An automatic wound measurement method based on AR calibration and AI semantic segmentation

Abstract

The present application relates to the technical field of medical wound evaluation, and specifically refers to a wound automatic measurement method based on AR calibration and AI semantic segmentation. The method completes identity binding and establishes a structured medical history archive by non-contact near-field induction reading of a patient wristband, starts an augmented reality shooting interface to superimpose a virtual scale grid on a viewfinder picture, uses an inertial measurement unit and a camera internal parameter matrix to detect perspective deformation in real time, guides a user to adjust a terminal posture through a prompt bar, and automatically locks a pixel and a physical size mapping ratio when deformation is below a threshold value. After image adaptive histogram equalization and Gaussian denoising, the image is input into a residual network semantic segmentation model to output a wound mask. According to this, the total number of pixels is counted, and the physical area and the lengths of the longest and widest axes are calculated in combination with the mapping ratio. The measurement result is transmitted to the cloud for desensitization storage and version snapshot generation. The present application eliminates the influence of perspective deformation on image proportion from the source of collection, and connects standardized collection and automatic measurement into a coherent process.

Description

Technical Field

[0001] This invention relates to the field of medical wound assessment technology, specifically to an automatic wound measurement method based on AR calibration and AI semantic segmentation. Background Technology

[0002] Wounds are a common core assessment object in clinical diagnosis and nursing care. Taking pressure ulcers as an example, also known as pressure injuries, their area and the dimensions of their longest and widest axes are core quantitative indicators for evaluating wound severity and tracking healing progress. Accurate wound measurement data are of great significance for the development of nursing plans, the evaluation of treatment effects, and the assessment of nursing quality.

[0003] Currently, wound measurement in clinical practice still relies primarily on manual methods. Healthcare workers use disposable paper rulers or transparent mesh films directly applied to the wound surface for visual estimation. While this method is simple, its accuracy is low, and the direct contact between the ruler and the wound poses a risk of cross-infection. Differences in measurement experience among different operators can also lead to significant fluctuations in measurement results for the same wound. In recent years, some studies have attempted to use digital cameras to photograph wounds and combine this with computer image processing technology for semi-automatic calculation of wound area. However, due to uncertainties in shooting distance and angle, the proportions of images of the same wound obtained from different photographs are inconsistent, resulting in poor repeatability of measurement results.

[0004] Chinese patent publication CN118592936A proposes a method to correct image proportions by placing a reference object next to the wound during photography. While this method improves the proportion distortion problem to some extent, it still requires placing a physical reference object near the wound, increasing the operational steps. Furthermore, placing the reference object and taking the photograph present difficulties when the wound is located in a positionally restricted area. In addition, if there is a height difference between the physical reference object and the plane of the wound, perspective effects will still introduce additional proportion errors.

[0005] In wound contour recognition, traditional image processing methods often rely on manual interactive drawing or semi-automatic segmentation based on color thresholds. Manual drawing is time-consuming and highly subjective, with different medical staff often making significant differences in their boundary judgments for the same wound image. Although deep learning semantic segmentation technology has been widely used in medical image analysis, the accuracy of the segmentation model is highly dependent on the quality of the input image. Images that have not undergone standardized calibration show significant differences in performance under different lighting conditions and shooting ratios, directly leading to a decrease in the stability of the segmentation results. The root cause of this problem lies in the disconnect between the image acquisition and image analysis stages. The acquisition end does not apply standardized scaling and lighting correction processing to the input image, making it difficult for the segmentation model to operate under consistent input conditions.

[0006] Regarding data security, medical privacy regulations have set forth clear requirements for the secure management of patient personal information and medical data. However, existing mobile assessment tools typically store wound images and patient identification information locally on the terminal device, lacking end-to-end encryption protection from the collection point to the storage point, thus posing a risk of leakage of sensitive patient information.

[0007] In summary, the existing technology faces the following pressing technical challenges. First, the uncertainty of shooting distance and angle in non-contact imaging leads to image distortion, resulting in a lack of reliable benchmarks for measuring wound physical dimensions, making direct comparison of measurement results from different devices and at different times difficult. Second, the wound image acquisition stage is disconnected from the wound contour recognition and physical dimension calculation stages, failing to form a complete technical chain from standardized acquisition to automated measurement. The lack of standardized preprocessing of input images at the acquisition end leads to unstable input quality of semantic segmentation models, resulting in insufficient measurement efficiency and consistency. Third, the lack of a unified security protection mechanism for medical assessment data in the acquisition, transmission, and storage stages exposes patients' privacy information to the risk of leakage. Summary of the Invention

[0008] The purpose of this invention is to provide an automated wound measurement method based on AR calibration and AI semantic segmentation, addressing technical problems in existing technologies such as image proportion distortion caused by uncertainties in shooting distance and angle during non-contact imaging, insufficient measurement consistency due to reliance on subjective human judgment for wound contour recognition, and the lack of unified security protection mechanisms for medical assessment data in all stages of acquisition, transmission, and storage. This invention eliminates the impact of perspective distortion on image proportions from the source of acquisition by using an augmented reality virtual ruler and a real-time detection and locking mechanism for perspective correction. It automatically extracts the wound area and calculates its physical dimensions through the synergistic cooperation of a deep learning semantic segmentation model and a standardized preprocessing workflow. End-to-end data encryption, desensitization processing, and version snapshot mechanisms ensure the security of medical data throughout its entire lifecycle. Therefore, this invention provides an automated wound assessment solution for clinical wound evaluation that offers stable measurement accuracy, a consistent operational process, and data security compliance.

[0009] To address the above problems, this invention provides an automatic wound measurement method based on AR calibration and AI semantic segmentation: An automated wound measurement method based on AR calibration and AI semantic segmentation includes the following steps: Step S1: Patient wound history record creation. The system provides a structured medical history entry template. Enter basic wound information, previous assessment summary, multidimensional description of wound bed status, risk factors and comorbidity markers, and recent treatment measures notes. For patients with multiple wounds, assign a unique identifier number to each wound site. After the medical history information is entered, the system generates an assessment file marked with the current timestamp. Step S2: Based on the identity binding and permission initialization of the medical wristband, the patient's wristband is read through non-contact near-field sensing, the patient's unique hospitalization number is parsed, the hospital information system is called to verify the legality of the identity, and the data access and editing permission policy is loaded according to the preset role of the currently logged-in user. The shared evaluation file with the patient ID as the main index and the permissions are isolated is initialized on the cloud server. Step S3: Standardized image acquisition based on AR virtual ruler and perspective correction. After selecting the wound area, the system starts the augmented reality assisted shooting interface and overlays an interactive virtual ruler control on the viewfinder. The initial physical size of the virtual ruler is calculated based on the screen parameters of the terminal device. The perspective correction module detects the geometric projection relationship between the virtual ruler and the skin plane where the wound is located in real time. When the perspective distortion of the virtual square is detected to be higher than the preset threshold, the interface prompt bar is displayed in red and guides the user to adjust the terminal posture. When the distortion is lower than the preset threshold, the prompt bar turns green and the system locks the pixel-physical size mapping ratio of the current image and records the mapping ratio as the basis for area calculation. Step S4: Image preprocessing. Adaptive histogram equalization and Gaussian denoising are performed on the calibration image acquired in step S3 to increase the contrast between the wound area and the surrounding normal skin. Step S5: Wound contour recognition based on deep learning semantic segmentation. Input the preprocessed image from step S4 into the pre-trained semantic segmentation model and output the binarized pixel mask of the wound region. Step S6: Calculate the physical size of the wound. Based on the binarized pixel mask obtained in step S5 and the pixel-physical size mapping ratio locked in step S3, calculate the actual physical area, the physical length of the longest axis and the physical length of the widest axis of the wound, and overlay the measurement results on the original image in the form of semi-transparent text. Step S7: Data encryption and uploading. The structured record, including patient identity information, wound location, measurement dimensions, area value, photo time, and auxiliary staging results, is encrypted and then pushed to the cloud server via the network. After receiving the data, the cloud server decrypts it, performs desensitization on the patient identity information, and stores it. It also generates a data snapshot with an incrementing version number for each data update.

[0010] Furthermore, in step S1, the medical history entry template includes a basic wound information module, a previous assessment summary module, a multidimensional description of the wound bed status module, and a risk factor and comorbidity marker module; The basic wound information module includes fields for wound location selection (supporting quick selection from preset anatomical atlases), initial date, and wound origin. The previous assessment summary module includes fields for the highest historical stage and the largest historical area. The multidimensional wound bed status description module includes clinical assessment parameters such as tissue type, exudate characteristics, surrounding skin condition, and pain level. The risk factors and comorbidity marker module supports multiple selections of wound-related medical history tags, including diabetes, hypoalbuminemia, and prolonged bed rest. The recent treatment measures notes module includes free text notes fields for dressing type and special medications. For patients with multiple wounds, the system assigns a unique identifier to each wound site, dynamically linking subsequent image acquisition and measurement data with the corresponding number. This allows each wound to have an independent historical record and trend tracking path. After the medical history information is entered, the system generates an assessment file marked with the current timestamp and dynamically links it with all subsequent image acquisition and area measurement data of the patient, forming a complete and traceable chain of clinical evidence.

[0011] Further, in step S2, the identity binding and permission initialization includes: reading the patient's wristband through non-contact near-field sensing to obtain the patient's unique hospitalization number; sending an identity verification request to the hospital information system to verify the legality of the patient's identity corresponding to the hospitalization number; the system loading data access and editing permission policies according to the preset roles of the currently logged-in user; the preset roles include responsible nurse, head nurse, and attending physician, with different roles corresponding to different data access scopes and operation permissions; and initializing a shared assessment file with the patient ID as the main index on the cloud server. In the shared assessment file, users of different roles can only access data fields within their access scope.

[0012] Furthermore, in step S3, the generation and perspective correction process of the virtual ruler includes the following steps: Step S31: The system dynamically generates a virtual ruler based on the screen resolution and pixel density parameters of the terminal device. The virtual ruler is a rectangular grid control and each square in the grid corresponds to a standard physical size. Step S32: The system obtains the current attitude data of the device from the inertial measurement unit of the terminal device and constructs the perspective projection transformation matrix in combination with the camera intrinsic parameter matrix. It calculates the coordinates of the four vertices of the virtual scale grid under the perspective projection in real time and calculates the deviation values ​​of each interior angle and right angle of the quadrilateral formed by the four vertices. Step S33: Compare the weighted average of the deviation values ​​as a quantitative indicator of the degree of perspective distortion with a preset threshold. When the deviation value is higher than the preset threshold, the prompt bar at the edge of the interface is displayed in red and a directional arrow prompt is generated to guide the user to adjust the terminal posture according to the deviation direction. Step S34: When the system detects that the quantitative index of the degree of perspective distortion is lower than the preset threshold, the prompt bar turns green and the system locks the pixel-physical size mapping ratio of the current image. Specifically, the system obtains the pixel side length of each standard square in the virtual ruler grid in the current frame image and calculates the ratio of the pixel side length to the known physical side length of the standard square. The system calculates the mapping ratio for multiple standard squares in the grid and takes the arithmetic mean. The locked mapping ratio value is associated with the timestamp and device identifier of the current image and stored in the local cache. The mapping ratio value is then directly called when calculating the wound area.

[0013] Further, in step S4, the image preprocessing includes: converting the acquired RGB image to the HSV color space and extracting the luminance component V; performing contrast-limited adaptive histogram equalization on the luminance component V to eliminate illumination non-uniformity; converting it back to the RGB color space to obtain an illumination-corrected image; and using a Gaussian filter to smooth and denoise the illumination-corrected image. The standard deviation of the Gaussian kernel is adaptively adjusted according to the image noise estimate to maintain the enhanced contrast between the wound area and the surrounding normal skin.

[0014] Further, in step S5, the semantic segmentation model's processing includes scaling the preprocessed image to the model's fixed input size, extracting multi-scale features using a deep convolutional neural network with an encoder-decoder architecture, the encoder using a pre-trained residual network as its backbone and extracting feature representations from low-level texture to high-level semantics layer by layer, the decoder restoring spatial resolution through layer-by-layer upsampling and skip connections to the corresponding layers of the encoder, and the model output layer using a sigmoid activation function to output a wound probability map with the same resolution as the input image, where each pixel value represents the probability that the pixel belongs to the wound region, and an adaptive threshold is used to adjust the probability. Figure 2 The adaptive threshold is dynamically determined based on the statistical distribution of the probability map. The binarized mask is subjected to morphological opening operation to remove isolated areas with an area smaller than a preset area threshold, and morphological closing operation is performed to fill holes with an area smaller than a preset area threshold inside the mask, so as to obtain a smooth and connected wound area mask.

[0015] Further, in step S6, the calculation of the physical dimensions of the wound includes the following steps: Step S61: Calculate the physical area of ​​the wound. Traverse all pixels with a value of 1 in the binarized pixel mask to count the total number of pixels N in the wound area. Calculate the actual physical area S of the wound according to the pixel-physical size mapping ratio k locked in step S3 using the following formula: in, This refers to the actual physical area of ​​the wound. This represents the total number of pixels in the wound area. This is the pixel-to-physical size mapping ratio, representing the ratio of the physical size to the pixel size; Step S62: Calculate the longest and widest axes of the wound. Perform principal component analysis on the coordinates of all wound pixels in the mask and obtain the first principal component direction as the direction of the longest axis of the wound. Obtain the maximum span in the direction orthogonal to the longest axis as the widest axis. Measure the pixel length of the mask along the longest axis direction and multiply it by the mapping ratio k to obtain the physical length of the longest axis of the wound. Measure the pixel width of the mask along the widest axis direction and multiply it by the mapping ratio k to obtain the physical length of the widest axis of the wound. Step S63: Measurement result annotation. Mark the outline of the wound area on the original image in a semi-transparent color overlay. Draw the longest axis annotation line and the widest axis annotation line outside the outline line, and display arrow marks at both ends of the annotation lines. Display the wound area value, the longest axis length value and the widest axis length value in a semi-transparent text box above or below the screen, and simultaneously display the currently used pixel-physical size mapping ratio information.

[0016] Further, in step S7, the data encryption and uploading includes: the system encapsulates the structured record, including patient identity information, wound location, measurement dimensions, area value, photo taking time and auxiliary staging results, into a JSON format data object, encrypts the data object using the AES-256 symmetric encryption algorithm, and transmits the encrypted data to the cloud server through a TLS secure transmission channel; After receiving the data, the cloud server decrypts it and performs desensitization on the patient's identity information before storing it in a secure database. The server generates a data snapshot with an incrementing version number for each received data, encrypts and stores the wound image data with annotation information, and performs automatic backup and disaster recovery.

[0017] Furthermore, in step S3, the standardized image acquisition also includes automatically adjusting the image acquisition parameters and enhancing the image brightness to maintain the quality of the acquired image when the ambient light is insufficient or uneven. When the handheld shooting is unstable, the system performs motion estimation and optimal frame selection on multiple consecutive frames and selects the frame with the highest clarity as the final acquired image, so that the measurement accuracy and wound recognition accuracy in complex clinical environments are maintained at a preset level.

[0018] A computer-readable storage medium storing a computer program, said computer program being executed by a processor to operate a method for automatic wound measurement based on AR calibration and AI semantic segmentation.

[0019] The advantages of this invention compared to the prior art are: 1. This invention eliminates the influence of perspective distortion on image proportion from the source of acquisition by using an AR virtual ruler and a real-time detection and locking mechanism for perspective correction. This allows wound images taken by different devices to restore physical dimensions based on a unified mapping ratio. Combined with illumination correction and multi-frame optimization processing, it maintains stable measurement accuracy even under complex conditions such as exudation and scab formation.

[0020] 2. This invention automatically extracts the binary pixel mask of the wound area through a deep learning semantic segmentation model, and directly calculates the physical area of ​​the wound and the lengths of the longest and widest axes by combining the mapping ratio locked in step S3. This connects wound contour recognition and physical size calculation into an automated process, eliminating the subjective error of manually drawing the wound boundary.

[0021] 3. This invention ensures the security of medical assessment data in all stages of collection, transmission and storage through end-to-end data encryption, desensitization processing and version snapshot mechanism, and meets the compliance requirements of medical privacy protection regulations.

[0022] 4. This invention integrates patient medical history filing, identity recognition, AR standardized data collection, AI automatic measurement, and secure data uploading into a coherent operation process, simplifying the operational steps in the clinical assessment process and improving the overall efficiency of a single wound assessment. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0024] In the attached diagram: Figure 1 This is an overall flowchart of an automatic wound measurement method based on AR calibration and AI semantic segmentation in Example 1.

[0025] Figure 2 This is a schematic diagram of the AR shooting interface of the mobile terminal in an automatic wound measurement method based on AR calibration and AI semantic segmentation in Example 1.

[0026] Figure 3 This is a schematic diagram of the measurement result annotation interface in an automatic wound measurement method based on AR calibration and AI semantic segmentation in Example 1.

[0027] Figure 4 This is a time sequence diagram of data synchronization between the terminal and the cloud server in an automatic wound measurement method based on AR calibration and AI semantic segmentation in Example 1. Detailed Implementation

[0028] The following detailed description of the embodiments is intended to exemplify the principles of this application, but should not be used to limit the scope of this application. That is, the automatic wound measurement method based on AR calibration and AI semantic segmentation in this application is not limited to the described embodiments.

[0029] Example 1 The method of the present invention will be described in detail below using a specific clinical application scenario as an example.

[0030] This embodiment uses the procedure of assessing a stage III sacral and coccygeal wound in an inpatient ward of a tertiary hospital as a scenario to demonstrate the complete execution flow of the method of the present invention, such as... Figure 1 As shown.

[0031] The patient was an elderly patient who had been bedridden for a long time and had a pressure ulcer on the sacrum and coccyx. Clinical assessment of the wound area and size changes was required regularly to determine healing progress. The responsible nurse used a mobile nursing terminal PDA provided in the department as the mobile terminal. This mobile nursing terminal PDA was equipped with a 12-megapixel rear camera and a six-axis inertial measurement unit, and the wound assessment application involved in this invention was installed. The mobile nursing terminal PDA had a screen resolution of 2560×1600 pixels and a pixel density of 265 PPI.

[0032] Step S1: Establishing a medical history record for the patient's wound. In this embodiment, after the nurse launches the wound assessment application, the system first guides the patient through the medical history creation process. Since this is the patient's first assessment record in this system, the nurse needs to fill in the wound's medical history information. The system interface presents a structured medical history entry template, containing five information modules: basic wound information module, previous assessment summary module, multidimensional description of the wound bed status module, risk factors and comorbidity marker module, and recent treatment measures notes module.

[0033] In the basic wound information module, the system provides a preset human anatomical atlas. Nurses select the sacrococcygeal region as the wound site from the atlas and enter the initial onset date as two months ago and the wound origin as a hospital-acquired infection. The anatomical atlas covers all major body surface regions, supporting rapid wound location.

[0034] In the previous assessment summary module, the nurse filled in that the highest stage recorded in the previous assessment of this wound was Stage III, and the largest previous area was approximately 12 cm. 2 These historical parameters provide a baseline reference for comparative analysis of treatment effects.

[0035] In the multidimensional status description module of the wound bed, nurses sequentially select: tissue type as predominantly granulation tissue with a small amount of necrotic tissue; exudate characteristics as moderate amount of pale yellow exudate; surrounding skin condition as mild redness and swelling at the wound edges; and a 0-10 numerical rating scale with a pain score of 4. This module comprehensively describes the clinical status of the wound bed from multiple dimensions, providing a structured assessment basis for clinical decision-making.

[0036] In the risk factor and comorbidity labeling module, the system's preset risk factor labels include diabetes, hypoalbuminemia, and prolonged bed rest. The nurse selects the diabetes and prolonged bed rest labels based on the patient's medical history. The system supports multi-selection, allowing nurses to select multiple risk factor labels simultaneously based on the patient's actual condition. These labels can be grouped by risk factor in data analysis to statistically analyze wound healing trends.

[0037] In the recent treatment measures notes module, the system provides a free-text notes field. Nurses can enter in this field that the currently used dressing type is hydrocolloid dressing and the special medication is topical silver ion antibacterial gel. This module records operational details related to wound management, ensuring the assessment file contains complete nursing background information.

[0038] After the medical history is entered, the system assigns a unique identifier, SA-001, to the wound site. Image acquisition and measurement data are dynamically linked to this identifier. The system generates an assessment profile tagged with the current timestamp. This profile is dynamically linked to all subsequent image acquisitions and area measurements for that patient, forming a traceable chain of clinical evidence. The medical history creation step integrates fragmented information required for wound care assessment, ensuring that each image acquisition is associated with complete clinical background information. For patients with multiple wounds, the system assigns a unique identifier to each wound site, ensuring that each wound has an independent historical record and trend tracking path.

[0039] Step S2: Identity binding and permission initialization based on medical wristband After the patient's medical history is established, the system proceeds to the identity binding stage. When the nurse arrives at the patient's bedside, the mobile nursing terminal (PDA) uses a non-contact near-field sensor to read the patient's wristband, extract the patient's hospital number, and send an identity verification request to the hospital information system. After verifying the patient information corresponding to the hospital number, the hospital information system returns a successful verification result, and the application loads the patient's basic information into the interface accordingly.

[0040] Simultaneously, the system loads the corresponding data access and editing permission policies based on the currently logged-in user's role, i.e., the responsible nurse. In this embodiment, three roles are preset: responsible nurse, head nurse, and attending physician. Different roles correspond to different data access scopes and operation permissions. The patient's attending physician has full operation permissions; the head nurse has the permission to view the ward's records and review assessment records; and the responsible nurse has the permission to perform assessments and upload images. The system initializes or opens the shared assessment file indexed by the patient ID on the cloud server. Users with different roles can only access data fields within their respective permission scopes.

[0041] The identity binding and permission initialization steps automate patient authentication, multi-role permission management, and cloud-shared file initialization, ensuring that all subsequent assessment operations are performed within a verified identity framework and clearly defined permission boundaries.

[0042] Step S3: Standardized image acquisition based on AR virtual ruler and perspective correction After completing the record-keeping and identity verification, the nurse selects the wound location for this assessment—the sacral and coccygeal region, wound number SA-001—in the application to enter the AR-assisted imaging interface. For example... Figure 2 As shown, the interface overlays a rectangular grid-like virtual ruler control onto the viewfinder of the mobile nursing terminal PDA. The generation and perspective correction process of this virtual ruler includes the following sub-steps: Step S31: Dynamic generation of the virtual ruler The system reads the tablet's screen pixel density of 265 PPI. Based on the preset standard physical dimensions (each square corresponds to 1 cm × 1 cm), it calculates that each square in the virtual ruler grid occupies a display area of ​​approximately 265 × 265 pixels on the screen. The system then renders a rectangular grid control with the standard physical dimensions onto the viewfinder, with the grid lines superimposed on the live view as semi-transparent lines. The initial physical dimensions of the virtual ruler are calculated based on the terminal device's screen parameters and do not rely on any external physical references.

[0043] Step S32: Constructing the perspective projection transformation matrix The nurse holds a mobile nursing terminal (PDA) and aims it at the sacral and coccygeal wound area. A virtual ruler grid is projected onto the wound in the viewfinder. The system's fluoroscopic correction module acquires the device's current attitude data in real time from the PDA's inertial measurement unit (IMU). This IMU includes a gyroscope and an accelerometer: the gyroscope outputs the device's three-axis angular velocity data, and the accelerometer outputs its three-axis linear acceleration data. The system obtains the device's pitch and roll angles relative to gravity by fusing and calculating the sensor data. A fluoroscopic projection transformation matrix is ​​constructed using the camera's intrinsic parameter matrix. This intrinsic parameter matrix includes factory calibration parameters such as focal length and principal point coordinates. In this embodiment, the camera's focal length is 4.2 mm, and the principal point coordinates are the image center position. The system calculates the coordinates of the four vertices of the virtual ruler grid under fluoroscopic projection in real time and calculates the deviation of each interior angle of the quadrilateral formed by the four vertices from the 90° right angle.

[0044] Step S33: Deformation Detection and Attitude Guidance The system takes the weighted average of four deviation values ​​as a quantitative indicator of the degree of fluoroscopic distortion and compares it with a preset threshold. In this embodiment, the preset threshold is 3°. At the initial stage of imaging, due to the tilt angle between the mobile nursing terminal (PDA) and the skin plane where the wound is located, the deviation value exceeds the 3° threshold. Figure 2 As shown, the prompt bars on the left and right sides of the interface are displayed in red; at the same time, directional arrows are generated on the interface according to the direction of the deviation, indicating to the nurse which direction to rotate and move the nursing terminal PDA to reduce the tilt. The nurse adjusts the posture of the nursing terminal PDA according to the arrow prompts, and the deviation value gradually decreases accordingly.

[0045] Step S34, Mapping ratio locking When the deviation value decreases to within 3°, the indicator bar changes from red to green, indicating that the current shooting angle meets the requirements for standardized acquisition. At this time, the system automatically locks the pixel-to-physical size mapping ratio of the current image. The specific steps of the locking process are as follows: obtain the pixel side length of each standard square in the virtual ruler grid in the current frame image, calculate the ratio of the pixel side length to the corresponding 1 cm physical side length; calculate the mapping ratio for each visible square in the grid and take the arithmetic mean. In this embodiment, the final locked mapping ratio is... This means that each pixel corresponds to a physical length of 0.0118 cm. This mapping ratio comprehensively reflects the impact of factors such as camera focal length, shooting distance, and terminal posture on the imaging ratio. The system associates the locked mapping ratio value with the timestamp and device identifier of the current image and stores it in the local cache. This mapping ratio value is directly called when calculating the wound area.

[0046] The aforementioned AR-assisted shooting process constitutes a complete detection-guided-lock workflow. In the detection phase, the perspective correction module continuously calculates the degree of perspective distortion of the virtual ruler grid; in the guidance phase, the abstract geometric deviations are transformed into intuitive operational guidance through color changes in the interface prompt bar and directional arrows; in the locking phase, the mapping ratio is automatically fixed and shooting is allowed when the distortion meets preset conditions. This mechanism solves the problem of image proportion distortion from the source of acquisition, enabling operators to complete standardized image acquisition without the need for external auxiliary tools.

[0047] Furthermore, the mixed lighting from natural light from the window and the ceiling fluorescent lights in the ward resulted in uneven illumination of the wound area. Upon detecting this uneven ambient lighting, the system automatically adjusted the image acquisition parameters and enhanced image brightness to maintain image quality. When the system detected ambient light intensity below a preset brightness threshold, it automatically increased the camera's exposure time and image sensor gain parameters to compensate for insufficient light, restoring the overall brightness of the captured image to a suitable level for shooting. Simultaneously, considering slight shaking during handheld shooting, the system performed motion estimation based on optical flow calculations on multiple consecutively acquired images and used the Laplacian variance index to evaluate the sharpness of each frame, selecting the frame with the highest variance value as the final acquired image. This illumination compensation and multi-frame selection mechanism ensured that the method maintained stable measurement accuracy and wound identification accuracy under complex wound conditions such as exudation and crusting, as well as in mixed lighting environments.

[0048] Step S4: Image preprocessing After image acquisition is complete, the system automatically enters the preprocessing stage. First, the acquired RGB image is converted to the HSV color space, and the luminance component is extracted. The reason for choosing the HSV color space instead of processing directly on the RGB channels is because of the luminance component. Independent of chroma and saturation information, separate processing of the luminance component does not change the color characteristics of the wound area, which is beneficial for preserving the color difference information between the wound and the surrounding skin.

[0049] For the luminance component The algorithm performs contrast-limited adaptive histogram equalization (CLAHE). CLAHE divides the image into several local regions; in this embodiment, the image is uniformly divided into... The image is processed in three local regions. A histogram is calculated for each region, and a contrast limiting factor is applied before equalization is performed. The contrast limiting factor is set to 2.0 to limit excessive noise amplification. Bilinear interpolation is then used to eliminate discontinuities at region boundaries, thus enhancing local contrast while avoiding excessive noise amplification. After processing, the image is converted back to the RGB color space to obtain the illumination-corrected image.

[0050] Subsequently, a Gaussian filter is used to smooth and denoise the illumination-corrected image. The standard deviation of the Gaussian kernel is adaptively adjusted based on the image noise estimate. In this embodiment, the noise estimation uses the median absolute deviation method of the high-frequency components of the image. Specifically, the Laplacian operator response is calculated for the image luminance component, and the median of its absolute value is taken as the noise level estimate. A larger standard deviation is used when the noise estimate is large to enhance the smoothing effect; a smaller standard deviation is used when the noise estimate is small to preserve the detail information of the wound edge. In this embodiment, the Gaussian kernel size is fixed. The pixels, with a standard deviation ranging from 0.5 to 2.0, are linearly mapped based on the noise estimate. In the preprocessed image, the contrast between the wound area and the surrounding normal skin is significantly increased, providing a stable input image for the semantic segmentation step.

[0051] Step S5: Wound contour recognition based on deep learning semantic segmentation After preprocessing, the system scales the image to a fixed input size required by the semantic segmentation model. In this embodiment, the fixed input size is... The image is scaled to maintain the original aspect ratio, with any insufficient pixels padded with zeros. The scaled image is then input into a pre-trained semantic segmentation model.

[0052] This model employs a deep convolutional neural network with an encoder-decoder architecture. The encoder uses a pre-trained ResNet-50 residual network as its backbone. ResNet-50 contains 49 convolutional layers and 1 fully connected layer, using residual connections to mitigate the vanishing gradient problem in deep networks. The encoder extracts feature representations layer by layer, from low-level texture to high-level semantics. Shallow features capture texture and color gradient information of the wound edges, while deep features capture the overall morphology and contextual semantic information of the wound region. The decoder restores spatial resolution through layer-by-layer upsampling. After each upsampling layer, a skip connection is used to concatenate the feature maps of the corresponding layers in the encoder. Skip connections directly pass the shallow features from the encoder that preserve spatial details to the corresponding layers in the decoder, compensating for the spatial resolution loss during upsampling.

[0053] The model's output layer uses the Sigmoid activation function, outputting a wound probability map with the same resolution as the input image. Each pixel value in the probability map represents the probability that the pixel belongs to the wound region, with values ​​ranging from [value range missing]. The system uses an adaptive threshold to determine the probability. Figure 2 Binarization. The adaptive threshold is dynamically determined based on the statistical distribution of the probability map. Specifically, the mean of the probability map plus 0.5 times the standard deviation is taken as the binarization threshold. Pixels with probability values ​​higher than this threshold are marked as wounds, with a corresponding pixel value of 1; pixels with probability values ​​lower than this threshold are marked as background, with a corresponding pixel value of 0.

[0054] Morphological opening operations are performed on the binarized mask to remove isolated regions smaller than 50 pixels. The opening operation involves erosion followed by dilation to eliminate small protrusions and discrete noise points caused by noise in the mask. Then, morphological closing operations are performed to fill holes smaller than 30 pixels inside the mask. The closing operation involves dilation followed by erosion to fill mis-segmented voids in the wound area caused by tissue color differences. After morphological processing, the wound area mask appears as a smooth and connected region, providing an accurate set of pixels for physical size calculations.

[0055] Step S6: Calculation of physical dimensions of the wound After obtaining the binarized pixel mask, the system calculates the physical dimensions of the wound, including the following sub-steps: Step S61: Calculation of the physical area of ​​the wound First, iterate through all pixels with a value of 1 in the mask and count the total number of pixels in the wound area. In this embodiment, the total number of pixels is calculated as follows: Based on the mapping ratio locked in step S3. Calculate the actual physical area of ​​the wound according to the formula. : ; in, The actual physical area of ​​the wound is expressed in square centimeters. This represents the total number of pixels in the wound area. This represents the pixel-to-physical size mapping ratio, expressed in centimeters per pixel. Substituting the data from this embodiment into the formula: ; This represents the physical area corresponding to each pixel. The actual physical area of ​​the wound is obtained by accumulating the pixel areas.

[0056] Step S62: Calculation of the longest and widest axes of the wound Subsequently, principal component analysis was performed on the coordinates of all wound pixels in the mask. Specifically, a two-dimensional coordinate matrix was constructed using the row and column coordinates of all pixels with a value of 1. The covariance matrix of this coordinate matrix was calculated, and eigenvalue decomposition was performed on the covariance matrix. The direction with the largest eigenvalue is the first principal component direction, representing the direction of the most dispersed distribution of wound pixels. This direction was taken as the longest axis of the wound. The pixel length of the mask was measured along this direction and multiplied by the mapping ratio. The physical length of the longest axis of the wound is obtained, which is approximately 2.36 cm in this embodiment. The maximum span in the direction orthogonal to the longest axis is obtained as the widest axis; the pixel width of the mask is measured along the widest axis and multiplied by the mapping ratio. The physical length of the widest axis of the wound is obtained, which is approximately 1.48 cm in this embodiment.

[0057] Step S63, Measurement Result Labeling like Figure 3 As shown, after the calculation is completed, the system overlays the wound area outline onto the original image in a semi-transparent green overlay. The longest and widest axis annotation lines are drawn outside the outline, with arrows at both ends. The wound area of ​​2.58 cm² is displayed in a semi-transparent text box at the top of the image. 2 The measurements include the longest axis (2.36 cm), the widest axis (1.48 cm), and the currently used mapping ratio. These annotations are provided for nurses to visually verify before submission, confirming that the measurements match the actual wound condition.

[0058] In practical clinical comprehensive assessment scenarios, since this method is based on two-dimensional image analysis of the body surface contour, the data obtained in steps S61 and S62 reflect the two-dimensional physical dimensions of the wound surface projection direction. This data can serve as a standardized quantitative indicator for tracking wound healing trends and comparing treatment efficacy. For complex wounds involving depth dimensions such as subcutaneous infiltration, sinuses, or cavities, clinicians can combine and analyze the two-dimensional surface data obtained by this method with the results of three-dimensional imaging examinations such as CT (computed tomography), MRI (magnetic resonance imaging), or ultrasound to construct a comprehensive wound assessment system covering both the body surface contour and deep structures.

[0059] Step S7: Data Encryption and Upload like Figure 4 As shown, after the measurement results are confirmed, the system will encapsulate all the data from this assessment in a structured format. The encapsulated content includes: patient identification information, wound location (sacral and coccygeal region), number SA-001, and wound area of ​​2.58 cm². 2 The system includes the longest axis length of 2.36 cm, the widest axis length of 1.48 cm, the photo capture time, and auxiliary installment results. The system encapsulates these structured records into a JSON data object, where each field is organized as a key-value pair.

[0060] The system uses the AES-256 symmetric encryption algorithm to encrypt the JSON data object. AES-256 uses a 256-bit key to encrypt the data in blocks. The key is rotated periodically by the system security management module. The encrypted data is transmitted to the cloud server through the TLS secure transmission channel. The TLS protocol encrypts and verifies the integrity of the communication data at the transport layer to prevent the data from being intercepted or tampered with during transmission.

[0061] After receiving the data, the cloud server decrypts it and performs anonymization processing on sensitive information such as the patient's name and ID number before storing it in a secure database. This ensures that the stored data cannot be directly associated with a specific patient. The server generates a data snapshot with an incrementing version number for each assessment. In this embodiment, the patient's version number increments from an initial value to the current value. A new snapshot record is generated for each assessment, providing a time-series data foundation for longitudinal comparative analysis of healing trends.

[0062] Wound image data with annotation information is also stored in a secure database after being encrypted with AES-256. Each time a new assessment record is received, the server automatically performs an incremental backup, synchronously writing the structured data and encrypted images of this assessment to an off-site disaster recovery storage node. Simultaneously, a full backup is performed daily according to a scheduled strategy to maintain the integrity of the disaster recovery node data. When the primary storage node fails, the system automatically switches to the disaster recovery node and restores the data from the latest data snapshot, ensuring business continuity. At this point, all operations for this wound assessment are complete. After the assessment data is automatically synchronized to the cloud, the attending physician and other authorized medical staff can view the measurement results and annotated images of this assessment in real time on their respective terminals.

[0063] The present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it automatically measures the pressure ulcer wound according to the operation flow of steps S1 to S7 described above. The computer-readable storage medium includes flash memory, read-only memory, random access memory, magnetic disk, and optical disk. The processor includes one or more of a general-purpose processor, a digital signal processor, and an application-specific integrated circuit.

[0064] It should be noted that the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0065] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. An automatic wound measurement method based on AR calibration and AI semantic segmentation, characterized in that, Includes the following steps, Step S1: Patient wound history record creation. The system provides a structured medical history entry template. Enter basic wound information, previous assessment summary, multidimensional description of wound bed status, risk factors and comorbidity markers, and recent treatment measures notes. For patients with multiple wounds, assign a unique identifier number to each wound site. After the medical history information is entered, the system generates an assessment file marked with the current timestamp. Step S2: Based on the identity binding and permission initialization of the medical wristband, the patient's wristband is read through non-contact near-field sensing, the patient's unique hospitalization number is parsed, the hospital information system is called to verify the legality of the identity, and the data access and editing permission policy is loaded according to the preset role of the currently logged-in user. The shared evaluation file with the patient ID as the main index and the permissions are isolated is initialized on the cloud server. Step S3: Standardized image acquisition based on AR virtual ruler and perspective correction. After selecting the wound area, the system starts the augmented reality assisted shooting interface and overlays an interactive virtual ruler control on the viewfinder. The initial physical size of the virtual ruler is calculated based on the screen parameters of the terminal device. The perspective correction module detects the geometric projection relationship between the virtual ruler and the skin plane where the wound is located in real time. When the perspective distortion of the virtual square is detected to be higher than the preset threshold, the interface prompt bar is displayed in red and guides the user to adjust the terminal posture. When the distortion is lower than the preset threshold, the prompt bar turns green and the system locks the pixel-physical size mapping ratio of the current image and records the mapping ratio as the basis for area calculation. Step S4: Image preprocessing. Adaptive histogram equalization and Gaussian denoising are performed on the calibration image acquired in step S3 to increase the contrast between the wound area and the surrounding normal skin. Step S5: Wound contour recognition based on deep learning semantic segmentation. Input the preprocessed image from step S4 into the pre-trained semantic segmentation model and output the binarized pixel mask of the wound region. Step S6: Calculate the physical size of the wound. Based on the binarized pixel mask obtained in step S5 and the pixel-physical size mapping ratio locked in step S3, calculate the actual physical area, the physical length of the longest axis and the physical length of the widest axis of the wound, and overlay the measurement results on the original image in the form of semi-transparent text. Step S7: Data encryption and uploading. The structured record, including patient identity information, wound location, measurement dimensions, area value, photo time, and auxiliary staging results, is encrypted and then pushed to the cloud server via the network. After receiving the data, the cloud server decrypts it, performs desensitization on the patient identity information, and stores it. It also generates a data snapshot with an incrementing version number for each data update.

2. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S1, the medical history entry template includes a basic wound information module, a previous assessment summary module, a multidimensional wound bed status description module, and a risk factor and comorbidity marker module; The basic wound information module includes fields for wound location selection (supporting quick selection from preset anatomical atlases), initial date, and wound origin. The previous assessment summary module includes fields for the highest historical stage and the largest historical area. The multidimensional wound bed status description module includes clinical assessment parameters such as tissue type, exudate characteristics, surrounding skin condition, and pain level. The risk factors and comorbidity marker module supports multiple selections of wound-related medical history tags, including diabetes, hypoalbuminemia, and prolonged bed rest. The recent treatment measures notes module includes free text notes fields for dressing type and special medications. For patients with multiple wounds, the system assigns a unique identifier to each wound site, dynamically linking subsequent image acquisition and measurement data with the corresponding number. This allows each wound to have an independent historical record and trend tracking path. After the medical history information is entered, the system generates an assessment file marked with the current timestamp and dynamically links it with all subsequent image acquisition and area measurement data of the patient, forming a complete and traceable chain of clinical evidence.

3. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S2, the identity binding and permission initialization includes: reading the patient's wristband through non-contact near-field sensing to obtain the patient's unique hospitalization number; sending an identity verification request to the hospital information system to verify the legality of the patient's identity corresponding to the hospitalization number; the system loading data access and editing permission policies according to the preset roles of the currently logged-in user, the preset roles including responsible nurse, head nurse, and attending physician, with different roles corresponding to different data access scopes and operation permissions; and initializing a shared assessment file with the patient ID as the main index on the cloud server, in which users of different roles can only access data fields within their access scope in the shared assessment file.

4. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S3, the process of generating the virtual ruler and correcting perspective includes the following steps: Step S31: The system dynamically generates a virtual ruler based on the screen resolution and pixel density parameters of the terminal device. The virtual ruler is a rectangular grid control and each square in the grid corresponds to a standard physical size. Step S32: The system obtains the current attitude data of the device from the inertial measurement unit of the terminal device and constructs the perspective projection transformation matrix in combination with the camera intrinsic parameter matrix. It calculates the coordinates of the four vertices of the virtual scale grid under the perspective projection in real time and calculates the deviation values ​​of each interior angle and right angle of the quadrilateral formed by the four vertices. Step S33: Compare the weighted average of the deviation values ​​as a quantitative indicator of the degree of perspective distortion with a preset threshold. When the deviation value is higher than the preset threshold, the prompt bar at the edge of the interface is displayed in red and a directional arrow prompt is generated to guide the user to adjust the terminal posture according to the deviation direction. Step S34: When the system detects that the quantitative index of the degree of perspective distortion is lower than the preset threshold, the prompt bar turns green and the system locks the pixel-physical size mapping ratio of the current image. Specifically, the system obtains the pixel side length of each standard square in the virtual ruler grid in the current frame image and calculates the ratio of the pixel side length to the known physical side length of the standard square. The system calculates the mapping ratio for multiple standard squares in the grid and takes the arithmetic mean. The locked mapping ratio value is associated with the timestamp and device identifier of the current image and stored in the local cache. The mapping ratio value is then directly called when calculating the wound area.

5. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S4, the image preprocessing includes: converting the acquired RGB image to the HSV color space and extracting the luminance component V; performing contrast-limited adaptive histogram equalization on the luminance component V to eliminate illumination non-uniformity; converting it back to the RGB color space to obtain the illumination-corrected image; and using a Gaussian filter to smooth and denoise the illumination-corrected image. The standard deviation of the Gaussian kernel is adaptively adjusted according to the image noise estimate to maintain the enhanced contrast between the wound area and the surrounding normal skin.

6. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S5, the semantic segmentation model's processing includes scaling the preprocessed image to the model's fixed input size, extracting multi-scale features using a deep convolutional neural network with an encoder-decoder architecture, the encoder using a pre-trained residual network as its backbone and extracting feature representations from low-level texture to high-level semantics layer by layer, the decoder restoring spatial resolution through layer-by-layer upsampling and skip connections with the corresponding layers of the encoder, the model output layer using a sigmoid activation function and outputting a wound probability map with the same resolution as the input image, where each pixel value represents the probability that the pixel belongs to the wound region, binarizing the probability map using an adaptive threshold, the adaptive threshold being dynamically determined based on the statistical distribution of the probability map, performing morphological opening operations on the binarized mask to remove isolated regions with an area smaller than a preset area threshold, and performing morphological closing operations to fill holes with an area smaller than a preset area threshold inside the mask, thereby obtaining a smooth and connected wound region mask.

7. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S6, the calculation of the physical dimensions of the wound includes the following steps: Step S61: Calculate the physical area of ​​the wound. Traverse all pixels with a value of 1 in the binarized pixel mask to count the total number of pixels N in the wound area. Calculate the actual physical area S of the wound according to the pixel-physical size mapping ratio k locked in step S3 using the following formula: in, This refers to the actual physical area of ​​the wound. This represents the total number of pixels in the wound area. This is the pixel-to-physical size mapping ratio, representing the ratio of the physical size to the pixel size; Step S62: Calculate the longest and widest axes of the wound. Perform principal component analysis on the coordinates of all wound pixels in the mask and obtain the first principal component direction as the direction of the longest axis of the wound. Obtain the maximum span in the direction orthogonal to the longest axis as the widest axis. Measure the pixel length of the mask along the longest axis direction and multiply it by the mapping ratio k to obtain the physical length of the longest axis of the wound. Measure the pixel width of the mask along the widest axis direction and multiply it by the mapping ratio k to obtain the physical length of the widest axis of the wound. Step S63: Measurement result annotation. Mark the outline of the wound area on the original image in a semi-transparent color overlay. Draw the longest axis annotation line and the widest axis annotation line outside the outline line, and display arrow marks at both ends of the annotation lines. Display the wound area value, the longest axis length value and the widest axis length value in a semi-transparent text box above or below the screen, and simultaneously display the currently used pixel-physical size mapping ratio information.

8. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S7, the data encryption and uploading includes: the system encapsulates the structured record, which includes patient identity information, wound location, measurement dimensions, area value, photo taking time and auxiliary staging results, into a JSON format data object, encrypts the data object using the AES-256 symmetric encryption algorithm, and transmits the encrypted data to the cloud server through a TLS secure transmission channel; After receiving the data, the cloud server decrypts it and performs desensitization on the patient's identity information before storing it in a secure database. The server generates a data snapshot with an incrementing version number for each received data, encrypts and stores the wound image data with annotation information, and performs automatic backup and disaster recovery.

9. The automatic wound measurement method based on AR calibration and AI semantic segmentation according to claim 1, characterized in that, In step S3, the standardized image acquisition also includes automatically adjusting the image acquisition parameters and enhancing the image brightness to maintain the quality of the acquired image when the ambient light is insufficient or uneven. When the handheld shooting is unstable, the system performs motion estimation and optimal frame selection on multiple consecutive frames and selects the frame with the highest clarity as the final acquired image, so that the measurement accuracy and wound recognition accuracy in complex clinical environments are maintained at a preset level.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, operates according to the steps of any one of claims 1 to 9, an automatic wound measurement method based on AR calibration and AI semantic segmentation.

Images