.

[0034] The six-axis IMU sensor continuously outputs the aforementioned six-axis data at a preset sampling rate. At each sampling moment, a set of six-dimensional real vectors is generated, forming a six-axis spatial attitude vector data sequence. After the six-axis spatial attitude vector data is processed by the attitude calculation algorithm, a quaternion or rotation matrix describing the current spatial orientation of the device is generated, which is used for subsequent pose mapping of the local surface geometry model.

[0035] To achieve complementary fusion of three-axis acceleration and three-axis angular velocity data, and thus accurately estimate the absolute spatial attitude of the device, this application employs a complementary filtering algorithm to fuse and solve the six-axis data.

[0036] Measurement timestamp data refers to the absolute time stamp recorded by the system for each frame of image acquisition. Specifically, it can be represented by an integer value with millisecond precision using the UTC (Coordinated Universal Time) standard.

[0037] The measurement timestamp data has a dual function in the system: first, it is used for the temporal alignment of image frames and IMU data, ensuring that each frame of image accurately corresponds to a set of six-axis attitude data through timestamp interpolation; second, as a metadata field of structured measurement record data, it supports the clinical recording system to perform temporal retrieval and comparison of healing progress of measurement results of the same patient at different times.

[0038] To ensure the synchronous acquisition of three data streams (image sequence, six-axis attitude vector, and timestamp), this application adopts a hardware-triggered synchronization mechanism: the exposure trigger signal of the image acquisition unit is simultaneously sent to the sampling latch pin of the six-axis IMU sensor to ensure that image exposure and IMU data latching are completed at the same physical moment.

[0039] In a preferred embodiment, when acquiring images of the right forearm wound of an emergency burn patient, the image acquisition unit continuously acquires 30 frames of RGB images at a frame rate of 30fps, with each frame having a resolution of 1920×1080, and the acquisition time is approximately 1.17 seconds. A six-axis IMU synchronously latches the data at a sampling rate of 200Hz, recording a total of 233 sets of six-axis vector data. Timestamps are recorded starting at 1718000000000ms (corresponding to a certain UTC time), with an interval of 33.3ms between adjacent frames, for a total of 35 timestamps. Hardware trigger signals bind the 35 exposure triggers to the corresponding IMU data latches, with a measured maximum time deviation of 3.2ms. After the three data streams are synchronously bound, they are encapsulated and entered into the feature modeling module.

[0040] Through the above technical solution, this application achieves precise synchronous acquisition of multimodal sensing data, ensuring that each frame of image and its corresponding spatial pose data are strictly consistent in the temporal dimension, eliminating the pose estimation error introduced by asynchronous acquisition. Compared with the software timestamp alignment scheme, the hardware-triggered synchronization mechanism compresses the synchronization error from tens of milliseconds at the software scheduling level to the physical limit of the IMU sampling interval (5ms), providing a timing guarantee for the high-precision construction of the subsequent local surface geometric model, ensuring the stability of pose estimation under both rapid device movement and long-term continuous use, and providing a reliable source of spatial pose data for the entire area compensation link. In combination with the overall system solution, the data synchronization acquisition module is the source of accuracy guarantee for the entire link. Its timing accuracy directly determines the construction accuracy of the surface geometric model in the subsequent feature modeling module, thus fundamentally affecting the accuracy of the non-homogeneous weight compensation matrix generation and the reliability of the final physical surface area value.

[0041] This application further proposes that the specific steps for extracting the wound mask pixel data of the wound area to be tested include:

[0042] The segmentation network is invoked to perform pixel-level probability mapping on the two-dimensional color image sequence data, identifying the target pixel coordinate set that conforms to the wound texture distribution characteristics. The segmentation network used in this application introduces a multi-scale feature fusion module (such as dilated spatial convolutional pooling, ASPP) at the encoder end. Its input layer not only contains the original RGB image but also introduces a and b channel data after color constancy processing through cascading, to enhance the system's ability to distinguish between blood-infiltrated areas and skin erythema areas. During model training, a weighted combination of Dice Loss and Focal Loss is used as the loss function to address the extremely imbalanced problem of emergency wounds, which constitute a very small proportion of the entire image. The model's output layer effectively captures the blurred diffusion characteristics of the wound edge by performing probability regression on each pixel.

[0043] Wound texture distribution features include color channels (e.g., the red channel response of the wound area in RGB space is higher than that of the surrounding normal skin), texture gradient features (gradient amplitude is concentrated at the wound edge), and morphological connectivity constraints. The segmentation network extracts the coordinates of all target pixels that satisfy the wound feature constraints by comprehensively judging the above multi-dimensional features, and summarizes them to form a set of target pixel coordinates.

[0044] A binarized feature matrix with the same resolution as the two-dimensional color image sequence data is generated based on the target pixel coordinate set, and this binarized feature matrix is ​​used as the wound mask pixel data. The binarized feature matrix is ​​a two-dimensional integer matrix with the same width and height as the input image. Specifically, it can be achieved by applying a 0.5 probability threshold to the probability heatmap: for pixel coordinates with a probability value greater than the threshold, the corresponding position in the matrix is ​​assigned a value of 1 (representing the wound area); for pixel coordinates with a probability value less than or equal to the threshold, a value of 0 is assigned (representing the background area). This matrix records the two-dimensional pixel distribution of the wound with precise spatial correspondence, providing a pixel-by-pixel mask index basis for subsequent projection compensation operations. In practical implementation, morphological opening operations (erosion followed by dilation) can also be performed on the binarized feature matrix to reduce the probability of isolated pixel interference caused by image noise in the segmentation result, and improve the smoothness and connectivity of the mask edges.

[0045] In a preferred embodiment, the segmentation network infers from a 1920×1080 resolution emergency wound image and outputs a probability heatmap of the same size. Pixel coordinates with a probability value greater than 0.5 in the heatmap are extracted, identifying a total of 47,832 target pixel coordinates, forming a target pixel coordinate set. A 1920×1080 binarized feature matrix is ​​generated based on this set, with 47,832 coordinate positions assigned a value of 1 and the remaining positions assigned a value of 0. After morphological opening processing, 317 isolated noise pixels are eliminated, resulting in wound mask pixel data with an area of ​​47,515 pixels, which is used for subsequent area compensation calculations.

[0046] Through the above technical solution, this application achieves precise localization of wound regions based on deep learning semantic segmentation. Compared with traditional segmentation methods based on color thresholds or manual features, the segmentation network can adapt to the visual features of different wound types (such as burns, lacerations, and abrasions), maintaining stable segmentation accuracy even in emergency scenarios with significant changes in lighting and skin color. The binarized feature matrix, as a unified representation format for wound mask pixel data, transforms the segmentation results into a standardized data structure suitable for subsequent matrix operations while maintaining pixel-level spatial resolution. This provides an efficient data interface for the dot product operation between the non-homogeneous weight compensation matrix and the mask data. In the overall system solution, accurate mask extraction is a direct determinant of the accuracy of subsequent area compensation. The accuracy of the mask boundary directly affects the correct allocation of edge pixel area coefficients, thus directly impacting the calculation error of the final physical surface area value.

[0047] This application further proposes that the specific steps for constructing a local surface geometric model corresponding to the wound area to be tested include:

[0048] Limb edge contour pixels refer to a set of pixel coordinates that describe the outer contour of a limb in an image. Specifically, they can be extracted using the Canny edge detection algorithm or a deep learning-based limb segmentation network.

[0049] A spatial transformation matrix is ​​constructed using the six-axis spatial attitude vector data. Preset prior parameters for human body parts are then mapped to the pose using this spatial transformation matrix to generate a local surface geometric model describing the physical undulations of the wound area under test. The spatial transformation matrix is ​​a 4×4 homogeneous transformation matrix that describes the spatial relationship between the camera coordinate system and the standard anatomical reference system. Specifically, it can be generated by quaternion fusion of the acceleration and angular velocity components output by the six-axis IMU to produce the corresponding rotational components, which are then combined with the known installation offset parameters of the device to synthesize the complete transformation matrix. The prior parameters for human body parts refer to a set of standard geometric parameters for each part pre-calibrated based on anatomical statistics. Specifically, this includes the statistical range of the major and minor axis radii of the ellipsoids for each part of the limbs (e.g., the average minor axis of the adult forearm is 25mm, and the average major axis is 38mm) and the corresponding natural curvature distribution. The pose mapping using the spatial transformation matrix transforms the standard prior parameters to a coordinate system consistent with the current image viewpoint, thereby generating a local surface geometric model reflecting the physical undulations of the wound area under test under actual acquisition conditions. The model is stored in the form of parametric surfaces (such as triaxial ellipsoids or non-uniform rational B-spline surfaces) and supports normal vector lookups for arbitrary image coordinates.

[0050] In a preferred embodiment, when processing an image containing a wound on a patient's left forearm, the system first performs elliptical fitting on the edge pixels of the limb in the image, identifying that the limb's cross-sectional contour is elliptical. Prior parameters for the forearm ellipsoid (minor axis radius 24mm, major axis radius 36mm, natural curvature 0.04 / mm) are retrieved from a priori parameter library. The quaternion output by the six-axis IMU is... The rotation matrix R is generated by solving the rotation matrix. After mapping the prior parameters of the ellipsoid through the rotation matrix, a local surface geometry model under the current viewpoint is generated. In this model, the angle between the surface normal vector in the major axis direction of the forearm and the optical axis direction of the camera is about 12° in the central region of the image and expands to 38° in the left and right edge regions of the image, reflecting the different distribution of surface projection shrinkage at different locations.

[0051] Through the above technical solution, this application achieves rapid construction of local surface geometric models based on anatomical prior knowledge and real-time spatial pose data, enabling the estimation of 3D surface morphology on a handheld device without the need for a depth sensor. Compared with 3D reconstruction methods relying on structured light or ToF depth cameras, this solution reduces hardware costs by utilizing existing statistical prior knowledge of human anatomy, while avoiding the risk of depth sensor failure in bright light emergency environments. Combined with the overall system solution, an accurate local surface geometric model is the direct determinant of the accuracy of the non-homogeneous weight compensation matrix generation. The degree to which the model restores the curvature distribution directly determines the accuracy of the normal vector calculation, thus fundamentally affecting the error propagation of the entire area compensation chain and improving the adaptive compensation capability for wounds at different anatomical locations.

[0052] This application further proposes that the specific steps for generating the non-homogeneous weight compensation matrix include:

[0053] The normal vectors corresponding to each pixel coordinate point on the local surface geometric model are calculated, and the camera optical axis vector is determined based on the six-axis spatial attitude vector data. The normal vector is a unit direction vector perpendicular to the tangent plane at a sampling point on the surface model and pointing outwards from the surface. Specifically, it can be calculated through partial differential operations on the parameterized surface model (i.e., taking the cross product of the partial derivatives of the surface parametric equations with respect to the two parameter directions and then normalizing). Each image coordinate point corresponds to a unique sampling point on the surface model, thus establishing a one-to-one mapping relationship between pixel coordinates and the normal vector. The camera optical axis vector is a unit direction vector extending along the imaging principal axis in the camera coordinate system. Specifically, it can be obtained through joint calculation of the camera intrinsic parameter matrix and the current six-axis IMU attitude solution. The spatial attitude vector output by the six-axis IMU includes three-axis acceleration and three-axis angular velocity. After being solved into quaternions through complementary filtering, it can accurately describe the camera's current orientation in the world coordinate system, thereby determining the direction of the principal optical axis in the world coordinate system.

[0054] The cosine of the angle between each normal vector and the camera optical axis vector is calculated, and the reciprocal of the cosine is defined as the pixel area compensation coefficient. This generates a non-homogeneous weight compensation matrix that matches the dimension of the wound mask pixel data. The cosine of the angle is obtained by the dot product of the normal vector and the camera optical axis vector: for an ideal orthophoto region where the curved surface tangent plane is perpendicular to the camera optical axis, the normal vector and optical axis direction are completely aligned, the cosine value is equal to 1, and the compensation coefficient is 1, indicating that the pixel does not require compensation; for a sloping curved surface region, as the angle θ increases, the cosine value cos(θ) decreases, and the compensation coefficient 1 / cos(θ) is greater than 1, indicating that the pixel's projection on the image plane is compressed and needs to be magnified to restore its corresponding true physical area contribution. The dimension of the non-homogeneous weight compensation matrix is ​​completely consistent with the wound mask pixel data (both are M×N). Each element in the matrix records the pixel area compensation coefficient at the corresponding coordinate position, thus forming a complete compensation field describing the spatial distribution of projection shrinkage in the entire image.

[0055] As a preferred embodiment, taking the forearm wound acquisition scenario as an example, the local curved surface geometric model in the image coordinates The corresponding normal vector is (Direction of normal illumination), the camera's optical axis vector is The dot product is 1, and the compensation coefficient is 1.00. (In the image coordinates...) At that point, the corresponding normal vector rotates to The angle between the coordinates and the optical axis is approximately 40°, with a cosine value of 0.77. The compensation coefficient is 1 / 0.77≈1.30, meaning that the projected area at this location is only 77% of the actual area, requiring a magnification of 1.30 times to restore the true value. Performing the above calculation on all pixel coordinates of the entire 1920×1080 image generates a non-homogeneous weight compensation matrix of the same size. The compensation coefficient in the central region of the matrix is ​​approximately 1.00, while the compensation coefficient in the edge regions can reach over 1.40.

[0056] It should be noted that, in order to prevent numerical instability caused by the compensation coefficient tending to infinity at extremely large angles at the edge of the curved surface, the system presets an included angle threshold. (For example (or maximum compensation weight limit) (For example When the calculation yields... When the value is less than the preset threshold, the system forces the compensation coefficient to equal the preset threshold. Alternatively, the pixel may be marked as an unreliable edge. This boundary suppression mechanism ensures that the physical surface area value output by the system is statistically robust under complex shooting angles in the emergency room.

[0057] Through the above technical solution, this application achieves accurate calculation of pixel-by-pixel area compensation coefficients based on the principle of geometric optical projection, completely transforming the spatial tilt information of the three-dimensional surface into a non-homogeneous compensation weight field in the two-dimensional image domain. Compared with the traditional globally uniform proportional compensation method, the non-homogeneous weight compensation matrix can accurately describe the differences in projection shrinkage at different locations on the surface, avoiding underestimation of the area caused by excessive projection compression in the edge region and overestimation of the area caused by overcompensation in the central region. Combined with the overall system solution, the non-homogeneous weight compensation matrix is ​​the core data structure connecting the local surface geometric model and the area integration operation. Its calculation accuracy directly determines the upper limit of the systematic error of the final physical surface area value, and also provides a direct data source for the gradient analysis of the subsequent measurement confidence score.

[0058] This application further proposes that the specific steps for calculating the physical surface area of ​​the target wound include:

[0059] Acquire preset pixel unit step size data. Pixel unit step size data refers to the actual physical size (unit: mm / pixel) of each pixel on the image sensor at a specific acquisition distance. This can be obtained through camera calibration experiments or proportional calculations based on known-sized reference objects, and dynamically corrected using a lookup table with the acquisition distance as the independent variable. The accuracy of the pixel unit step size data directly determines the baseline scaling factor when converting pixel counts to physical area values. Therefore, it needs to be rigorously calibrated at the factory and continuously corrected in real-time during subsequent use in conjunction with autofocus distance estimation.

[0060] Each pixel in the wound mask pixel data of the area to be tested is multiplied by the pixel area compensation coefficient at the corresponding coordinate in the non-homogeneous weight compensation matrix to obtain the corrected sub-pixel area unit. The dot product operation is an element-wise multiplication of two matrices with the same dimension, specifically implemented using the Hadamard product: after performing the Hadamard product operation on the wound mask pixel data (a binary matrix with values ​​of 0 or 1) and the non-homogeneous weight compensation matrix (a floating-point matrix with values ​​greater than or equal to 1), only pixels with a mask value of 1 retain the corresponding compensation coefficient value, while the result for the background area (mask value of 0) remains 0, thus achieving accurate extraction of the wound area compensation coefficient. Each non-zero result value represents the relative area contribution of that pixel on the real surface of the curved surface, and is called the corrected sub-pixel area unit.

[0061] The sum of all sub-pixel area units is calculated, and the physical surface area of ​​the target wound is obtained by combining the pixel unit step size data. The total area of ​​all sub-pixel area units is summed to obtain the compensated equivalent pixel area. This total is then multiplied by the square of the pixel unit step size data (because the dimension of area is the square of length). The dimension of the total obtained by summing the sub-pixel area units is the dimensionless equivalent pixel number. After multiplying by the square of the pixel unit step size (unit: mm / pixel), the dimension is converted to mm², thus obtaining the physical surface area of ​​the target wound (unit: mm² or cm²).

[0062] In a preferred embodiment, based on the aforementioned 47,515-pixel wound mask, after performing a Hadamard product operation with the non-homogeneous weight compensation matrix, the sum of the area units of each sub-pixel is 55,243.6 (equivalent number of compensated pixels). With a pixel unit step size corrected for distance to 0.18 mm / pixel, the physical surface area of ​​the target wound is 55,243.6 × (0.18)² ≈ 1,790 mm² ≈ 17.90 cm². Without applying the compensation matrix, directly multiplying 47,515 pixels by the square of the step size yields 1,539 mm² ≈ 15.39 cm², resulting in a systematic underestimation bias of approximately 14%, verifying the necessity of the non-homogeneous weight compensation matrix for improving measurement accuracy.

[0063] Through the above technical solution, this application achieves a precise conversion from discrete pixel counting to continuous physical area values, fully reflecting the compensation effect of the non-homogeneous weight compensation matrix in the final area output through pixel-by-pixel multiplication and global summation operations. Compared to the traditional calculation method of directly counting the number of mask pixels and then multiplying by a fixed ratio, the introduction of sub-pixel area units ensures that each wound pixel contributes its actual physical area to the accumulation, thereby reducing the non-uniform shrinkage error caused by curved surface projection. Combined with the overall system solution, accurate physical area values ​​are the core output driving subsequent confidence assessment and structured recording. Their accuracy directly affects clinical injury grading (such as calculating the percentage of burn area to body surface area), fluid resuscitation plan formulation, and longitudinal comparative analysis of treatment effects, and has direct clinical application value.

[0064] This application further proposes that the specific steps for calculating the measurement confidence score include:

[0065] First-order or second-order differencing operations are performed on the non-homogeneous weight compensation matrix to extract gradient feature maps describing the degree of weight abrupt changes. First-order differencing calculates the difference between adjacent elements in the matrix. Specifically, the Sobel or Prewitt operators can be used to perform convolution operations on the compensation matrix in the horizontal and vertical directions, respectively, and then the gradient magnitude is taken to quantify the rate of change of the compensation coefficients in space. Second-order differencing (such as the Laplacian operator) can further amplify the weight jumps caused by local curvature abrupt changes, making it more sensitive to regions of extreme distortion. The value of each pixel in the gradient feature map reflects the degree of spatial change in the compensation weights at that location: regions where the compensation weights transition smoothly in space (corresponding to regions with gentle surface curvature and stable projection distortion) have smaller gradient values; regions where the compensation weights change abruptly in space (corresponding to regions where the angle between the local surface normal vector and the optical axis changes drastically and the projection height is unstable) have larger gradient values.

[0066] Connected regions in the gradient feature map whose values ​​exceed a preset rate of change threshold are identified and designated as high-distortion masks. The preset rate of change threshold is a scalar value pre-set by the system during the calibration phase based on the statistical distribution of distortion features in historical data. Specifically, it can be set to the 90th percentile of the global distribution of the gradient feature map; pixels with gradient values ​​higher than this threshold are identified as regions experiencing projection distortion. Connected region extraction and merging are performed on all pixels meeting the threshold condition using a connected component analysis algorithm (such as the 8-neighborhood labeling algorithm), ultimately forming a high-distortion mask that records the spatial distribution of high distortion.

[0067] The percentage of the area of ​​the high-distortion mask relative to the pixel data of the wound mask is calculated, and this percentage is inversely mapped to a measurement confidence score. The area percentage is obtained by dividing the number of pixels in the high-distortion mask by the number of pixels in the wound mask. A higher ratio indicates a larger proportion of pixels in the wound area affected by severe projection distortion, and a lower reliability of the current measurement result. The inverse mapping relationship can be defined as: Where K is the normalization coefficient, ensuring that the confidence score is within the range specified in the original text. The distribution is linear within the interval. When the percentage of the high-distortion mask area is zero, the confidence score is 1.0 (full marks); when the high-distortion mask completely covers the wound mask, the confidence score approaches 0.

[0068] In a preferred embodiment, a Sobel first-order difference operation is performed on the compensation matrix of the forearm wound scene to generate a 1920×1080 gradient feature map, where the 90th percentile gradient threshold is 0.15. After identifying connected regions with gradient values ​​exceeding 0.15, the high-distortion mask contains a total of 6,823 pixels, accounting for 14.4% of the 47,515 pixels of the wound mask.

[0069] Substituting into the inverse proportional mapping formula, K takes the value of 1. If the score is higher than the preset measurement threshold of 0.80, the current measurement result is deemed reliable, triggering the structured record encapsulation process.

[0070] Through the above technical solution, this application realizes a quantitative assessment of measurement confidence based on compensation matrix gradient analysis. This transforms the measurement uncertainty, originally implicit in geometric modeling errors, into a quantifiable and comparable confidence scalar, providing a reliable basis for automated monitoring of measurement quality. Compared to traditional quality control methods that rely on operator subjective judgment, the confidence score objectively and in real-time reflects the degree of matching between the current acquisition posture and the surface morphology, reducing subjective errors in manual assessment. In the overall system design, confidence assessment is a crucial link connecting area calculation output and acquisition quality control: only measurement results that pass the confidence threshold test are stored in structured records, ensuring that all archived records in the clinical database have verifiable quality assurance, thus improving the reliability of the system output and the safety of clinical applications.

[0071] This application further proposes that the specific steps for generating the acquisition guidance instruction data include:

[0072] Locate the centroid coordinates of the high-distortion mask in the image coordinate system, and calculate the offset vector of the centroid coordinates relative to the image center point. The centroid coordinates refer to the geometric center position of all pixel coordinates in the high-distortion mask, which can be calculated by averaging the x and y coordinates of all pixels in the high-distortion mask. , Where N is the total number of pixels contained in the high-distortion mask. The image center point refers to the standard reference coordinate located at half the width and height of the image in the image coordinate system. For a 1920×1080 resolution image, the image center point is... .

[0073] The offset vector is calculated by subtracting the image center coordinates from the centroid coordinates:

[0074] Offset vector = (centroid x - 960, centroid y - 540); the direction of the offset vector indicates which side the current high distortion region is biased towards in the image, and its magnitude indicates the degree of deviation from the center. Together, they determine the basis for the calculation of the direction and intensity of the guidance adjustment.

[0075] The offset vector is converted into directional guidance feature image data, which serves as the acquisition guidance instruction data. The directional guidance feature image data is a visual guidance graphic superimposed on the real-time image preview interface. Specifically, it can be implemented using a dynamic directional arrow image with the image center as the starting point and the normalized offset vector direction as the arrow's direction. The offset vector is first normalized so that the arrow's direction information is retained only; the length or thickness of the arrow can increase with the increase of the offset vector amplitude, forming an intuitive feedback of the guidance force. Its physical meaning is: the area of ​​high distortion concentrated on a particular side of the image indicates that the surface tilt angle on that side is too large. Medical personnel need to adjust the acquisition device in the opposite direction of that direction, that is, translate or rotate the device along the opposite direction of the offset vector, so that the camera's optical axis is as perpendicular as possible to the current wound surface normal vector, thereby reducing the degree of projection distortion in that area. The guidance instructions can also include text prompts, such as: Please move the device to the left or please rotate it 15° clockwise, to further reduce the operational threshold.

[0076] As a preferred embodiment, the centroid coordinates of the high-distortion mask in the current frame are calculated as follows: The center of the image is The offset vector is The direction is the positive x-axis (right side of the image). The system draws a guide arrow pointing to the right in the center of the preview interface, with a length corresponding to a normalized medium intensity of 320 pixels, and adds the text prompt: Please move the device to the right or adjust the optical axis angle. After the medical staff adjusts the device to the right as prompted, the centroid of the high-distortion mask shifts to the left. The offset vector magnitude is reduced to approximately 60 pixels, the confidence score is increased to 0.84, exceeding the threshold of 0.80, the boot process is terminated, and the system enters the recording and encapsulation process.

[0077] Through the above technical solution, this application realizes a real-time acquisition guidance function based on spatial distortion geometric analysis, transforming the projection distortion distribution, which was originally invisible to medical staff, into intuitive operation prompts, filling the gap in the acquisition quality guidance stage of traditional measurement systems. Compared with manual adjustment methods that rely on operational experience, the guidance instructions can accurately calculate the adjustment direction, enabling ordinary medical staff lacking three-dimensional spatial perception training to complete high-quality acquisition in the shortest possible time. Combined with the overall system solution, the acquisition guidance function and the confidence assessment module together constitute a closed-loop self-correction mechanism for measurement quality: the confidence assessment is responsible for identifying the state of insufficient measurement quality, and the acquisition guidance is responsible for guiding the operator to correct the acquisition posture. The two work together to ensure that every measurement data entering the structured recording meets the preset accuracy threshold, thereby ensuring the overall measurement accuracy of the system without increasing the operational skill threshold.

[0078] This application further proposes that it also includes:

[0079] The non-homogeneous weight compensation matrix corresponding to each frame in the two-dimensional color image sequence data is extracted, and the extreme value of weight fluctuation for each pixel in the temporal dimension is calculated. Due to the continuous change in camera pose during acquisition, the projection compensation coefficients corresponding to the same wound pixel at different time steps are not the same. After generating the non-homogeneous weight compensation matrix for each frame in the image sequence, for each pixel coordinate position, its compensation coefficient sequence across all time steps is extracted, and the minimum value of this sequence is calculated. This minimum value is the extreme value of weight fluctuation (minimum distortion extreme value) for that pixel in the temporal dimension. The minimum distortion extreme value corresponds to the moment when the projection of that pixel is closest to orthogonal (i.e., the angle between the normal vector and the optical axis is smallest) during the entire acquisition process, representing the optimal geometric projection condition for that pixel in the temporal sequence.

[0080] The minimum distortion pixel feature among the extreme values ​​of the weight fluctuations is identified, and the minimum distortion pixel features at different time steps are spatially aligned and recombined to generate orthorectified fused image data after geometric distortion correction. The minimum distortion pixel feature refers to the pixel color value corresponding to the pixel position in the original image at the moment when the compensation coefficient is smallest in the temporal dimension (i.e., the projection distortion is lightest). Spatial alignment refers to reducing the probability of pixel coordinate offset caused by camera micro-displacement for pixels at the same physical location in images acquired at different time steps through image registration algorithms (such as affine transformation based on feature points or optical flow methods), ensuring that pixel contributions from different time steps are projected onto a unified reference coordinate system. Pixel recombination refers to extracting the minimum distortion pixel feature of each pixel coordinate position and filling it into the corresponding coordinates to construct a new synthetic image, i.e., orthorectified fused image data. Each pixel in this image originates from the moment of minimum projection distortion in the temporal sequence, thus approximating an ideal image of the wound viewed from the orthogonal direction.

[0081] The area fitting correction module performs secondary area correction on the orthorectified fused image data. Secondary area correction refers to re-inputting the orthorectified fused image data into the segmentation network to extract the wound mask, and then using the new mask and the corresponding optimal compensation coefficient to perform area integration again to obtain a more accurate final area estimate after fusion of temporal multi-frame information.

[0082] In a preferred embodiment, a compensation matrix is ​​generated frame by frame from a sequence of 30 continuously acquired wound images, and the image coordinates are adjusted accordingly. The compensation coefficient sequence of the pixel across 30 frames was statistically analyzed, and the minimum value appeared in frame 17 (compensation coefficient 1.05). The RGB color value of this pixel in frame 17 was... The selected pixel feature is the one with the smallest distortion. After performing the same operation on all pixels in the entire image, the corresponding pixels of each frame with the smallest distortion are spatially aligned using optical flow and then filled into the reference coordinate system to generate an orthorectified fused image. After re-segmenting the fused image and performing a second area correction, the area is corrected from 1,790 mm² to 1,762 mm², reflecting the further correction of the single-frame result by multi-frame fusion.

[0083] Through the above technical solution, this application realizes an orthorectified fusion method based on the selection of optimal projection conditions for multiple time-series frames. This method fully utilizes the multi-view redundant information caused by posture changes during continuous handheld acquisition, transforming posture jitter, originally considered a source of noise, into information resources that improve measurement accuracy. Compared to traditional methods that rely on single-frame images for area calculation, the multi-frame fusion method reduces the geometric distortion residual in the entire wound image by selecting the optimal projection time pixel by pixel, achieving an approximate orthorectified imaging effect without additional hardware. Combined with the overall system solution, the orthorectified fused image data not only improves the accuracy of area calculation but also provides a high-quality wound image with minimal distortion. This image can serve as a standard reference image in clinical records for longitudinal comparison of measurement results at different times, possessing significant clinical archival value.

[0084] This application further proposes that it also includes:

[0085] The edge coordinate set of the wound mask pixel data is extracted, and the boundary curvature feature value corresponding to the edge coordinate set is calculated. The edge coordinate set refers to the set of pixel coordinates located at the boundary between the wound area and the background area in the wound mask. Specifically, it can be extracted by performing morphological gradient operations (dilation result minus erosion result) on the wound mask pixel data or by Canny edge detection. The boundary curvature feature value refers to the curvature scalar sequence calculated point-by-point along the wound edge contour. Specifically, it can be calculated by fitting a parametric curve (such as a B-spline curve) to the edge coordinate set and then taking the ratio of the second derivative to the first derivative magnitude of the fitted curve. The boundary curvature feature value reflects the local curvature of the wound contour in the image coordinate system. When the wound edge passes through an area with large surface curvature, it should exhibit a high rate of curvature change in the image.

[0086] The local surface curvature data of the local surface geometric model at corresponding coordinate positions is obtained, and the deviation between the boundary curvature feature value and the local surface curvature data is compared to generate an edge fitting residual. The local surface curvature data refers to the theoretical curvature prediction values ​​corresponding to each edge coordinate position extracted from the constructed local surface geometric model, obtained by calculating the principal curvature of the parametric surface model at the corresponding parameter positions. The edge fitting residual is characterized by calculating the absolute difference between the measured boundary curvature feature value and the theoretical local surface curvature data. This residual reflects the degree of conformity between the current mask edge contour and the surface geometric model under local geometric continuity constraints.

[0087] If the edge fitting residual is detected to be greater than a preset constraint threshold, pixel trajectory interpolation is performed on the corresponding feature fracture region using the local surface curvature data to generate a topology correction mask. The target wound's physical surface area is then updated using this topology correction mask. The feature fracture region refers to the edge coordinate segment where the edge fitting residual exceeds the preset constraint threshold; it typically corresponds to the mask edge fracture location in the image caused by local occlusion, specular reflection, or unevenness in the segmentation network's output boundary. Pixel trajectory interpolation refers to performing cubic spline interpolation along a surface with equal curvature constraint trajectory between the two control points of the feature fracture region, based on the edge direction predicted by the local surface curvature data, to generate a complete edge coordinate sequence that conforms to the surface's geometric continuity constraints. The edges of the feature fracture region in the original wound mask are corrected to the interpolated topology correction edges, forming a topology correction mask. The topology correction mask is then used to replace the original mask, and area integration is re-executed to update the target wound's physical surface area.

[0088] If the edge fitting residual is detected to be less than or equal to a preset constraint threshold, it is determined that the edge coordinate set meets the local geometric continuity constraint. The mask edge does not need to be corrected, and the subsequent calculation process can be continued directly with the original mask.

[0089] As a preferred embodiment, in a forearm wound scenario, the mask edge is located at the image coordinates A fracture gap of approximately 35 pixels exists nearby. The residual between the boundary curvature feature value of this segment and the theoretical curvature predicted by the surface model is 0.172 / mm, exceeding the preset constraint threshold of 0.111 / mm. The system identifies this fracture region and uses cubic spline interpolation based on the local surface curvature constraint of the surface model at the corresponding location to complete the 35-pixel edge gap, generating a topology correction mask. The area is recalculated using the corrected mask, and the final area is corrected from 1,790 mm² to 1,824 mm², a correction of approximately 1.9%, reducing the probability of area underestimation due to the mask edge fracture.

[0090] Through the above technical solution, this application realizes an automatic wound mask edge repair function based on surface geometric continuity constraints. It explicitly and automatically corrects edge geometric inconsistencies originally hidden in the segmentation output through a cross-validation mechanism with the surface model. Compared to traditional methods that rely on the accuracy of the segmentation network's own output, geometric constraint-driven topology correction provides an additional layer of physical rationality verification for the mask edges, reducing the probability of mask breakage caused by imaging interference (such as strong light or skin folds) and improving the robustness of area calculation in complex emergency scenarios. Combined with the overall system solution, the topology-corrected mask updates the physical surface area value of the target wound, forming a dual accuracy guarantee with the compensation matrix-driven area correction mechanism: the former solves pixel-level projection ratio errors, and the latter solves mask edge geometric continuity defects. Together, they further reduce the overall measurement error of the system in complex acquisition scenarios.

[0091] This application further proposes that it also includes:

[0092] The closed-loop curvature features of pixels belonging to the boundary region in the wound mask pixel data are extracted. Closed-loop curvature features refer to the curvature distribution sequence extracted along the complete outer perimeter of the wound mask. Their closed-loop attribute means that the start and end points of the curvature sequence are spatially continuous and closed, forming a complete quantitative description of the overall geometry of the wound boundary. Specifically, this can be achieved by parametrically fitting the set of coordinates of the wound mask edge (e.g., arc length parametrication), calculating the curvature point-by-point on the closed curve, and forming an ordered curvature sequence. Closed-loop curvature features can reflect the differences in the degree of curvature of the wound contour in different directions, including the larger curvature component in areas of higher surface curvature (e.g., the side of the limb) and the smaller curvature component in areas of lower surface curvature (e.g., the center of the front of the limb). This feature sequence comprehensively encodes the spatial projection relationship between the wound contour and the underlying surface geometry, and is a key observation driving the correction of geometric model parameters.

[0093] The closed-loop curvature feature is compared with the local surface curvature of the local surface geometric model to generate curvature fitting error data. Local surface curvature refers to the theoretical curvature prediction value extracted from the current local surface geometric model at each edge coordinate position, obtained by performing second-order differential operations on the surface model parametric equations at the corresponding parameter coordinates. The curvature fitting error data is generated by calculating the difference between the measured closed-loop curvature feature value and the theoretical local surface curvature value point-by-point, forming an error sequence of the same length as the edge coordinate set. This error sequence reveals the fitting deviation of the current local surface geometric model at each edge position: areas with smaller errors indicate that the model accurately describes the surface morphology at that location; areas with larger errors indicate that the actual surface curvature at that location deviates from the prior parameters, possibly due to individual anatomical differences or local deformation caused by specific body positions.

[0094] The local surface geometry model is adjusted based on the curvature fitting error data to update the non-homogeneous weight compensation matrix. The model adjustment process refers to an optimization process that iteratively corrects the control parameters (such as the ellipsoidal axis-to-length ratio and local curvature coefficients) of the local surface geometry model using the curvature fitting error as a feedback signal. Specifically, gradient descent or Newton's method can be used to update the parameters with the goal of minimizing the overall mean square value of the curvature fitting error. The updated local surface geometry model can more accurately describe the actual wound area geometry of the current patient, thereby re-executing the normal vector calculation process of the projection analysis module to generate a new version of the non-homogeneous weight compensation matrix after geometric model correction, thus improving the accuracy of subsequent area integration. This dynamic model update mechanism enables the system to continuously optimize the geometric estimation of the target surface during multi-frame acquisition, achieving a progressive accuracy improvement from prior-driven to data-driven approaches.

[0095] In a preferred embodiment, in the forearm wound acquisition scenario, the curvature sequence of the closed-loop edge of the wound mask was extracted, yielding measured curvature values ​​from 1,248 edge sampling points. Comparison with the theoretical curvature values ​​predicted by the prior ellipsoidal model revealed a mean curvature fitting error of 0.018 / mm at 320 sampling points on the radial side of the forearm (left side of the image) (exceeding the set error threshold of 0.012 / mm), while the mean error in other areas was 0.006 / mm (below the threshold). Using the 320 high-error sampling points as supervisory signals, gradient descent correction was applied to the minor axis radius parameter of the ellipsoid. After three iterations, the minor axis radius was updated from 24mm to 21.5mm, and the mean curvature fitting error decreased to 0.009 / mm. The compensation matrix was regenerated using the corrected geometric model. The compensation coefficient for the radial edge region of the forearm was corrected from 1.36 to 1.43, and the final area was updated from 1,824 mm² to 1,851 mm², achieving a closed-loop improvement in the accuracy of individualized geometric modeling.

[0096] Through the above technical solution, this application realizes an online adaptive update mechanism for the local curved surface geometric model driven by wound contour geometric observation. This mechanism transforms the real edge curvature information acquired in each acquisition into a basis for correcting the prior geometric model, enabling the system to gradually reduce the systematic deviation between statistical prior parameters and the actual anatomical morphology of a specific patient. Compared to methods that rely entirely on fixed prior parameters, adaptive model updates allow the area measurement accuracy to iteratively converge with the increase in the number of acquisition frames, exhibiting stronger adaptability to individual differences and postural deformations. Combined with the overall system solution, the curvature fitting error-driven geometric model update, orthorectification fusion, and edge topology correction together constitute a triple accuracy optimization mechanism at the multi-frame sequence level: orthorectification fusion solves the problem of selecting the optimal temporal frame, edge topology correction addresses mask continuity defects, and geometric model updates address the deviation between prior parameters and the actual individual morphology. The contributions of these three to the overall system measurement accuracy are independent and additive, achieving a high-precision and robust intelligent measurement target for emergency trauma wound area.

[0097] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A deep learning-based intelligent measurement and recording system for emergency trauma wound area, characterized in that, Includes the following modules: The data synchronization acquisition module is used to synchronously acquire two-dimensional color image sequence data, six-axis spatial attitude vector data and measurement timestamp data of the wound area to be measured. The feature modeling module is used to input the two-dimensional color image sequence data into a preset segmentation network to extract the wound mask pixel data of the wound area to be tested; and to construct a local surface geometric model of the corresponding wound area to be tested based on the six-axis spatial pose vector data and preset prior parameters of human body parts. The projection analysis module is used to calculate the angle distribution between the spatial normal vector and the camera optical axis vector of each sampling point on the local curved surface geometric model, and to map the angle distribution to the image coordinate system to generate a non-homogeneous weight compensation matrix corresponding to the wound mask pixel data. The area fitting correction module is used to perform projection deformation correction on the physical area contribution of each pixel in the wound mask pixel data through the non-homogeneous weight compensation matrix, and to perform integral operation on the corrected pixel data to obtain the target wound physical surface area value. The evaluation feedback module is used to calculate the measurement confidence score of the current wound mask pixel data by using the gradient change rate of the calculated non-homogeneous weight compensation matrix; if the measurement confidence score is less than a preset measurement threshold, then acquisition guidance instruction data is generated; if the measurement confidence score is greater than or equal to the preset measurement threshold, then the target wound physical surface area value, the two-dimensional color image sequence data and the measurement timestamp data are associated and encapsulated to generate structured measurement record data.

2. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, The process of extracting the wound mask pixel data of the wound area to be tested includes: A preset segmentation network is invoked to perform pixel-level probability mapping on the two-dimensional color image sequence data to identify a set of target pixel coordinates that conform to the distribution characteristics of wound texture; a binarized feature matrix with the same resolution as the two-dimensional color image sequence data is generated based on the target pixel coordinate set, and the binarized feature matrix is ​​used as the wound mask pixel data.

3. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, The process of constructing a local surface geometric model corresponding to the wound area to be tested includes: A spatial transformation matrix is ​​constructed using the six-axis spatial attitude vector data, and the preset prior parameters of human body parts are mapped to the pose using the spatial transformation matrix to generate a local surface geometric model describing the physical undulation state of the wound area to be tested.

4. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, The process of generating a non-homogeneous weight compensation matrix includes: Calculate the normal vector of each pixel coordinate point on the local curved surface geometric model, and determine the camera optical axis vector based on the six-axis spatial attitude vector data; Calculate the cosine of the angle between each of the normal vectors and the camera optical axis vector, and define the reciprocal of the cosine of the angle as the pixel area compensation coefficient to generate a non-homogeneous weight compensation matrix that matches the dimension of the wound mask pixel data.

5. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 4, characterized in that, The process of calculating the physical surface area of ​​the target wound includes: Obtain the preset pixel unit step size data; Each pixel in the wound mask pixel data of the area to be tested is multiplied by the pixel area compensation coefficient of the corresponding coordinate in the non-homogeneous weight compensation matrix to obtain the corrected sub-pixel area unit. The sum of all the sub-pixel area units is calculated, and the physical surface area of ​​the target wound is obtained by combining the pixel unit step size data.

6. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, The process of calculating the measurement confidence score includes: Perform a difference operation on the non-homogeneous weight compensation matrix to extract a gradient feature map describing the degree of weight mutation; identify connected regions in the gradient feature map whose values ​​exceed a preset rate of change threshold, and record them as high distortion masks; Calculate the percentage of the area of ​​the high-distortion mask relative to the pixel data of the wound mask, and inversely map the percentage of the area to a measurement confidence score.

7. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 6, characterized in that, The process of generating acquisition guidance command data includes: Locate the centroid coordinates of the high-distortion mask in the image coordinate system, and calculate the offset vector of the centroid coordinates relative to the image center point; The offset vector is converted into acquisition guidance command data.

8. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, Also includes: Extract the non-homogeneous weight compensation matrix corresponding to each frame in the two-dimensional color image sequence data, and calculate the extreme value of weight fluctuation of each pixel in the temporal dimension. Identify the minimum distortion pixel features in the extreme values ​​of the weight fluctuations, and spatially align and reconstruct the minimum distortion pixel features at different time steps to generate orthorectified fused image data after geometric distortion removal processing; The area fitting correction module performs secondary area correction on the orthorectified fused image data.

9. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, Also includes: Extract the set of edge coordinates from the pixel data of the wound mask, and calculate the boundary curvature feature value corresponding to the set of edge coordinates; The local surface curvature data of the local surface geometric model at the corresponding coordinate position is obtained, and the boundary curvature feature value is compared with the local surface curvature data to generate edge fitting residuals. If the edge fitting residual is detected to be greater than a preset constraint threshold, the edge position corresponding to the edge fitting residual is determined as a feature fracture region. Pixel trajectory interpolation is then performed on the feature fracture region using the local surface curvature data to generate a topology correction mask. The physical surface area value of the target wound is then updated using the topology correction mask.

10. The intelligent measurement and recording system for emergency trauma wound area based on deep learning according to claim 1, characterized in that, Also includes: Extract the closed-loop curvature features of pixels belonging to the boundary region from the pixel data of the wound mask; The closed-loop curvature feature is compared with the local surface curvature of the local surface geometric model to generate curvature fitting error data; The local surface geometry model is adjusted based on the curvature fitting error data to update the non-homogeneous weight compensation matrix.

Images