Machine vision assisted diabetic wound image noise suppression method and system

Abstract

The application discloses a machine vision assisted diabetic wound image noise suppression method and system, relates to the technical field of medical imaging, and comprises the following steps: extracting a reflection image feature of a reflection area of a diabetic foot wound image; analyzing and determining a reflection interference type; if the reflection interference type is exudate reflection, performing wound area calculation error prediction based on the reflection image feature; and formulating an adaptive image denoising mechanism according to the predicted area error, and performing reflection noise suppression processing on the diabetic foot wound image. The technical problem that the existing diabetic foot wound image noise suppression method cannot effectively solve the interference caused by the reflection of the wound area, resulting in insufficient wound area calculation accuracy, is solved.

Description

Technical Field

[0001] This application relates to the field of medical imaging technology, specifically to a machine vision-assisted method and system for noise suppression of diabetic wound images. Background Technology

[0002] With the increasing incidence of diabetes year by year, diabetic foot is characterized by high disability and mortality rates, posing a serious threat to patients' quality of life and health. In actual clinical settings, due to factors such as imaging conditions, wound characteristics, and patient positioning, images of diabetic wounds often suffer from various types of noise interference, among which reflective noise is one of the most common types. Reflective noise not only obscures the true structural information of the wound, blurring the boundary between the wound area and normal tissue in the image, but also interferes with subsequent image segmentation, feature extraction, and area calculation processes.

[0003] However, in the existing technology, most image noise suppression methods are general denoising techniques that fail to fully consider the special characteristics of reflective noise in diabetic wound images and its correlation with wound tissue features. While suppressing reflective noise, they are prone to causing the loss of wound detail information or excessive smoothing, resulting in insufficient accuracy in wound area calculation and thus unsatisfactory denoising effect. Summary of the Invention

[0004] This application provides a machine vision-assisted method and system for noise suppression of diabetic foot wound images, which solves the technical problem that existing methods for noise suppression of diabetic foot wound images cannot effectively solve the interference caused by reflections in the wound area, resulting in insufficient accuracy in wound area calculation.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows:

[0006] In a first aspect, this application provides a machine vision-assisted method for noise suppression of diabetic wound images, the method comprising:

[0007] The reflective image features of reflective areas in diabetic foot lesion images were extracted using a convolutional neural network.

[0008] The type of reflective interference is determined based on the characteristics of the reflected image.

[0009] If the type of reflective interference is exudate reflection, the wound area calculation error is predicted based on the reflective image features to obtain the predicted area error.

[0010] Based on the predicted area error, an adaptive image denoising mechanism is developed to suppress reflective noise in the image of the diabetic foot lesion.

[0011] Secondly, this application provides a machine vision-assisted noise suppression system for diabetic wound images, comprising:

[0012] The image feature extraction module is used to extract reflective image features of reflective areas in diabetic foot wound images through a convolutional neural network.

[0013] An interference type determination module is used to determine the type of reflective interference based on the characteristics of the reflective image.

[0014] The wound error prediction module is used to predict the wound area calculation error based on the reflective image features if the reflective interference type is exudate reflection, and to obtain the predicted area error.

[0015] The wound image processing module is used to formulate an adaptive image denoising mechanism based on the predicted area error, and to perform reflective noise suppression processing on the diabetic foot wound image.

[0016] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0017] This application provides a machine vision-assisted method and system for noise suppression in diabetic foot wound images. First, a convolutional neural network is used to extract reflective image features from reflective areas in diabetic foot wound images. Second, based on a preset feature identification rule table, the type of reflective interference is determined, and differentiated processing strategies are adopted to ensure targeted suppression of different types of reflective interference. Third, for exudate reflective types, the wound area calculation error is predicted based on reflective image features, and a suitable image denoising mechanism is matched from an area error-denoising strategy mapping table based on the prediction error to achieve precise suppression of reflective noise. Finally, based on the predicted area error, a pre-constructed exudate reflective area error-denoising strategy mapping table is queried to match and determine a suitable image denoising mechanism. The denoising intensity of the image reflective denoising algorithm in this mechanism is positively correlated with the area error, thereby achieving dynamic suppression of exudate reflective noise.

[0018] Through the above technical solution, this application can effectively distinguish and process different types of reflective interference. By dynamically adjusting the denoising strategy through error prediction, the impact of reflective noise on wound area calculation can be minimized, and reflective interference in wound images can be accurately suppressed, thereby improving the quality of diabetic foot wound images and the accuracy of subsequent area calculation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the machine vision-assisted noise suppression method for diabetic wound images provided in this application embodiment;

[0021] Figure 2 This is a schematic diagram of the structure of the machine vision-assisted noise suppression system for diabetic wound images provided in the embodiments of this application.

[0022] The components represented by each number in the attached diagram are explained below:

[0023] Image feature extraction module 11, interference type judgment module 12, wound error prediction module 13, wound image processing module 14. Detailed Implementation

[0024] This application provides a machine vision-assisted method and system for noise suppression of diabetic foot wound images, which addresses the technical problem that existing methods for noise suppression of diabetic foot wound images cannot effectively solve the interference caused by reflections in the wound area, resulting in insufficient accuracy in wound area calculation.

[0025] Example 1, as Figure 1 As shown, this application provides a machine vision-assisted method for noise suppression in diabetic wound images, including:

[0026] S10: Extract reflective image features of reflective areas from diabetic foot lesion images using a convolutional neural network;

[0027] In this embodiment, the input diabetic foot wound image is first preprocessed, such as image size normalization, grayscale conversion, or color space conversion. Then, using the constructed reflective image feature extraction model, features are extracted from the preprocessed diabetic foot wound image, outputting a reflective image feature vector that characterizes the reflective area.

[0028] Furthermore, the reflective image feature vector contains information such as color saturation, hue, texture uniformity, spatial location, brightness gradient features, shape features, and size features.

[0029] Specifically, step S10 in the method includes:

[0030] Configure reflective image feature indicators, wherein the reflective image feature indicators include at least color saturation, hue, texture uniformity, spatial location, brightness gradient features, shape features, and size features;

[0031] Based on historical diabetes diagnosis records, a set of images of reflective areas of wounds in diabetic samples was collected, and the reflective image feature sets of different samples were labeled according to the reflective image feature indicators.

[0032] Using the sample reflective region wound image set and the sample reflective image feature set as training data, a convolutional neural network was trained until convergence to construct a reflective image feature extraction model.

[0033] The reflective image feature extraction model is used to extract reflective image features of reflective areas in diabetic foot lesion images.

[0034] In this embodiment of the application, firstly, a reflective image feature index system is configured, including shape features such as color saturation, hue, texture uniformity, spatial location coordinates, brightness gradient magnitude and direction, aspect ratio of the bounding rectangle of the region, and size features such as the percentage of pixels, to ensure that the physical and visual attributes of the reflective area can be fully depicted.

[0035] Secondly, images of diabetic wounds containing typical reflective phenomena are selected from historical diabetes diagnosis records. The reflective areas are manually labeled as a sample reflective area wound image set. Based on preset reflective image feature indicators, each sample reflective area is quantitatively labeled to form a sample reflective image feature set. For example, color saturation can be quantified by the S channel value of the HSV color space, and texture uniformity can be characterized by the energy value of the gray-level co-occurrence matrix.

[0036] Then, the set of wound images of the reflective areas of the samples is used as input, and the corresponding set of reflective image features is used as output to train the convolutional neural network. During the training process, the mean squared error loss function is used, and the network parameters are adjusted through the backpropagation algorithm until the feature extraction error of the model on the validation set converges to below a preset threshold, thereby constructing the reflective image feature extraction model.

[0037] For example, a feature extraction model for reflective images is constructed and trained based on a convolutional neural network, and the specific steps are as follows:

[0038] First, data preparation: based on historical diabetes diagnosis records, a set of wound images of reflective areas of samples was collected, and the wound images of reflective areas of different samples were labeled to obtain the feature set of sample reflective images.

[0039] Secondly, for model construction, the VGG16 network was chosen as the basic architecture. Its original fully connected layers were removed, retaining only the first five convolutional blocks, which contain 13 convolutional layers and 5 pooling layers. A global average pooling layer was added after the last pooling layer to output a fixed-dimensional feature vector. Each convolutional layer uses a 3×3 kernel with a stride of 1, and the pooling layers use 2×2 max pooling with a stride of 2.

[0040] Next, for model training, the extracted reflective image features are used as output. The sample set of wound images of reflective areas is divided into training and validation sets in an 8:2 ratio, and the image input size is uniformly adjusted to 224×224×3. The initial learning rate is set to 0.001, the Adam optimizer is used, the batch size is 32, and the training epochs are 50. After each training epoch, the feature extraction accuracy of the model is evaluated using the validation set. When the loss on the validation set no longer decreases after 5 consecutive epochs, an early stopping strategy is used to terminate training, and the optimal model parameters are saved.

[0041] Finally, the trained reflective image feature extraction model is used to perform forward propagation on the input diabetic foot wound image. The global average pooling layer outputs a reflective image feature vector with a dimension of 512, which contains multi-dimensional feature information such as color, texture, and shape of the reflective area.

[0042] S20: Determine the type of reflective interference based on the reflected image features;

[0043] Specifically, based on a preset feature identification rule table, the type of interference is determined by judging the type of interference based on the reflective image features. The types of interference include exudate reflection, granulation tissue reflection, and eschar reflection.

[0044] In this embodiment, the interference type is determined based on a preset feature identification rule table, and the type of reflective interference is identified. The preset feature identification rule table includes the feature threshold range and combination conditions corresponding to different reflective interference types.

[0045] For example, the reflection of exudate is usually characterized by a color saturation in the range of [30%, 60%], a hue concentrated in the warm hue range of [20°, 40°], a texture uniformity higher than 0.7, and is quantified by the energy value of the gray-scale co-occurrence matrix. Spatially, it is mostly distributed in the concave area in the middle of the wound. The reflection of granulation tissue generally has a color saturation greater than 55%, a hue biased towards orange-red tones of [10°, 25°], a texture uniformity between 0.4 and 0.6, and is mostly located in the newly formed tissue area at the edge of the wound. The reflection of eschar has low saturation, usually below 25%, high brightness gradient characteristics, a gradient amplitude average greater than 150, and is mostly irregularly shaped, with its size usually accounting for less than 15% of the total wound area.

[0046] Furthermore, by comparing the extracted reflective image features with the thresholds and combination conditions in the preset feature identification rule table, the interference type of the current reflective area can be determined when all feature conditions of a certain reflective interference type are met.

[0047] If the type of reflective interference is granulation tissue reflection, the reflective component of the diabetic foot wound image is separated and reconstructed according to the first image processing strategy. The first image processing strategy is to eliminate highlights by estimating and compressing local illumination components, while protecting and enhancing the inherent reflective components of the object to preserve the color and texture features of the granulation tissue.

[0048] In this embodiment of the application, when the type of reflective interference is determined to be granulation tissue reflective, a first image processing strategy is implemented.

[0049] Specifically, firstly, a multi-scale Retinex algorithm based on Retinex theory is used to decompose the diabetic foot wound image, separating it into an illumination component and a reflection component. The illumination component primarily represents the highlight areas and overall brightness distribution in the image, while the reflection component corresponds to the inherent color and texture features of the wound tissue. For the reflective areas of granulation tissue, by analyzing their brightness channels in the HSV color space, a dynamic threshold is set to identify highlight pixels, and adaptive compression processing is applied to the corresponding areas in the illumination component to reduce highlight intensity.

[0050] Among them, the multi-scale Retinex algorithm is an image enhancement algorithm based on the human visual system's perception mechanism of object color. It enhances the local contrast and detail information of the image by decomposing the image into reflection and illumination components and suppressing uneven highlight areas in the illumination component.

[0051] Simultaneously, edge-preserving enhancement is performed on the reflection component, using a bilateral filtering algorithm to smooth noise while preserving the texture details of the granulation tissue edges. During the reflection component reconstruction process, the processed illumination component and the enhanced reflection component are proportionally fused to reconstruct a wound image that eliminates reflection interference from granulation tissue while preserving tissue details.

[0052] Furthermore, if the type of reflective interference is eschar reflection, then the diabetic foot wound image is subjected to edge-aware brightness suppression and boundary preservation processing according to the second image processing strategy. The second image processing strategy is to reduce the brightness of the reflective area at the edge of the eschar while maintaining a clear physical structural boundary between the eschar and normal tissue through guided filtering.

[0053] In this embodiment of the application, when the type of reflective interference is determined to be eschar reflection, a second image processing strategy is implemented.

[0054] Specifically, the Canny edge detection algorithm is first used to extract edges from images of diabetic foot wounds, identifying the physical boundaries between the eschar region and the surrounding normal tissue, and constructing an edge mask. The Canny edge detection algorithm is a classic algorithm for edge detection based on image gradient information. It extracts edge contours from the image through steps such as Gaussian filtering to smooth the image, calculating gradient magnitude and direction, non-maximum suppression, and double threshold detection and connection.

[0055] Subsequently, a guided filtering algorithm was used to process the image. This algorithm uses the original image as the guide map and the edge mask as the constraint condition. When suppressing brightness in the reflective area of ​​the eschar, the smoothing intensity of the filter kernel is adaptively adjusted according to the edge mask information. For the highlight pixels inside the reflective area of ​​the eschar, the smoothing radius of the guided filter is reduced and the regularization parameter is increased to achieve precise suppression of local brightness. For the boundary area between the eschar and normal tissue, the smoothing radius is increased and the regularization parameter is decreased to ensure that the boundary details are not blurred.

[0056] During the brightness suppression process, the brightness histogram of the reflective area of ​​the eschar is analyzed to determine the dynamic brightness adjustment threshold. The brightness value of the highlight pixel is compressed to a range that matches the average brightness value of the surrounding non-reflective area. At the same time, the boundary preservation factor is used to avoid the appearance of halos or artifacts at the edge of the eschar during the processing. Finally, the wound image with effective suppression of eschar reflective noise and clear tissue boundaries is output.

[0057] S30: If the type of reflective interference is exudate reflection, predict the wound area calculation error based on the reflective image features and obtain the predicted area error.

[0058] In this embodiment, when the reflective interference type is exudate reflection, a wound area calculation error prediction model is constructed. This model takes the reflective image feature vector extracted in step S10 as input and the difference between the actual wound area and the area calculated when exudate reflection interference exists as output, thereby obtaining the predicted area error.

[0059] The method of predicting the wound area calculation error based on the reflected image features and obtaining the predicted area error includes:

[0060] Based on historical diabetes diagnosis records, and using exudate reflection as a constraint, a sample reflection image feature set was collected. The historical wound area calculation error in the historical detection process of different sample reflection image features was statistically analyzed as the sample area error, and a sample area error set was obtained.

[0061] The sample reflective image feature set and sample area error set are used as training data, and K-fold cross-division is performed to obtain K sample training sets, where K is an integer greater than or equal to 8;

[0062] The deep learning models are trained to convergence using the K sample training sets, generating K area error predictors;

[0063] Based on the reflective image features, the area error prediction complexity is obtained through area error prediction complexity analysis. Based on the error prediction complexity, K area error predictors, and reflective image features, the wound area calculation error is predicted, and the predicted area error is output.

[0064] In this embodiment, firstly, wound image cases marked as exudate reflection are selected from historical diabetes diagnosis records as sample data sources. For each sample case, its reflective image feature vector is extracted. This vector contains information such as color saturation, hue, texture uniformity, spatial location, brightness gradient features, shape features, and size features, forming a sample reflective image feature set. Simultaneously, for each sample, the actual wound area data is obtained through clinical measurement and compared with the wound area data calculated using conventional image segmentation algorithms without removing exudate reflection interference. The difference between the two is the historical wound area calculation error for that sample. The error values ​​of all samples are summarized to form a sample area error set.

[0065] Secondly, the sample reflective image feature set and the corresponding sample area error set are used as the overall training data, and K-fold cross-split is performed. The value of K is set to an integer greater than or equal to 8, for example, K=10, which means that all training data are randomly and uniformly divided into 10 non-overlapping subsets. In each split, 9 subsets are selected as the training set, and the remaining 1 subset is selected as the validation set. This process is repeated to ensure that each subset has the opportunity to participate in the model evaluation as a validation set.

[0066] Next, K independent deep learning models are trained using K sample training sets. For example, a multilayer perceptron neural network structure is chosen as the deep learning model, with the number of neurons in the input layer matching the dimension of the reflective image feature vector, such as 512 dimensions. Two to three hidden layers are set in the middle, and the number of neurons in each layer is adjusted experimentally, for example, 256 neurons in the first layer, 128 neurons in the second layer, and 1 neuron in the output layer, which is used to predict the area error value.

[0067] During training, mean squared error is used as the loss function, and the Adam optimizer updates the parameters. The learning rate is initialized to 0.001 and can be dynamically adjusted based on the validation set loss. Each model is trained until the loss converges on its respective validation set, meaning that the validation set loss no longer decreases significantly after multiple consecutive rounds, thus generating K independent area error predictors.

[0068] Finally, for newly input images of diabetic foot wounds with exudate reflection, after extracting their reflection image feature vectors, the complexity analysis of area error prediction is first performed based on indicators such as the degree of dispersion of each feature in the reflection image feature vector, the correlation between features, and the degree of deviation of feature values ​​from the distribution of feature values ​​in historical samples.

[0069] Specifically, the steps to obtain the sample area error include:

[0070] The initial wound area of ​​the original diabetic foot wound image without reflection noise suppression processing, corresponding to the reflection image features of the sample;

[0071] Standard wound area of ​​a standard diabetic foot lesion image corresponding to the reflective image features of the collected sample, after reflective noise suppression processing;

[0072] The ratio of the difference between the initial wound area and the standard wound area to the standard wound area is used as the sample area error.

[0073] In this embodiment, firstly, for the original diabetic foot wound image corresponding to the sample reflective image features, i.e., the image without reflective noise suppression processing, a conventional image segmentation algorithm, such as edge detection-based segmentation, is used to segment the wound region and calculate the initial wound area. Specifically, the original diabetic foot wound image is first preprocessed, including grayscale conversion, converting the color image to a grayscale image to reduce computational load, and then smoothing the grayscale image using Gaussian filtering to remove some high-frequency noise. The Canny edge detection algorithm is used to extract the edge contour of the wound region, and then morphological processing, such as erosion and dilation operations, is performed on the detected edge contour to eliminate burrs and holes in the contour, making the edges more continuous and complete. Finally, by calculating the number of pixels enclosed by the edge contour and combining it with the image resolution parameters, such as the actual area represented by each pixel, the number of pixels is converted into the actual initial wound area value. For example, if the image resolution is 0.01 mm² / pixel and the number of pixels within the contour is 10,000, then the initial wound area is 10,000 × 0.01 = 100 mm².

[0074] Secondly, a standard diabetic foot wound image corresponding to the reflective image features of the sample, after professional reflection noise suppression processing, is obtained. This standard image is usually obtained by medical experts through manual annotation or processing by high-precision professional equipment, reflecting the real wound area. The standard wound area is calculated using the same method as the initial wound area calculation mentioned above.

[0075] Finally, the difference between the initial wound area and the standard wound area, i.e., the area difference divided by the standard wound area, is the sample area error of the sample. This quantifies the degree of interference of exudate reflection on the wound area calculation.

[0076] For example, if the initial wound area of ​​a sample is 12cm² and the standard wound area is 10cm², then the area difference is 2cm², and the sample area error is 2 / 10 = 0.2, or 20%.

[0077] Further, based on the reflective image features, an area error prediction complexity analysis is performed to obtain the error prediction complexity. Based on this error prediction complexity, K area error predictors, and the reflective image features, the wound area calculation error is predicted, and the predicted area error is output, including:

[0078] A first feature discrimination matrix is ​​constructed based on the reflected image features;

[0079] Construct a set of sample feature discrimination matrices based on the sample reflective image feature set;

[0080] The first feature identification matrix is ​​compared with multiple sample feature identification matrices in the set of sample feature identification matrices to determine the similarity of multiple samples, and the average value is used to calculate the comprehensive sample similarity.

[0081] The reciprocal of the comprehensive sample similarity is used as the error prediction complexity;

[0082] Multiply the error prediction complexity by P and round down to get J, where P is the initial number of area error predictors selected, P is 3. If the calculated result of J is less than 1, then J is equal to 1. If the calculated result of J is greater than K, then J is equal to K.

[0083] J area error predictors are randomly selected from the K area error predictors. The wound area calculation error is predicted based on the reflective image features. The mean of the J prediction results is then calculated to obtain the predicted area error.

[0084] In this embodiment, the reflective image feature vector is first converted into a matrix form to construct a first feature discrimination matrix. Specifically, the 512-dimensional reflective image feature vector is reshaped into an 8×64 matrix structure, where rows represent different categories of features, such as color features, texture features, and shape features, and columns represent specific feature parameters under each category, forming a first feature discrimination matrix with a row and column structure.

[0085] Secondly, for each sample in the sample reflective image feature set, its feature vector is converted into a corresponding sample feature discrimination matrix in the same way. The sample feature discrimination matrices of all samples together constitute the sample feature discrimination matrix set. For example, if the sample reflective image feature set contains N samples, then the sample feature discrimination matrix set contains N 8×64 sample feature discrimination matrices.

[0086] Then, the similarity between the first feature identification matrix and each sample feature identification matrix in the sample feature identification matrix set is calculated. The similarity calculation uses the cosine similarity algorithm, which measures the similarity by calculating the cosine value of corresponding elements between two matrices. The closer the cosine value is to 1, the more similar the feature distributions of the two matrices are. The similarity is calculated iteratively across all sample feature identification matrices, resulting in multiple sample similarity values. These similarity values ​​are then arithmetically averaged to obtain the overall sample similarity. For example, if the average similarity with 100 sample feature identification matrices is 0.75, then the overall sample similarity is 0.75.

[0087] Furthermore, the reciprocal of the overall sample similarity is used as the error prediction complexity. For example, when the overall sample similarity is 0.75, the error prediction complexity is 1 / 0.75≈1.333. This complexity value reflects the difficulty of matching the current reflective image features with historical sample features. The smaller the similarity, the higher the complexity, indicating that the difference between the current features and historical samples is greater, and the prediction difficulty increases accordingly.

[0088] Next, the error prediction complexity is multiplied by the initial selection number P and rounded down, P=3, to obtain the J value. If the calculated result is less than 1, J is set to 1; if it is greater than K, where K is the total number of area error predictors (e.g., 10), then J is set to K.

[0089] For example, when the error prediction complexity is 1.333, 1.333 × 3 ≈ 4, and J = 4 after rounding; if the error prediction complexity is 0.2, then 0.2 × 3 = 0.6, and J = 1 after rounding; if the error prediction complexity is 4.5, then 4.5 × 3 = 13.5, and K = 10, then J = 10.

[0090] Finally, J predictors are randomly selected from the K area error predictors. The feature vector of the current reflected image is input into each of the J predictors to obtain J area error prediction results. The arithmetic mean of the J prediction results is the final output predicted area error. For example, if four predictors are selected and the prediction results are 15%, 18%, 16%, and 17% respectively, then the predicted area error is (15% + 18% + 16% + 17%) / 4 = 16.5%.

[0091] By dynamically selecting the number of predictors based on similarity, the model's adaptability to data with different feature distributions can be improved while ensuring prediction accuracy.

[0092] S40: Based on the predicted area error, an adaptive image denoising mechanism is formulated to suppress reflective noise in the image of the diabetic foot lesion.

[0093] In this embodiment of the application, when the type of reflective interference is exudate reflection, the formulation of the adaptive image denoising mechanism is based on the predicted area error obtained in step S30. By dynamically adjusting the intensity of the denoising strategy and the algorithm parameters, the noise of exudate reflection is accurately suppressed.

[0094] Specifically, an adapted image denoising mechanism is formulated based on the predicted area error, including:

[0095] A pre-constructed mapping table of area error and denoising strategy for exudate reflection is provided, wherein the denoising strategy includes an image reflection denoising algorithm, and the denoising intensity of the image reflection denoising algorithm is positively correlated with the area error.

[0096] Using the area error-denoising strategy mapping table, the appropriate image denoising mechanism is determined based on the predicted area error matching.

[0097] In this embodiment, firstly, a mapping table of area error and denoising strategy for exudate reflection is pre-constructed through clinical trial data and algorithm simulation tests. This mapping table uses the numerical range of the predicted area error as the index key, with each index key corresponding to a specific denoising strategy. The core component of the denoising strategy is the image reflection denoising algorithm, and it is explicitly stated that the denoising intensity of this algorithm is positively correlated with the area error.

[0098] For example, when the predicted area error is at a low error level, such as ≤10%, the denoising algorithm recorded in the mapping table is based on Gaussian filtering, and its denoising strength parameters, such as the filter kernel size and standard deviation σ, are set to a low level; when the error enters the medium error level, such as 10%-20%, the mapping table points to a composite filtering algorithm that includes adaptive threshold segmentation, and the denoising strength parameters are increased accordingly; while for high error levels, such as >20%, the mapping table will match a deep denoising algorithm that combines multi-scale analysis with morphological operations, and its denoising strength parameters reach the highest level.

[0099] In summary, compared with existing technologies, this application designs differentiated noise suppression strategies for different types of reflective interference, achieving precise and adaptive suppression of reflective noise in diabetic wound images. This effectively solves the problem of wound area calculation errors caused by exudate reflection, thus improving the quality of diabetic wound images.

[0100] In summary, the embodiments of this application have at least the following technical effects:

[0101] This application provides a machine vision-assisted method for noise suppression in diabetic foot wound images. First, a convolutional neural network is used to extract reflective image features from reflective areas in diabetic foot wound images. Second, based on a pre-defined feature identification rule table, the type of reflective interference is determined, and differentiated processing strategies are adopted to ensure targeted suppression of different types of reflective interference. Third, for exudate reflective types, the wound area calculation error is predicted based on reflective image features, and a suitable image denoising mechanism is matched from an area error-denoising strategy mapping table based on the prediction error to achieve precise suppression of reflective noise. Finally, based on the predicted area error, a pre-constructed exudate reflective area error-denoising strategy mapping table is queried to match and determine a suitable image denoising mechanism. The denoising intensity of the image reflective denoising algorithm in this mechanism is positively correlated with the area error, thereby achieving dynamic suppression of exudate reflective noise.

[0102] Through the above technical solution, this application can effectively distinguish and process different types of reflective interference. By dynamically adjusting the denoising strategy through error prediction, the impact of reflective noise on wound area calculation can be minimized, and reflective interference in wound images can be accurately suppressed, thereby improving the quality of diabetic foot wound images and the accuracy of subsequent area calculation.

[0103] Example 2, as Figure 2 As shown, based on the same inventive concept as the machine vision-assisted noise suppression method for diabetic wound images provided in Embodiment 1, this application also provides a machine vision-assisted noise suppression system for diabetic wound images, including:

[0104] Image feature extraction module 11 is used to extract reflective image features of reflective areas in diabetic foot wound images through a convolutional neural network;

[0105] Interference type determination module 12 is used to determine the type of reflective interference based on the reflective image feature analysis;

[0106] Wound error prediction module 13 is used to predict the wound area calculation error based on the reflective image features if the reflective interference type is exudate reflection, and obtain the predicted area error.

[0107] The wound image processing module 14 is used to formulate an adaptive image denoising mechanism based on the predicted area error and to perform reflective noise suppression processing on the diabetic foot wound image.

[0108] In one embodiment, the image feature extraction module 11 is specifically used for:

[0109] Configure reflective image feature indicators, wherein the reflective image feature indicators include at least color saturation, hue, texture uniformity, spatial location, brightness gradient features, shape features, and size features;

[0110] Based on historical diabetes diagnosis records, a set of images of reflective areas of wounds in diabetic samples was collected, and the reflective image feature sets of different samples were labeled according to the reflective image feature indicators.

[0111] Using the sample reflective region wound image set and the sample reflective image feature set as training data, a convolutional neural network was trained until convergence to construct a reflective image feature extraction model.

[0112] The reflective image feature extraction model is used to extract reflective image features of reflective areas in diabetic foot lesion images.

[0113] Furthermore, in one embodiment of the application, the type of reflective interference is determined by judging the type of interference based on the reflective image features according to a preset feature identification rule table, wherein the type of reflective interference includes exudate reflection, granulation tissue reflection and eschar reflection.

[0114] Furthermore, in one embodiment of the application, if the type of reflective interference is granulation tissue reflection, the reflective component of the diabetic foot wound image is separated and reconstructed according to a first image processing strategy. The first image processing strategy is to eliminate highlights by estimating and compressing local illumination components, while protecting and enhancing the inherent reflective components of the object to preserve the color and texture features of the granulation tissue.

[0115] Furthermore, if the type of reflective interference is eschar reflection, then the diabetic foot wound image is subjected to edge-aware brightness suppression and boundary preservation processing according to the second image processing strategy. The second image processing strategy is to reduce the brightness of the reflective area at the edge of the eschar while maintaining a clear physical structural boundary between the eschar and normal tissue through guided filtering.

[0116] Furthermore, in one embodiment of the application, the method of predicting the wound area calculation error based on the reflective image features and obtaining the predicted area error includes:

[0117] Based on historical diabetes diagnosis records, and with exudate reflection as a constraint, a sample reflection image feature set was collected. The historical wound area calculation error in the historical detection process was not statistically analyzed for different sample reflection image features, and a sample area error set was obtained.

[0118] The sample reflective image feature set and sample area error set are used as training data, and K-fold cross-division is performed to obtain K sample training sets, where K is an integer greater than or equal to 8;

[0119] The deep learning models are trained to convergence using the K sample training sets, generating K area error predictors;

[0120] Based on the reflective image features, the area error prediction complexity is obtained through area error prediction complexity analysis. Based on the error prediction complexity, K area error predictors, and reflective image features, the wound area calculation error is predicted, and the predicted area error is output.

[0121] The steps for obtaining the sample area error include:

[0122] The initial wound area of ​​the original diabetic foot wound image without reflection noise suppression processing, corresponding to the reflection image features of the sample;

[0123] Standard wound area of ​​a standard diabetic foot lesion image corresponding to the reflective image features of the collected sample, after reflective noise suppression processing;

[0124] The ratio of the difference between the initial wound area and the standard wound area to the standard wound area is used as the sample area error.

[0125] Further, based on the reflective image features, an area error prediction complexity analysis is performed to obtain the error prediction complexity. Based on this error prediction complexity, K area error predictors, and the reflective image features, the wound area calculation error is predicted, and the predicted area error is output, including:

[0126] A first feature discrimination matrix is ​​constructed based on the reflected image features;

[0127] Construct a set of sample feature discrimination matrices based on the sample reflective image feature set;

[0128] The first feature identification matrix is ​​compared with multiple sample feature identification matrices in the set of sample feature identification matrices to determine the similarity of multiple samples, and the average value is used to calculate the comprehensive sample similarity.

[0129] The reciprocal of the comprehensive sample similarity is used as the error prediction complexity;

[0130] Multiply the error prediction complexity by P and round down to get J, where P is the initial number of area error predictors selected, P is 3. If the calculated result of J is less than 1, then J is equal to 1. If the calculated result of J is greater than K, then J is equal to K.

[0131] J area error predictors are randomly selected from the K area error predictors. The wound area calculation error is predicted based on the reflective image features. The mean of the J prediction results is then calculated to obtain the predicted area error.

[0132] Furthermore, an adaptive image denoising mechanism is formulated based on the predicted area error, including:

[0133] A pre-constructed mapping table of area error and denoising strategy for exudate reflection is provided, wherein the denoising strategy includes an image reflection denoising algorithm, and the denoising intensity of the image reflection denoising algorithm is positively correlated with the area error.

[0134] Using the area error-denoising strategy mapping table, the appropriate image denoising mechanism is determined based on the predicted area error matching.

Claims

1. A machine vision-assisted method for noise suppression in diabetic wound images, characterized in that, The methods include: The reflective image features of reflective areas in diabetic foot lesion images were extracted using a convolutional neural network. The type of reflective interference is determined based on the analysis of the reflected image features. If the type of reflective interference is exudate reflection, the wound area calculation error is predicted based on the reflective image features to obtain the predicted area error. Based on the predicted area error, an adaptive image denoising mechanism is formulated to suppress reflective noise in the image of the diabetic foot lesion. The method of predicting the wound area calculation error based on the reflected image features and obtaining the predicted area error includes: Based on historical diabetes diagnosis records, and using exudate reflection as a constraint, a sample reflection image feature set was collected. The historical wound area calculation error in the historical detection process of different sample reflection image features was statistically analyzed as the sample area error, and a sample area error set was obtained. The sample reflective image feature set and sample area error set are used as training data, and K-fold cross-division is performed to obtain K sample training sets, where K is an integer greater than or equal to 8; The deep learning models are trained to convergence using the K sample training sets, generating K area error predictors. Based on the reflective image features, the area error prediction complexity is obtained through area error prediction complexity analysis. Based on the error prediction complexity, K area error predictors, and reflective image features, the wound area calculation error is predicted, and the predicted area error is output.

2. The machine vision-assisted noise suppression method for diabetic wound images according to claim 1, characterized in that, The reflective image features of reflective areas in diabetic foot wound images were extracted using a convolutional neural network, including: Configure reflective image feature indicators, wherein the reflective image feature indicators include at least color saturation, hue, texture uniformity, spatial location, brightness gradient features, shape features, and size features; Based on historical diabetes diagnosis records, a set of images of reflective areas of wounds in diabetic samples was collected, and the reflective image feature sets of different samples were labeled according to the reflective image feature indicators. Using the sample reflective region wound image set and the sample reflective image feature set as training data, a convolutional neural network was trained until convergence to construct a reflective image feature extraction model. The reflective image feature extraction model is used to extract reflective image features of reflective areas in diabetic foot lesion images.

3. The machine vision-assisted noise suppression method for diabetic wound images according to claim 1, characterized in that, Based on the preset feature identification rule table, the type of interference is determined by judging the interference type according to the reflective image features. The types of interference include exudate reflection, granulation tissue reflection and eschar reflection.

4. The machine vision-assisted noise suppression method for diabetic wound images according to claim 3, characterized in that, If the type of reflective interference is granulation tissue reflection, then the reflective component of the diabetic foot wound image is separated and reconstructed according to the first image processing strategy. The first image processing strategy is to eliminate highlights by estimating and compressing local illumination components, while protecting and enhancing the inherent reflective components of the object to preserve the color and texture features of the granulation tissue.

5. The machine vision-assisted noise suppression method for diabetic wound images according to claim 3, characterized in that, If the type of reflective interference is eschar reflection, then the diabetic foot wound image is subjected to edge-aware brightness suppression and boundary preservation processing according to the second image processing strategy. The second image processing strategy is to reduce the brightness of the reflective area at the edge of the eschar while maintaining a clear physical structural boundary between the eschar and normal tissue through guided filtering.

6. The machine vision-assisted noise suppression method for diabetic wound images according to claim 1, characterized in that, The steps to obtain the sample area error include: The initial wound area of ​​the original diabetic foot wound image without reflection noise suppression processing, corresponding to the reflection image features of the sample; Standard wound area of ​​a standard diabetic foot lesion image corresponding to the reflective image features of the collected sample, after reflective noise suppression processing; The ratio of the difference between the initial wound area and the standard wound area to the standard wound area is used as the sample area error.

7. The machine vision-assisted noise suppression method for diabetic wound images according to claim 6, characterized in that, Based on the reflected image features, an area error prediction complexity analysis is performed to obtain the error prediction complexity. Based on this error prediction complexity, K area error predictors, and the reflected image features, the wound area calculation error is predicted, and the predicted area error is output, including: A first feature discrimination matrix is ​​constructed based on the reflected image features; Construct a set of sample feature discrimination matrices based on the sample reflective image feature set; The first feature identification matrix is ​​compared with multiple sample feature identification matrices in the set of sample feature identification matrices to determine the similarity of multiple samples, and the average value is used to calculate the comprehensive sample similarity. The reciprocal of the comprehensive sample similarity is used as the error prediction complexity; Multiply the error prediction complexity by P and round down to get J, where P is the initial number of area error predictors selected, P is 3. If the calculated result of J is less than 1, then J is equal to 1. If the calculated result of J is greater than K, then J is equal to K. J area error predictors are randomly selected from the K area error predictors. The wound area calculation error is predicted based on the reflective image features. The mean of the J prediction results is then calculated to obtain the predicted area error.

8. The machine vision-assisted noise suppression method for diabetic wound images according to claim 1, characterized in that, An adaptive image denoising mechanism is formulated based on the predicted area error, including: A pre-constructed mapping table of area error and denoising strategy for exudate reflection is provided, wherein the denoising strategy includes an image reflection denoising algorithm, and the denoising intensity of the image reflection denoising algorithm is positively correlated with the area error. Using the area error-denoising strategy mapping table, the appropriate image denoising mechanism is determined based on the predicted area error matching.

9. A machine vision-assisted noise suppression system for diabetic wound images, characterized in that, A method for performing machine vision-assisted noise suppression of diabetic wound images according to any one of claims 1-8, comprising: The image feature extraction module is used to extract reflective image features of reflective areas in diabetic foot wound images through a convolutional neural network. An interference type determination module is used to determine the type of reflective interference based on the characteristics of the reflective image. The wound error prediction module is used to predict the wound area calculation error based on the reflective image features if the reflective interference type is exudate reflection, and to obtain the predicted area error. The wound image processing module is used to formulate an adaptive image denoising mechanism based on the predicted area error, and to perform reflective noise suppression processing on the diabetic foot wound image; The method of predicting the wound area calculation error based on the reflected image features and obtaining the predicted area error includes: Based on historical diabetes diagnosis records, and using exudate reflection as a constraint, a sample reflection image feature set was collected. The historical wound area calculation error in the historical detection process of different sample reflection image features was statistically analyzed as the sample area error, and a sample area error set was obtained. The sample reflective image feature set and sample area error set are used as training data, and K-fold cross-division is performed to obtain K sample training sets, where K is an integer greater than or equal to 8; The deep learning models are trained to convergence using the K sample training sets, generating K area error predictors. Based on the reflective image features, the area error prediction complexity is obtained through area error prediction complexity analysis. Based on the error prediction complexity, K area error predictors, and reflective image features, the wound area calculation error is predicted, and the predicted area error is output.

Images