Diabetic foot early detection method combining infrared and visible light imaging

Abstract

The invention discloses a diabetic foot early-stage detection method combining infrared and visible light imaging, and relates to the technical field of diabetic foot medical imaging diagnos.The diabetic foot early-stage detection method comprises the steps that multi-mode image collection and standardization processing are conducted, infrared images and visible light images are synchronously collected, and standardization processing such as size normalization is conducted; image registration and space alignment are carried out, marking points are set based on foot anatomical features, and feature extraction, mismatching point elimination and transformation matrix calculation are carried out; extracting and screening multi-dimensional features, extracting temperature and structural features, and screening by using a Relief-F algorithm; feature lesion recognition and classification are fused, and lesion probability is output through a double-branch convolutional neural network; carrying out detection result verification and feedback optimization, and comparing a clinical diagnosis optimization model; and generating a detection report and storing data, and generating a report containing the fused image. The early lesion detection precision is improved through multi-modal fusion, individual and environment differences are adjusted and adapted in a personalized mode, and reliable technical support is provided for clinic.

Description

Technical Field

[0001] This invention relates to the field of medical imaging diagnostic technology for diabetic foot, and more particularly to a method for early detection of diabetic foot combining infrared and visible light imaging. Background Technology

[0002] Current detection methods largely rely on visual observation and palpation by medical personnel, qualitatively assessing skin color, temperature, and presence of damage. This approach is highly subjective and lacks precision, making it difficult to identify microcirculatory abnormalities or minute structural lesions invisible to the naked eye. The application of single-modality imaging technology also has significant limitations: infrared thermal imaging can only capture abnormal temperature distribution and cannot correlate with changes in skin structure, easily misinterpreting ambient temperature interference as lesions; while visible light imaging can display skin surface structures, it struggles to reflect subcutaneous microcirculation and has weak detection capabilities for deeper issues such as early inflammation and vascular lesions. Furthermore, the registration technology between the two modalities is immature, lacking anatomical constraints and flexible region adaptation, resulting in inaccurate correlation between temperature and structural information, affecting the accuracy of lesion localization.

[0003] Furthermore, existing technologies have significant shortcomings in feature extraction, threshold setting, and long-term monitoring. Feature extraction is often limited to a single dimension, focusing primarily on temperature or structural indicators, failing to fully integrate multi-dimensional information such as texture, color, and gradient, making it difficult to comprehensively characterize the complex features of early lesions. Threshold determination uses fixed standards, neglecting individual differences such as patient age, disease duration, and baseline temperature, and failing to compensate for the influence of season and ambient temperature, resulting in high false positive and false negative rates. There is a lack of targeted monitoring of high-risk areas, making it easy to miss small lesions in common sites such as below the metatarsal heads and the heel. Additionally, existing methods can only perform single tests, unable to trace lesion progression or predict disease changes, making it difficult to support the development of personalized clinical intervention plans and meet the clinical needs of long-term management of diabetic foot. Summary of the Invention

[0004] This invention proposes an early detection method for diabetic foot combining infrared and visible light imaging to solve the problems mentioned in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for early detection of diabetic foot combining infrared and visible light imaging, comprising the following steps: Multimodal image acquisition and standardization processing: Simultaneous acquisition of infrared thermal imaging and visible light images of the patient's foot; infrared imaging wavelength 8-14μm; visible light imaging band 500-700nm, equipped with a ring LED light source to eliminate shadows; images are normalized in size, calibrated in temperature, corrected in white balance and enhanced in contrast. Image registration and spatial alignment: 12 marker points were set based on foot anatomical features. Feature points were extracted by an improved scale-invariant feature transformation algorithm. Mismatched points were removed by a random sampling consistency algorithm. The affine transformation matrix was calculated to unify the pixel spacing. Multi-dimensional feature extraction and screening: temperature-related features are extracted from infrared images, and structure-related features are extracted from visible light images. The Relief-F algorithm is used to screen core features. By integrating feature-based lesion identification and grading, a dual-branch convolutional neural network with an attention mechanism is constructed to output lesion probabilities and classify them into four levels. The model is then optimized and trained. Dynamic threshold adaptive adjustment: Based on the patient's individual baseline data, the judgment threshold is corrected through a multiple linear regression algorithm to adapt to individual differences; Detection results are validated and feedback is optimized by comparing clinical diagnostic results, calculating relevant performance indicators, and accumulating clinical data to continuously optimize model parameters and feature selection thresholds. Test report generation and data storage: Test reports are generated, and related data is stored using encrypted distributed storage.

[0006] Furthermore, it also includes a step for calculating the weight of temperature anomaly regions, the weight calculation formula being: ,in The overall weight of the temperature anomaly region, for The area of ​​the temperature anomaly at any given time. The spatial diffusion attenuation coefficient, To monitor the time window, As an integral variable, the cumulative impact of temperature anomalies in time and space is quantified through integral calculations, making lesion identification more focused on areas of continuously spreading temperature anomalies.

[0007] Furthermore, it also includes a step for analyzing dynamic changes in skin texture, using... Optical flow method is used to track subtle changes in the stratum corneum texture in visible light images and extract texture motion vectors, deformation rates, and continuity features. Based on the analysis results, the accuracy of structural feature extraction is optimized, and infrared temperature features are combined to determine whether it is an early sign of ulceration.

[0008] Furthermore, in the multi-dimensional feature extraction step, a dynamic calculation mechanism for feature fusion weights is added, and a fusion formula is constructed based on the correlation between the two modal features. ,in To fuse feature vectors, This is the infrared temperature feature vector. This is the visible light structure feature vector. For infrared feature weighting coefficients, For visible light features, weighting coefficients. + =1; The correlation between the two modal characteristics is calculated using the Pearson correlation coefficient. When the correlation coefficient is >0.6, the temperature anomaly characteristic is considered significant, and when the correlation coefficient is <0.4, the structural anomaly characteristic is considered significant.

[0009] Furthermore, in the image registration step, an anatomical constraint registration mechanism is added, setting 6 rigid constraint points based on the anatomical structure of the foot bones; and an elastic registration algorithm is used for flexible regions.

[0010] Furthermore, in the dynamic threshold adaptive adjustment step, a seasonal and ambient temperature compensation mechanism is added. The ambient temperature and seasonal coefficient during detection are collected to compensate and correct the individual's baseline temperature. The corrected baseline temperature is... ,in The original baseline temperature at the center of the dorsum of the foot is used to ensure consistency in the determination of temperature anomalies under different environmental conditions.

[0011] Furthermore, in the detection result verification and feedback optimization steps, an iterative optimization formula for model parameters is introduced, which is: ,in These are the network weight parameters after iteration. The current weight parameters, For learning rate, for The detection error corresponding to the current parameters at any given time. for target error at any time, For the iteration time window, As an integral variable, the cumulative effect of error is quantified through integration, making parameter adjustments more aligned with the optimization needs of long-term detection performance.

[0012] Furthermore, in the multimodal image acquisition step, a motion artifact elimination mechanism is added. A sequential acquisition method is adopted, and the motion region is identified by the continuous frame difference method. A GAN-based image inpainting algorithm is used for the motion region, which uses the texture and color information of three adjacent still frames to fill in the motion blur region and repair pixels.

[0013] Furthermore, it also includes a high-risk area key monitoring mechanism, which identifies high-incidence areas of diabetic foot based on clinical data statistics, and adopts a local high-resolution mode for these areas during image acquisition; increases the feature sampling density of high-risk areas during feature extraction, and assigns a 1.2 times weight to the lesion probability of high-risk areas during lesion identification.

[0014] Furthermore, in the test report generation process, a lesion development trend prediction module is added. A time series prediction model is constructed based on the patient's historical test data, using a 3-layer long short-term memory network. The input layer is the historical core feature vector, and the output layer is the probability of lesion grade change within the next 3 months. The prediction results are combined with clinical intervention measures to generate personalized intervention enhancement suggestions. Patients with high risk of deterioration are advised to have a follow-up examination once a week, and at the same time, microcirculation improvement treatment is added. Patients with stable risk are advised to have a follow-up examination once a month and maintain the existing care plan. Patients with improved risk are advised to have a follow-up examination every 3 months to optimize the care frequency.

[0015] Compared with existing technologies, the beneficial effects of this invention are: This invention constructs a precise, comprehensive, and personalized early detection system for diabetic foot by optimizing multimodal fusion and the entire process, effectively overcoming the core deficiencies of existing technologies. The multimodal image acquisition and standardized processing workflow clarifies the key parameters and calibration standards for infrared and visible light imaging. Combined with targeted light source and stage design, it eliminates interference from ambient light and reflections, ensuring the consistency and reliability of image data and laying a solid foundation for subsequent analysis. The improved image registration technology introduces anatomical constraint points and elastic registration algorithms, taking into account both bone rigidity and soft tissue deformation characteristics, achieving high-precision spatial alignment of the two modalities. This enables precise correlation between temperature abnormalities and structural abnormalities, improving the accuracy of lesion localization.

[0016] A multi-dimensional feature extraction and screening mechanism integrates multiple core features such as temperature, texture, color, and stratum corneum state. Through scientific algorithms, it filters high-contribution features to comprehensively capture the complex manifestations of early lesions, avoiding the biased judgments caused by single features. A dual-branch neural network model based on an attention mechanism dynamically allocates feature weights for two modalities, focusing on dominant abnormality types and significantly improving the identification ability of early microcirculatory disorders, minor lesions, and inflammation, reducing missed diagnoses and misdiagnoses. A personalized threshold adjustment mechanism combines individual patient data with environmental and seasonal compensations to adapt detection standards to different populations and scenarios, further reducing the risk of false positives and false negatives.

[0017] The high-risk area-focused monitoring mechanism targets common sites of diabetic foot, employing high-resolution imaging and high-density feature sampling to enhance the detection of minute lesions and meet clinical needs for focused prevention and control in high-incidence areas. A closed-loop system for verifying and optimizing test results continuously iterates model parameters using clinical data, improving the method's generalization ability and clinical adaptability. The lesion progression prediction module constructs a time-series model based on historical data, enabling the prediction of disease progression and providing prospective guidance for clinical intervention. Encrypted database storage and multi-dimensional traceability support the entire process of lesion development, providing data support for adjusting personalized intervention plans and long-term management.

[0018] Overall, this invention achieves accurate detection, precise localization, comprehensive assessment, and trend prediction of early diabetic foot lesions, significantly improving the detection sensitivity and specificity of early lesions, buying valuable time for clinical intervention, and effectively reducing the risk of serious complications such as ulcers and amputation. Simultaneously, personalized testing and intervention recommendations cater to the clinical needs of different patients and are suitable for long-term monitoring and management scenarios, providing reliable technical support for the clinical diagnosis and health management of diabetic foot, and possessing significant clinical application value and promising prospects for wider application. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of an early detection method for diabetic foot combining infrared and visible light imaging proposed in this invention. Figure 2 A bar chart comparing the detection rates of early lesions using different imaging modalities; Figure 3 Line graph comparing detection sensitivity for different lesion grades; Figure 4 A bar chart comparing the false positive rate under different ambient temperatures; Figure 5 A bar chart comparing the detection rates of minute lesions in high-risk and non-high-risk areas. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 5 A method for early detection of diabetic foot combining infrared and visible light imaging, comprising the following steps: Multimodal image acquisition and standardization processing: Simultaneous acquisition of infrared thermal and visible light images of the patient's foot. Infrared imaging utilizes an infrared camera with a wavelength range of 8-14μm, a temperature measurement range of 25-45℃, a temperature measurement accuracy of ±0.1℃, a resolution of 640×512 pixels, and a fixed acquisition distance of 50cm. During acquisition, the patient's foot is placed on a black, non-reflective stage to obtain temperature distribution data across the entire foot area. Visible light imaging employs a 500-700nm visible light band camera with a focal length of 50mm, a fixed ISO parameter of 100, and a resolution of 1920×1080 pixels, equipped with a 50W power, 5500K color temperature ring camera. An LED light source, positioned 30cm from the foot, eliminates shadow interference and acquires structural data such as foot skin texture, color, stratum corneum state, and minor skin lesions. The acquired images are standardized, with dimensions normalized to 1024×1024 pixels (using bilinear interpolation). Infrared images are calibrated to Celsius using a blackbody radiation source (calibration points set at 30℃, 35℃, and 40℃). Visible light images undergo white balance correction using a standard grayscale card, and contrast is enhanced using the CLAHE algorithm (clipLimit=2.0, grid size 8×8) to remove ambient light interference, ensuring consistent and comparable image data. Image registration and spatial alignment were performed using a set of 12 registration markers based on foot anatomical features (phalangeal contours, arch curves, and calcaneal bony landmarks). An improved scale-invariant feature transform algorithm was used to extract key feature points from both infrared and visible light images. A Gaussian difference pyramid was set to 6 layers, with a feature point detection threshold of 0.03. A random sampling consensus algorithm was used to remove mismatched points (1000 iterations, inlier threshold of 2 pixels). An affine transformation matrix (6 degrees of freedom) was calculated to align the spatial coordinates of the two modalities. After registration, a resampling algorithm was used to unify the pixel spacing (0.1 mm / pixel), keeping the registration error within 1.5 pixels, ensuring accurate correspondence between temperature and structural information at the same anatomical location. Multi-dimensional feature extraction and filtering were also performed, extracting temperature features from the infrared images, including the regional average temperature (calculated using a 10×10 pixel sliding window), temperature standard deviation, and the distribution density of the highest temperature points (number of highest temperature points per square centimeter). Temperature gradient change rate (mean of temperature difference between adjacent pixels, unit: ℃ / mm) is used to identify areas with abnormal temperature increases or decreases (areas with a temperature difference exceeding 2℃ and an area ≥ 5mm² are marked as suspicious areas). Structural features are extracted from visible light images: skin texture entropy is calculated using a 5×5 gray-level co-occurrence matrix (step size 1 pixel, angles 0° / 45° / 90° / 135°), color channel mean (RGB channels are weighted averaged by region pixels), stratum corneum thickness estimate (based on a quadratic fitting model of gray values ​​and known thickness samples, R² ≥ 0.95), and the number of microcracks detected (using Canny edge detection, low threshold 100, high threshold 200, cracks ≥ 3mm in length are included). The Relief-F algorithm is used to rank the importance of the extracted 28-dimensional features (100 iterations, 10 neighboring samples), and the top 30% of 8-9 features are selected to form a core feature set, reducing the computational complexity of subsequent calculations. This study integrates feature-based lesion recognition and grading, constructing a dual-branch convolutional neural network model based on an attention mechanism. Both the infrared and visible light feature branches contain four encoder layers (3×3 kernels, stride 1, padding 1, ReLU activation function) and two pooling layers (2×2, stride 2). A CBAM attention module (channel attention + spatial attention) is inserted in between to dynamically allocate weights for the two modalities. The decoder uses four transposed convolutional layers (3×3 kernels, stride 2, padding 1) to restore feature map resolution, and the output layer uses the Softmax activation function. During model training, the batch size is set to 8, and the number of iterations is [number missing]. The model was run 200 times with an initial learning rate of 0.001 (using the Adam optimizer) and a cross-entropy loss function. The model outputs the probability values ​​of lesions in each area of ​​the foot. Based on the probability values, the detection results are divided into four levels: normal (probability < 5%), mild abnormality (5% ≤ probability < 30%), moderate abnormality (30% ≤ probability < 70%), and severe abnormality (probability ≥ 70%). Mild abnormality corresponds to early inflammation or microcirculation disorders, moderate abnormality corresponds to stratum corneum abnormality (thickness > 2 mm) or early ulcer precursors (epidermal damage < 2 mm), and severe abnormality corresponds to obvious ulcers (damage ≥ 2 mm) or infection. Dynamic threshold adaptive adjustment is implemented. A personalized threshold adjustment model is constructed based on individual patient baseline data (age, duration of diabetes, glycated hemoglobin value, and baseline temperature at the center of the dorsum of the foot). The adjustment coefficient is calculated using a multiple linear regression algorithm, with the weights of each input feature being 0.15 for age, 0.3 for duration of diabetes, 0.25 for glycated hemoglobin value, and 0.3 for baseline temperature. The disease probability threshold is dynamically corrected for different individuals. For patients with a duration of diabetes exceeding 10 years, the threshold for mild abnormality is lowered to 3%. For patients with a baseline temperature higher than 34℃, the temperature difference for abnormality is increased to 2.5℃. For patients with a glycated hemoglobin value >9%, the threshold for moderate abnormality is increased to 35%, so that the test results are adapted to individual differences. The detection results are validated and optimized by comparing the automatic detection results with the clinical physician's diagnosis results (combined with physical examination and ultrasound microcirculation detection). The accuracy, sensitivity, and specificity indicators are calculated (accuracy ≥90%, sensitivity ≥88%, specificity ≥92%). The accuracy is the proportion of correctly identified samples out of the total number of samples, the sensitivity is the proportion of correctly detected samples in actual lesion samples, and the specificity is the proportion of correctly identified samples in normal samples. Based on the validation results, the gradient descent algorithm is used to optimize the neural network model parameters. The learning rate is set to 0.0005. The model is updated every 50 clinical data points, and the contribution threshold of feature selection is adjusted synchronously to continuously improve the recognition accuracy. The system generates and stores test reports, including registered infrared-visible fusion images (using weighted fusion, with infrared image weight 0.4 and visible image weight 0.6), lesion area markers (abnormal areas are marked with red rectangles, with a frame width of 3 pixels, indicating the abnormality type and probability value), a list of feature parameters (specific values ​​such as temperature, texture entropy, and stratum corneum thickness), lesion grading results, and risk warnings. Risk warnings include daily care priorities (such as moisturizing frequency and avoiding pressure points) and follow-up examination cycles (every 2 weeks for mild abnormalities, weekly for moderate abnormalities). Data is stored in a distributed database encrypted with AES-256. Stored fields include patient ID, test time, original image data, registered data, core feature vectors, lesion grading results, clinical diagnosis conclusions, and intervention records. The system supports tracing the development trend of foot lesions by time series, providing data support for long-term monitoring.

[0024] This invention also includes a step for calculating the weight of temperature anomaly regions, taking into account the duration and spatial diffusion trend of the temperature anomaly. The weight calculation formula is as follows: ,in The overall weight of the temperature anomaly region, for Area of ​​the temperature anomaly at any given time (unit: mm², area change measured every minute). This is the spatial diffusion attenuation coefficient, with a value of 0.01. The monitoring time window is set to 30 minutes. As an integral variable, the cumulative impact of temperature abnormalities in time and space is quantified through integral calculation, making lesion identification focus more on areas of continuously spreading temperature abnormalities and improving the detection sensitivity of early microcirculatory disorders.

[0025] This invention also includes a skin texture dynamic change analysis step, employing... Optical flow was used to track subtle changes in the stratum corneum texture in visible light images. The window size was set to 15×15 pixels, and the number of iterations was 5. Texture motion vectors (displacement in the x / y directions), deformation rate (the proportion of texture morphology change per unit time, with a calculation cycle of 10 seconds) and continuity features (number of break points in the texture contour / total contour length) were extracted. Based on the analysis results, the accuracy of structural feature extraction was optimized. For areas with a deformation rate exceeding 5%, the texture sampling density was increased from 1 pixel / point to 0.5 pixels / point. Areas with discontinuous continuity were marked as potential crack areas. Combined with infrared temperature features (the area with a temperature more than 2°C higher than the surrounding area was considered high risk), it was determined whether it was an early sign of ulceration, thus improving the ability to identify subtle structural anomalies.

[0026] In this invention, a dynamic calculation mechanism for feature fusion weights is added to the multi-dimensional feature extraction step. A fusion formula is constructed based on the correlation between the two modal features. ,in To fuse feature vectors, , is the infrared temperature feature vector (normalized to the [0,1] interval). , is the visible light structure feature vector (normalized to the [0,1] interval). For infrared feature weighting coefficients, For visible light features, weighting coefficients. + =1, The value range is 0.3-0.7; the correlation between the two modal characteristics is calculated using the Pearson correlation coefficient, and a correlation coefficient > 0.6 indicates that the temperature anomaly characteristics are significant. A correlation coefficient between 0.6 and 0.7, and a correlation coefficient < 0.4, is considered a significant structural anomaly. By setting the weights to 0.3-0.4, the fusion features are more focused on the dominant abnormality type, thereby improving the targeting of lesion identification.

[0027] In this invention, an anatomical constraint registration mechanism is added to the image registration step. Six rigid constraint points are set based on the anatomical structure of the foot skeleton (first metatarsal, calcaneus, and base of the fifth metatarsal). The registration error of the constraint points shall not exceed 0.5 pixels. For the flexible area (soft tissue of the foot and skin between the toes), an elastic registration algorithm is used to achieve fine local alignment through B-spline interpolation. The spacing between the B-spline control points is set to 10 pixels, the elastic coefficient is set to 0.02, and the registration error is iteratively optimized to within 1 pixel. This ensures that the registration result conforms to the rigidity of the skeleton and adapts to the deformation characteristics of the soft tissue, reducing the registration deviation caused by slight changes in foot posture.

[0028] In this invention, a seasonal and ambient temperature compensation mechanism is added to the dynamic threshold adaptive adjustment step, and the ambient temperature at the time of detection is collected ( (accuracy ±0.5℃) and seasonal coefficient ( (Using 1.2 for winter (December-February), 1.0 for spring (March-May), 0.8 for summer (June-August), and 1.0 for autumn (September-November), the individual basal temperature is compensated and corrected, and the corrected basal temperature is: ,in The original baseline temperature of the center of the dorsum of the foot (measurement time ≥ 5 minutes, avoid measurement immediately after exercise) is used to ensure that the judgment of temperature abnormalities is consistent under different environmental conditions, and to avoid false positives caused by excessively low ambient temperature in winter or false negatives caused by excessively high ambient temperature in summer.

[0029] In this invention, the detection result verification and feedback optimization step introduces an iterative optimization formula for model parameters. This formula dynamically adjusts the network weight parameters based on the detection error. ,in These are the network weight parameters after iteration. This is the current weight parameter (initial values ​​are randomly generated in [-0.01, 0.01]). The learning rate is 0.001. for The detection error corresponding to the current parameters at any given time (the proportion of error-determined samples is calculated using the confusion matrix). for Time target error (set to 3%). The iteration time window is set to 7 days. As an integral variable, the cumulative effect of error is quantified through integration, making parameter adjustments more aligned with the optimization needs of long-term detection performance and accelerating model convergence to the target accuracy.

[0030] In this invention, a motion artifact elimination mechanism is added to the multimodal image acquisition step. An image sequence acquisition method with a frame rate of 30fps is adopted, and 10 frames of images are continuously acquired. The motion region is identified by the continuous frame difference method (the region where the pixel gray value changes by more than 5% between frames is determined to be the motion region). A GAN-based image inpainting algorithm (the generator is a U-Net structure and the discriminator is a convolutional neural network) is used for the motion blur region. The texture and color information of three adjacent still frames are used to fill the motion blur region. The inpainting radius is set to 3-5 pixels. The peak signal-to-noise ratio of the inpainted image is ≥35dB, so that clear image data can still be obtained when the patient moves slightly, reducing the impact of motion interference on subsequent feature extraction.

[0031] This invention also includes a high-risk area key monitoring mechanism. Based on clinical data statistics, high-incidence areas of diabetic foot are identified (coordinate range below the first metatarsal head: x: 300-500 pixels, y: 600-800 pixels; coordinate range of the heel: x: 400-600 pixels, y: 800-1000 pixels; coordinate range of the interdigital region: x: 200-800 pixels, y: 400-600 pixels). During image acquisition, a local high-resolution mode is used for these areas (infrared images are upscaled to 1280×1024 pixels with a pixel pitch of 0.05 mm / pixel; visible light images are upscaled to 3840×2160 pixels with a pixel pitch of 0.02 mm / pixel). During feature extraction, the feature sampling density of high-risk areas is increased (the number of sampling points per square millimeter is twice that of non-high-risk areas). During lesion identification, the lesion probability of high-risk areas is assigned a weight of 1.2 times, thereby improving the detection rate of early small lesions (such as epidermal lesions <1 mm) in high-incidence areas.

[0032] In this invention, a lesion development trend prediction module is added to the test report generation step. A time series prediction model is constructed based on the patient's historical test data (at least 3 test records, with an interval of ≥2 weeks). A 3-layer Long Short-Term Memory (LSTM) network is used, with the input layer being the historical core feature vector, the number of hidden layer units being 128, and the dropout rate being 0.2. The output layer is the probability of change in lesion grade within the next 3 months. The probability of change is divided into deterioration (probability > 30%), stability (10% ≤ probability ≤ 30%), and improvement (probability < 10%). The prediction results are combined with clinical intervention measures (type of medication used, frequency of foot care, and wearing of pressure relief devices) to generate personalized intervention enhancement suggestions. Patients with a high risk of deterioration are advised to have a follow-up examination once a week, and microcirculation improvement treatment is added simultaneously. Patients with a stable risk are advised to have a follow-up examination once a month and maintain the existing care plan. Patients with an improvement risk are advised to have a follow-up examination every 3 months and optimize the care frequency, providing prospective guidance for clinical intervention.

[0033] The following two examples further illustrate the specific implementation of this system: Example 1: Routine screening of elderly diabetic patients in a community hospital (winter environment) This embodiment focuses on early screening of diabetic foot in elderly diabetic patients aged 60 and above in community hospitals. A total of 80 patients were included, including 45 males and 35 females, with a diabetes duration of 5-20 years, glycated hemoglobin levels of 6.5%-10.2%, and winter ambient temperatures of 18-22℃. The focus was on detecting early microcirculatory disorders and minor epidermal lesions to verify the practicality and accuracy of the method in routine screening in primary healthcare institutions.

[0034] During the multimodal image acquisition and standardized processing stage, patients need to enter the screening room 30 minutes in advance to acclimatize to the ambient temperature. After removing their shoes and socks, they should place their feet on a black, non-reflective platform, keeping their feet relaxed and toes apart to avoid pressure. Infrared imaging uses an 8-14μm wavelength infrared camera with a temperature measurement range of 25-45℃, a temperature measurement accuracy of ±0.1℃, a resolution of 640×512 pixels, and a fixed acquisition distance of 50cm, simultaneously recording the temperature distribution data of the entire foot area. Visible light imaging uses a 500-700nm band camera with a focal length of 50mm, ISO parameter 100, and a resolution of 1920×1080 pixels, equipped with a 50W, 5500K color temperature ring LED light source. The light source is 30cm away from the foot to eliminate shadows between the toes and on the sole, acquiring skin texture, color, and stratum corneum condition. Image standardization was performed using bilinear interpolation to normalize both types of images to 1024×1024 pixels. Infrared images were calibrated using a blackbody radiation source at three calibration points: 30℃, 35℃, and 40℃. Visible light images underwent white balance correction using a standard grayscale chart, contrast was enhanced using the CLAHE algorithm, and clipLimit was set to 2.0 with a grid size of 8×8 to remove ambient light interference. To address motion artifacts in winter, 10 frames were continuously acquired at 30fps. Motion regions were identified by a standard of over 5% change in pixel grayscale value between frames. For motion regions caused by slight jitter, a GAN-based image inpainting algorithm was used to fill in the gaps using texture and color information from three adjacent still frames, with a repair radius of 3-5 pixels. The peak signal-to-noise ratio of the repaired image reached over 36dB, ensuring image clarity.

[0035] In the image registration and spatial alignment stage, 12 registration markers were set based on foot anatomical features, including key bony landmarks such as the head of the first metatarsal, the base of the fifth metatarsal, the calcaneal tuberosity, and the distal ends of each phalanx. An improved scale-invariant feature transformation algorithm was used to extract feature points, with a 6-layer Gaussian difference pyramid and a feature point detection threshold of 0.03. A random sampling consensus algorithm was used to remove mismatched points, with 1000 iterations and an inlier threshold of 2 pixels. An affine transformation matrix was calculated to achieve spatial coordinate alignment. An anatomical constraint registration mechanism was added. For 6 rigid constraint points, including the first metatarsal and calcaneus, the registration error was controlled to be ≤0.5 pixels. For flexible areas such as the soft tissue of the foot and the skin between the toes, B-spline interpolation was used for elastic registration, controlling the point spacing to 10 pixels and the elasticity coefficient to 0.02. After iterative optimization, the registration error was controlled to within 1 pixel. After resampling, a uniform pixel spacing of 0.1 mm / pixel was applied to ensure accurate correspondence between areas of abnormal heel temperature and areas of rough skin.

[0036] In the multi-dimensional feature extraction and screening stage, the average temperature of the infrared image was calculated using a 10×10 pixel sliding window. The average temperature of the right heel region of a patient was 34.8℃, while the surrounding normal area was 32.2℃, resulting in a temperature difference of 2.6℃ and an area of ​​8mm², which was marked as a suspicious area. Simultaneously, the temperature standard deviation (0.8℃), the distribution density of the highest temperature point (2 points / cm²), and the temperature gradient change rate (0.3℃ / mm) were extracted. For the visible light image, the texture entropy value of the heel region was calculated using a 5×5 gray-level co-occurrence matrix with a step size of 1 pixel and 4 angles, resulting in a value of 0.75. The RGB channel mean values ​​were 165, 152, and 148, respectively. The stratum corneum thickness was estimated to be 2.3mm using a quadratic fitting model. Canny edge detection with a low threshold of 100 and a high threshold of 200 detected a 4mm long microcrack. The Relief-F algorithm was used to sort the 28-dimensional features, iterating 100 times with 10 neighboring samples, and selecting 8 core features, including temperature difference, texture entropy, stratum corneum thickness, and crack length, to form a feature vector. When dynamically calculating the feature fusion weights, the correlation between infrared and visible light features is calculated using the Pearson correlation coefficient, which is 0.65, indicating that the temperature anomaly feature is significant. Therefore, α is set to 0.65, β to 0.35, and the fusion formula is as follows: ,in and All values ​​are standardized to the [0,1] interval to enhance the weight of temperature anomaly features.

[0037] The weighting of temperature anomaly regions is calculated using the formula. ,in The area of ​​the abnormal region is measured every minute. When the area is 0, the area is 8 mm². =9mm² at 10 minutes, =9.5mm² at 20 minutes, =10mm² at 30 minutes =0.01, =30 minutes, calculated by integral. , =245.6, this region has a higher weight than other temporary temperature fluctuation regions, which improves the sensitivity of early microcirculation disorder detection.

[0038] A dual-branch convolutional neural network model was constructed to integrate the identification and grading of characteristic lesions. Both the infrared and visible light characteristic branches contained 4 encoder layers and 2 pooling layers, with a CBAM attention module inserted. The decoder contained 4 transposed convolutional layers. The model was trained with a batch size of 8, 200 iterations, and an initial learning rate of 0.001, employing the Adam optimizer and cross-entropy loss function. For dynamic threshold adaptive adjustment, considering the patient's individual data (age 65, disease duration 12 years, glycated hemoglobin 8.5%, and basal foot temperature 33.5℃), adjustment coefficients were calculated using multiple linear regression. The disease duration weight was highest at 0.3, lowering the threshold for mild abnormalities to 3%. The model output a 38% probability of lesions in the patient's heel region, classifying it as moderate abnormality, corresponding to keratinization abnormalities and early ulcer precursors, consistent with clinical ultrasound microcirculation detection results.

[0039] In the result verification and feedback optimization phase, the automated detection results are compared with the clinical physician's diagnosis results, and the accuracy, sensitivity, and specificity are calculated. Iterative optimization of model parameters uses formulas. ,in The current weight parameters, =0.001, The detection error is 8% within 7 days. =3%, =7 days, calculated by integral. , = -0.001 × 0.35 = -0.00035, precisely adjusting network weights. The test report includes the infrared-visible light fusion image with an infrared weight of 0.4 and a visible light weight of 0.6, a 3-pixel wide border for abnormal areas marked with red rectangles, a list of feature parameters, a moderate abnormality grading result, and a risk warning suggesting weekly follow-up examinations and daily heel moisturizing and stress-reducing care. Data is stored in an AES-256 encrypted distributed database, containing fields such as patient ID, test time, image data, feature vectors, and diagnostic results, supporting time-series traceability.

[0040] The lesion progression prediction module, based on three patient test records with a two-week interval, uses a 3-layer LSTM network with 128 hidden layer units and a dropout rate of 0.2. It predicts a 25% probability of lesion deterioration, a 60% probability of stabilization, and a 15% probability of improvement over the next three months. Combined with the intervention of daily foot care once a day, it is recommended to maintain the current care plan and have a follow-up examination once a month.

[0041] Table 1 Comparison of early screening effects for diabetic foot in elderly patients at community hospitals Table 1 shows that the present invention has significant advantages in community screening scenarios. Existing methods rely on visual observation and single infrared imaging, with an accuracy of only 75% and a sensitivity of 68% for early lesions. Due to the inability to correlate temperature and structural information, the false positive rate reaches 28%. The present invention achieves precise correspondence between temperature abnormalities and structural lesions through multimodal fusion and accurate registration. Core feature screening and attention mechanism models improve recognition accuracy, raising the accuracy rate to 94%. The sensitivity for early lesions reaches 92%, effectively detecting microcirculatory disorders and microcracks. The false positive rate is reduced to 5% due to personalized threshold adjustment and seasonal temperature compensation, avoiding misjudgments caused by winter environmental temperature interference. The false negative rate of 8% is far lower than existing methods, thanks to key monitoring of high-risk areas and elimination of motion artifacts, ensuring that no micro-lesions are missed, fully meeting the clinical needs of routine community screening.

[0042] Example 2: Precise monitoring of high-risk diabetic foot patients in a tertiary hospital (springtime environment) This embodiment focuses on the follow-up monitoring of high-risk diabetic foot patients in a tertiary hospital. It includes 60 patients, all of whom have a disease course of more than 5 years and have peripheral neuropathy. Among them, 30 patients have a history of foot ulcers. The ambient temperature is 22-25℃ in spring. The focus is on monitoring the ulcer healing and recurrence risk, and verifying the performance of the method in accurate monitoring and trend prediction.

[0043] During the multimodal image acquisition and standardization process, pressure should be avoided during the healing period of the patient's foot ulcer. An adjustable stage was used to support the foot during acquisition, exposing the ulcer and a 5cm surrounding area. Infrared imaging parameters were consistent with the previous example. Visible light imaging used a local high-resolution mode for the ulcer area, upscaling to 3840×2160 pixels with a pixel pitch of 0.02mm / pixel, clearly capturing the epidermal growth state at the ulcer edge. During image standardization, given the stable spring temperatures, the visible light image contrast enhancement parameter was adjusted to clipLimit=1.8, and the grid size was set to 8×8. Motion artifact removal was performed at a frame rate of 30fps. For areas of slight skin movement around the ulcer, a repair radius of 3 pixels was applied to ensure clear details at the ulcer edge.

[0044] In the image registration and spatial alignment stage, ulcer edge feature points are added to the registration markers. The registration error of rigid constraint points is ≤0.5 pixels, and the flexible registration in flexible areas is iteratively optimized to a registration error within 0.8 pixels, ensuring that temperature changes in the ulcer area accurately correspond to the epidermal repair status. A key monitoring mechanism for high-risk areas is activated. The x-coordinate range below the first metatarsal head is 300-500 pixels, and the y-coordinate range is 600-800 pixels. The x-coordinate range of the heel is 400-600 pixels, and the y-coordinate range is 800-1000 pixels. Local high-resolution imaging is used, and the infrared image is upscaled to 1280×1024 pixels with a pixel pitch of 0.05 mm / pixel. The feature sampling density is twice that of non-high-risk areas.

[0045] In the multi-dimensional feature extraction and screening stage, monitoring data from a patient with a history of heel ulcers showed that the average infrared temperature of the heel area was 33.6℃, ​​while the surrounding normal area was 32.1℃, with a temperature difference of 1.5℃ and a temperature gradient change rate of 0.2℃ / mm. In the visible light image, the texture entropy value of the ulcer healing area was 0.62, the mean RGB channel values ​​were 158, 146, and 140, the stratum corneum thickness was 1.8mm, and no new cracks were observed. When calculating the feature fusion weights, the Pearson correlation coefficient was 0.38, indicating significant structural abnormalities. α was set to 0.35, β to 0.65, and the fusion formula was used. Strengthen the structural feature weights of the ulcer healing area.

[0046] During the dynamic threshold adaptive adjustment phase, the ambient temperature was 24℃, and the spring seasonal coefficient was used. =1.0, the patient's original basal foot temperature was 33.8℃, and the corrected basal temperature... =33.8 + 0.05 × (25 - 24) × 1.0 = 33.85℃. Considering the patient's individual data (age 58, disease duration 15 years, glycated hemoglobin 7.8%, glycated hemoglobin value > 7%), the threshold for moderate abnormality is raised to 35% to avoid overdiagnosis.

[0047] In the fusion-feature lesion identification and grading stage, the model output a lesion probability of 28% in the heel area, which was judged as a mild abnormality, corresponding to mild local microcirculatory disturbance, consistent with the clinical physical examination results. In the test result verification and feedback optimization stage, the model parameters were updated every 50 accumulated data points, and the optimization further improved the identification accuracy for patients with a history of ulcers.

[0048] The test report includes fused images, ulcer healing area markers, characteristic parameters, and mild abnormality grading results. Risk warnings recommend follow-up every two weeks and continued use of the pressure relief device. Data storage incorporates multi-timepoint records of the ulcer healing process, supporting the tracking of dynamic changes in temperature and structural characteristics. The lesion progression prediction module, based on the patient's four test records, predicts a 12% probability of lesion deterioration, a 75% probability of stabilization, and a 13% probability of improvement over the next three months. Combined with the patient's use of microcirculation-improving medications and twice-daily foot care interventions, it is recommended to maintain the current treatment plan and have a monthly follow-up.

[0049] Table 2 Comparison of the effectiveness of accurate monitoring of diabetic foot in high-risk patients in tertiary hospitals Table 2 shows that the data demonstrates the outstanding performance of this invention in the precise monitoring of high-risk patients. Existing methods cannot accurately capture subtle changes during the ulcer healing period, with a lesion identification accuracy of 78%, an ulcer recurrence warning sensitivity of only 70%, and a trend prediction consistency of less than 70%. This invention, through high-resolution imaging and high-density feature sampling in high-risk areas, increases the detection rate of subtle lesions to 91%, effectively identifying microcirculatory fluctuations at the ulcer edge; multimodal fusion and personalized threshold adjustment achieve a lesion identification accuracy of 96% and a specificity of 97%, avoiding misjudgment of normal changes during the healing period. The 92% trend prediction consistency stems from an LSTM network based on multi-timepoint data, which can accurately predict the direction of lesion development, providing prospective guidance for clinical intervention. In particular, the ulcer recurrence warning sensitivity of 93% can detect early signs of recurrence, allowing for timely adjustment of intervention plans, reducing the risk of ulcer recurrence, and fully meeting the clinical needs for long-term precise monitoring of high-risk patients.

[0050] Reference Figure 2 This figure clearly demonstrates the core advantages of multimodal fusion. Single infrared imaging has a detection rate of less than 60% for structure-related microcracks and stratum corneum abnormalities, while single visible light imaging has a detection rate of only 45% for deep lesions such as microcirculatory disturbances, both exhibiting significant limitations. This invention, through precise fusion of infrared temperature characteristics and visible light structural characteristics, dynamically allocates weights, taking into account both deep microcirculation and surface structural abnormalities, achieving a detection rate exceeding 90% for various early lesions. In particular, for the crucial early microcirculatory disturbances in diabetic foot, the detection rate has increased from a maximum of 65% for single-modality imaging to 92%, effectively solving the problem of single-modality imaging's "missing some aspects" and providing a comprehensive basis for early intervention.

[0051] Reference Figure 3 This figure highlights the high detection capability of this invention for early, mild lesions. Existing methods have a sensitivity of only 68% for mild abnormalities (microcirculatory disturbances), making it difficult to detect early lesions without typical symptoms, which is a core challenge in the progression of diabetic foot. This invention enhances the ability to identify early lesion features through focused monitoring of high-risk areas, multi-dimensional feature screening, and an attention mechanism model. The sensitivity for mild abnormalities is increased to 92%, and the sensitivity for all lesion grades exceeds 90%. As the lesion grade increases, the sensitivity continues to improve to 99%, ensuring that early lesions are not missed and that moderate to severe lesions are accurately diagnosed, comprehensively covering the detection needs of different stages of diabetic foot development.

[0052] Reference Figure 4This figure illustrates the environmental adaptability of this invention. Existing methods are significantly affected by ambient temperature, with false positive rates exceeding 30% at low or high temperatures. Due to the lack of a temperature compensation mechanism, environmentally-induced temperature fluctuations are easily misdiagnosed as lesions. This invention, through seasonal coefficient and ambient temperature compensation, corrects the individual's baseline temperature threshold, keeping the false positive rate below 7% under different ambient temperatures, with minimal fluctuations. Even under extreme conditions of 15°C (low temperature) or 35°C (high temperature), the false positive rates are only 6% and 7%, respectively, far lower than existing methods. This solves the problem of inconsistent test results in different seasons and environments, making it suitable for diverse application scenarios such as communities and hospitals.

[0053] Reference Figure 5 This figure highlights the effectiveness of focused monitoring in high-risk areas. Diabetic foot lesions are concentrated in high-risk areas such as below the first metatarsal head and the heel. Existing methods are not specifically optimized for these areas, resulting in small differences in detection rates between high-risk and non-high-risk areas, both below 70%, and making it easy to miss lesions smaller than 1 mm. This invention employs high-resolution imaging, high-density feature sampling, and weighted enhancement for high-risk areas, achieving a detection rate of 96% in the heel region and 94% below the first metatarsal head, with an overall average of 94% for high-risk areas, significantly higher than the 85% in non-high-risk areas. This differentiated monitoring strategy accurately matches the pathogenesis of diabetic foot, significantly improves the detection capability of small lesions in high-incidence areas, and reduces the risk of lesion progression in core areas.

[0054] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early detection of diabetic foot combining infrared and visible light imaging, characterized in that, Includes the following steps: Multimodal image acquisition and standardization processing: Simultaneous acquisition of infrared thermal imaging and visible light images of the patient's foot; infrared imaging wavelength 8-14μm; visible light imaging band 500-700nm, equipped with a ring LED light source to eliminate shadows; images are normalized in size, calibrated in temperature, corrected in white balance and enhanced in contrast. Image registration and spatial alignment: 12 marker points were set based on foot anatomical features. Feature points were extracted by an improved scale-invariant feature transformation algorithm. Mismatched points were removed by a random sampling consistency algorithm. The affine transformation matrix was calculated to unify the pixel spacing. Multi-dimensional feature extraction and screening: temperature-related features are extracted from infrared images, and structure-related features are extracted from visible light images. The Relief-F algorithm is used to screen core features. By integrating feature-based lesion identification and grading, a dual-branch convolutional neural network with an attention mechanism is constructed to output lesion probabilities and classify them into four levels. The model is then optimized and trained. Dynamic threshold adaptive adjustment: Based on the patient's individual baseline data, the judgment threshold is corrected through a multiple linear regression algorithm to adapt to individual differences; Detection results are validated and feedback is optimized by comparing clinical diagnostic results, calculating relevant performance indicators, and accumulating clinical data to continuously optimize model parameters and feature selection thresholds. Test report generation and data storage: Test reports are generated, and related data is stored using encrypted distributed storage.

2. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, It also includes a step for calculating the weight of temperature anomaly regions, and the weight calculation formula is as follows: ,in The overall weight of the temperature anomaly region, for The area of ​​the temperature anomaly at any given time. The spatial diffusion attenuation coefficient, To monitor the time window, As an integral variable, the cumulative impact of temperature anomalies in time and space is quantified through integral calculations, making lesion identification more focused on areas of continuously spreading temperature anomalies.

3. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, It also includes a step for analyzing dynamic changes in skin texture, using... Optical flow method is used to track subtle changes in the stratum corneum texture in visible light images and extract texture motion vectors, deformation rate, and continuity features; Based on the analysis results, the accuracy of structural feature extraction was optimized, and infrared temperature characteristics were combined to determine whether it was an early sign of ulceration.

4. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the multi-dimensional feature extraction step, a dynamic calculation mechanism for feature fusion weights is added, and a fusion formula is constructed based on the correlation between the two modal features. ,in To fuse feature vectors, This is the infrared temperature feature vector. This is the visible light structure feature vector. For infrared feature weighting coefficients, For visible light features, weighting coefficients. + =1; The correlation between the two modal characteristics is calculated using the Pearson correlation coefficient. When the correlation coefficient is >0.6, the temperature anomaly characteristic is considered significant, and when the correlation coefficient is <0.4, the structural anomaly characteristic is considered significant.

5. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the image registration step, an anatomical constraint registration mechanism is added, setting 6 rigid constraint points based on the anatomical structure of the foot bones; and an elastic registration algorithm is used for flexible regions.

6. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the dynamic threshold adaptive adjustment step, a seasonal and ambient temperature compensation mechanism is added. The ambient temperature and seasonal coefficient during detection are collected to compensate and correct the individual's baseline temperature. The corrected baseline temperature is... ,in The original baseline temperature at the center of the dorsum of the foot is used to ensure consistency in the determination of temperature anomalies under different environmental conditions.

7. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the verification and feedback optimization steps of the detection results, an iterative optimization formula for model parameters is introduced, which is: ,in These are the network weight parameters after iteration. The current weight parameters, For learning rate, for The detection error corresponding to the current parameters at any given time. for target error at any time, For the iteration time window, As an integral variable, the cumulative effect of error is quantified through integration, making parameter adjustments more aligned with the optimization needs of long-term detection performance.

8. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the multimodal image acquisition step, a motion artifact elimination mechanism is added. A sequential acquisition method is adopted, and the motion region is identified by the continuous frame difference method. A GAN-based image inpainting algorithm is used for the motion region to fill the motion blur region and repair pixels by using the texture and color information of three adjacent still frames.

9. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, It also includes a high-risk area key monitoring mechanism, which identifies high-incidence areas of diabetic foot based on clinical data statistics, and uses a local high-resolution mode for these areas during image acquisition; increases the feature sampling density of high-risk areas during feature extraction, and assigns a 1.2 times weight to the lesion probability of high-risk areas during lesion identification.

10. The method for early detection of diabetic foot combining infrared and visible light imaging according to claim 1, characterized in that, In the test report generation process, a lesion development trend prediction module has been added. A time series prediction model is built based on the patient's historical test data, using a 3-layer long short-term memory network. The input layer is the historical core feature vector, and the output layer is the probability of lesion grade change within the next 3 months. The prediction results are combined with clinical intervention measures to generate personalized intervention enhancement suggestions. Patients with high risk of deterioration are advised to have a follow-up examination once a week, and at the same time increase microcirculation improvement treatment; patients with stable risk are advised to have a follow-up examination once a month and maintain the existing care plan. Patients with improved risk are advised to have a follow-up examination every 3 months to optimize the frequency of care.

Images