Pressure injury classification system based on image processing

Abstract

This invention discloses a pressure injury classification system based on image processing, belonging to the field of image processing technology. The system includes: a pressure injury image acquisition module, a pressure injury image preprocessing module, a pressure injury classification model construction module, and a pressure injury classification module. Specifically, this invention obtains comprehensive noise feature values ​​based on pixel variance, LBP value variance, and distance weights; calculates adaptive filtering weights; and fuses bilateral filtering and Gaussian filtering for denoising. It then combines local entropy values ​​and Sobel gradient magnitudes to generate detail enhancement factors, enhancing details and local contrast, significantly improving image quality. Furthermore, it extracts and fuses multi-scale features, introduces a spatial-channel attention mechanism to obtain global weights, and obtains global and local enhancement fusion features, fusing them to generate comprehensive features, significantly improving the ability to identify pressure injuries at different stages.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically referring to a pressure injury classification system based on image processing. Background Technology

[0002] The pressure injury classification system is an intelligent system that uses advanced image processing technology to learn from a large amount of pressure injury image data of patients' skin, explore the characteristics and patterns, and accurately classify pressure injuries. It provides medical staff with objective and accurate basis for injury staging, helps to develop more targeted treatment and care plans, effectively improves the efficiency and quality of diagnosis and treatment of pressure injuries, and improves the patient's recovery process.

[0003] However, existing pressure injury classification systems suffer from problems such as unstable image acquisition quality, severe background interference, uneven lighting conditions, large differences in noise distribution, unclear damage details, and insufficient contrast, making image feature extraction difficult and affecting the accuracy of pressure injury classification. Furthermore, existing pressure injury classification systems exhibit characteristics of multi-scale, multi-morphological, blurred boundaries, and subtle feature differences at different stages. Traditional single-scale feature extraction methods are unable to fully capture these complex features, resulting in insufficient recognition capabilities for pressure injuries at different stages. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an image processing-based pressure injury classification system. Addressing the problems of unstable image acquisition quality, severe background interference, uneven lighting conditions, large noise distribution variations, unclear injury details, and insufficient contrast in existing pressure injury classification systems, which hinder image feature extraction and consequently affect the accuracy of pressure injury classification, this solution eliminates background interference by segmenting the pressure injury image. It then uses pixel variance, LBP variance, and distance weights to obtain comprehensive noise feature values ​​to assess noise intensity, avoiding information loss or residual noise caused by uniform denoising. Adaptive filtering weights are calculated, and bilateral filtering and Gaussian filtering are fused for denoising, enhancing image usability and feature discriminability. Finally, a detail enhancement factor is generated by combining local entropy and Sobel gradient magnitude to enhance details and local contrast. This approach extracts and fuses multi-scale features to enhance information exchange between different scales. A spatial-channel attention mechanism is introduced to obtain global weights, strengthening high-importance features and weighting low-importance features to obtain globally enhanced fusion features. Local enhanced fusion features are then obtained through channel grouping and weighting. These fusions generate comprehensive features, which are then used to classify and output image labels, enhancing the model's ability to comprehensively discriminate between different stages of pressure injuries and significantly improving the recognition capability of pressure injuries at different stages. The existing pressure injury classification systems suffer from the characteristics of multi-scale, multi-morphological, blurred boundaries, and subtle feature differences across different stages. Traditional single-scale feature extraction methods struggle to fully capture these complex features, leading to insufficient recognition capability for pressure injuries at different stages.

[0005] The pressure injury classification system based on image processing provided by this invention includes a pressure injury image acquisition module, a pressure injury image preprocessing module, a pressure injury classification model construction module, and a pressure injury classification module.

[0006] The pressure injury image acquisition module acquires historical images of pressure injuries to the patient's skin and the corresponding injury categories;

[0007] The pressure injury image preprocessing module segments the pressure injury image, obtains a comprehensive noise feature value based on pixel variance, LBP value variance and distance weight, evaluates the noise intensity, calculates adaptive filtering weights, and fuses bilateral filtering and Gaussian filtering for noise reduction. It also generates a detail enhancement factor by combining local entropy value and Sobel gradient magnitude to enhance details and local contrast, thus obtaining a processed image of the skin region.

[0008] The module for constructing a pressure injury classification model extracts and fuses multi-scale features, introduces a spatial-channel attention mechanism to obtain global weights, strengthens high-importance features and weights low-importance features to obtain globally enhanced fusion features, obtains locally enhanced fusion features through channel grouping and weighting, fuses to generate comprehensive features, classifies and outputs image labels;

[0009] The pressure injury classification module acquires real-time pressure injury images of the patient's skin, preprocesses them, and inputs them into the pressure injury classification model. Based on the output image labels, the corresponding injury category is obtained.

[0010] Furthermore, the pressure injury image acquisition module acquires historical images of pressure injuries to the patient's skin and the corresponding injury categories, and uses the injury categories as image labels to annotate the pressure injury images.

[0011] Furthermore, the pressure injury image preprocessing module includes a noise intensity assessment unit, an adaptive filtering and denoising unit, and an enhancement unit, specifically comprising the following:

[0012] The noise intensity assessment unit performs preliminary filtering on the pressure injury image to obtain a preliminary denoised image, and segments it into skin region images and background region images according to the optimal segmentation threshold. The skin region image is divided into non-overlapping initial blocks, and the pixel variance and LBP value variance of each initial block are calculated to obtain the dual feature heterogeneity. A splitting threshold is set. If the dual feature heterogeneity of the initial block is greater than the splitting threshold, the initial block is split into 4 sub-blocks. For each block, the pixel variance and LBP value variance are normalized, and then combined with the distance weight between the block center coordinates and the center coordinates of the skin region image, the comprehensive noise feature value of each block is obtained. The noise intensity is divided into three categories to obtain low noise intensity blocks, medium noise intensity blocks, and high noise intensity blocks.

[0013] Adaptive filtering denoising unit: For each location coordinate in the skin region image, the adaptive filtering weight is calculated based on the noise intensity of the block in which it is located. The results of bilateral filtering and Gaussian filtering are linearly weighted and fused using the adaptive filtering weight to obtain the denoised skin region image.

[0014] Enhancement unit: In the denoised image of the skin region, a 5×5 pixel neighborhood is taken centered at each location coordinate, the local entropy value and Sobel gradient magnitude of the neighborhood are calculated, and then combined with the generated detail enhancement factor, the skin region detail enhancement image is obtained by weighting the difference between the initial denoised image and the denoised image of the skin region, and local contrast enhancement is performed to obtain the processed image of the skin region.

[0015] Furthermore, the module for constructing the pressure injury classification model includes a multi-scale feature extraction unit, a fusion unit, and an attention feature fusion network unit, specifically comprising the following:

[0016] Multi-scale feature extraction unit: Extracts multi-scale features from skin region processed images, obtains four-layer features, and unifies the number of channels for all features;

[0017] The fusion unit upsamples the upper-layer features and adds them to the lower-layer features of the same size. It then repairs discontinuous edges through convolution to obtain coarse-grained, medium-grained, and fine-grained features. Finally, it uses sampling to allow features from different levels to interact at the same size. Combined with BN, ReLU, and 3×3 convolution optimization, it obtains the corresponding optimized features.

[0018] Attention feature fusion network unit; includes the following:

[0019] Spatial-channel attention mechanism: The size and number of channels of the optimized features are uniformly optimized to obtain coarse-grained unified features F1, medium-grained unified features F2 and fine-grained unified features F3, and channels are concatenated to generate multi-grained unified features; spatial pooling and channel pooling are performed on the multi-grained unified features respectively to obtain spatial attention weights and channel attention weights. After the size of both is matched with the multi-grained unified features, they are fused to obtain global weights.

[0020] Global and local enhancement: Global weights are averaged using pooling to obtain a global threshold. Based on this global threshold, F1, F2, and F3 are divided into element-wise thresholds to obtain their respective high-importance and low-importance features. High-importance features are concatenated, enhanced by depthwise separable convolution, and multiplied by the global weights. Low-importance features are concatenated and weighted using Softmax reconstruction. The enhanced high-importance features and weighted low-importance features are added element-wise to obtain the global enhancement fusion feature. Local enhancement features are obtained based on multi-granularity unified features. These features are evenly divided into 8 groups according to channels. Each group is generated with a weight distribution matching its own channel number through 1×1 convolution. After Softmax normalization, this weight is used as the channel weight for that group. The channel weights of each group are used to weight the features of that group element-wise, and the concatenation results in the local enhancement fusion feature.

[0021] Classify and output; add the global enhancement fusion feature and the local enhancement fusion feature element by element to obtain the comprehensive feature; perform global average pooling on the optimized feature and the comprehensive feature respectively to obtain 4 feature vectors, concatenate them and pass them through two fully connected layers and Softmax normalization, and output the damage category with the highest probability as the image label.

[0022] Furthermore, the pressure injury classification module acquires real-time pressure injury images of the patient's skin, preprocesses them, and inputs them into the pressure injury classification model. Based on the output image labels, the corresponding injury category is obtained, thus completing the real-time pressure injury classification of the patient's skin.

[0023] The beneficial effects achieved by adopting the above solution are as follows:

[0024] (1) To address the problems in existing pressure injury classification systems, such as unstable image acquisition quality, severe background interference, uneven lighting conditions, large differences in noise distribution, unclear injury details, and insufficient contrast, which make image feature extraction difficult and affect the accuracy of pressure injury classification, this solution segments pressure injury images to eliminate background interference and improve the targeting and accuracy of feature extraction; obtains comprehensive noise feature values ​​based on pixel variance, LBP value variance, and distance weight to evaluate noise intensity and achieve quantitative evaluation of noise intensity, avoiding information loss or residual noise caused by uniform denoising and improving the consistency and controllability of image quality; calculates adaptive filtering weights and integrates bilateral filtering and Gaussian filtering for denoising to enhance image usability and feature distinguishability; generates detail enhancement factors by combining local entropy value and Sobel gradient magnitude to enhance details and improve the model's ability to identify injuries at different stages; and enhances local contrast to obtain skin region processed images, significantly improving image quality and thus improving the accuracy of pressure injury staging.

[0025] (2) In response to the problem that existing pressure injury classification systems exhibit multi-scale, multi-morphological, blurred boundary, and subtle feature differences in pressure injuries at different stages, traditional single-scale feature extraction methods are unable to fully capture these complex features, resulting in insufficient recognition ability for pressure injuries at different stages. This scheme extracts multi-scale features and fuses them to fully cover the details and structural features of the injury and enhance the information interaction between features at different scales. A spatial-channel attention mechanism is introduced to obtain global weights, strengthen high-importance features and weighted low-importance features to obtain globally enhanced fusion features, strengthen features that contribute greatly to classification, and retain the auxiliary information of low-importance features. Local enhanced fusion features are obtained by weighting through channel grouping to enhance the model's ability to distinguish damage details. The fusion generates comprehensive features, classifies and outputs image labels, enhances the model's comprehensive ability to distinguish damage at different stages, and significantly improves the ability to recognize pressure injuries at different stages. Attached Figure Description

[0026] Figure 1 A schematic diagram of the image processing-based pressure injury classification system provided by the present invention;

[0027] Figure 2This is a schematic diagram of the pressure injury image preprocessing module;

[0028] Figure 3 A schematic diagram of the module for constructing a pressure injury classification model.

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", 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.

[0032] Example 1, see Figure 1 The pressure injury classification system based on image processing provided by the present invention includes a pressure injury image acquisition module, a pressure injury image preprocessing module, a pressure injury classification model construction module, and a pressure injury classification module.

[0033] The pressure injury image acquisition module acquires historical images of pressure injuries to the patient's skin and the corresponding injury categories, and sends the data to the pressure injury image preprocessing module;

[0034] The pressure injury image preprocessing module receives data sent by the pressure injury image acquisition module, segments the pressure injury image, obtains a comprehensive noise feature value based on pixel variance, LBP value variance and distance weight, evaluates the noise intensity, calculates adaptive filtering weights, and fuses bilateral filtering and Gaussian filtering for noise reduction. It also generates a detail enhancement factor by combining local entropy value and Sobel gradient magnitude to enhance details and local contrast, obtaining a processed skin region image, and sends the data to the pressure injury classification model construction module.

[0035] The pressure injury classification model module receives data sent by the pressure injury image preprocessing module, extracts and fuses multi-scale features, introduces a spatial-channel attention mechanism to obtain global weights, strengthens high-importance features and weights low-importance features to obtain global enhanced fusion features, obtains local enhanced fusion features through channel grouping and weighting, generates comprehensive features, classifies and outputs image labels, and sends the data to the pressure injury classification module.

[0036] The pressure injury classification module receives data sent by the pressure injury classification model building module, collects real-time pressure injury images of the patient's skin, preprocesses them, and inputs them into the pressure injury classification model. Based on the output image labels, the corresponding injury category is obtained.

[0037] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the pressure injury image acquisition module, historical pressure injury images of patients' skin and corresponding injury categories are acquired. The injury categories are used as image labels to annotate the pressure injury images. The injury categories include the ecchymosis stage, inflammatory infiltration stage, superficial ulcer stage, necrotic ulcer stage, unstageable stage, and deep tissue injury stage.

[0038] Example 3, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, includes a noise intensity evaluation unit, an adaptive filtering and denoising unit, and an enhancement unit in the pressure injury image preprocessing module, specifically comprising the following:

[0039] Noise intensity assessment unit: Pressure injury images come from diverse sources, with inconsistent formats and pixel ranges, and often contain irrelevant backgrounds that interfere with feature extraction from the injury area. Furthermore, the uneven noise distribution in pressure injury images makes uniform denoising prone to smoothing out damage details or incomplete denoising of high-noise areas. This unit divides the image into blocks and calculates comprehensive noise feature values ​​to accurately segment low, medium, and high noise intensity blocks, quantifying the noise intensity of different regions. This provides a basis for subsequent adaptive denoising, ensuring that high-noise intensity blocks are fully denoised while low-noise intensity blocks retain details, avoiding feature loss. The pressure injury images are then uniformly converted to 8-bit unsigned... The image is formatted as an integer and mapped to a standard range of [0, 255]. A 3×3 window median filtering algorithm is used to initially filter the pressure injury image, resulting in a preliminary denoised image. This preliminary denoised image is then converted to grayscale, and the Otsu algorithm is applied to calculate the optimal segmentation threshold. Based on the optimal segmentation threshold, the preliminary denoised image is segmented into skin region images and background region images. The skin region image is divided into non-overlapping 16×16 initial blocks. The pixel variance and LBP value variance of each initial block are calculated to obtain the dual-feature heterogeneity. The average dual-feature heterogeneity of all blocks is then used to determine the heterogeneity. Set split threshold If the heterogeneity of the initial block's dual features is greater than the splitting threshold, the initial block is split into four 8×8 sub-blocks. For each block, the pixel variance and LBP value variance are normalized, and then combined with the distance weight between the block center coordinates and the skin region image center coordinates to obtain the comprehensive noise feature value of each block. Based on the magnitude of the comprehensive noise feature value, the noise intensity is divided into three categories: low noise intensity block, medium noise intensity block, and high noise intensity block. The formula used is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] In the formula, and These are the pixel variance and LBP value variance of block A, respectively, and C. A L A and K A These are the dual-feature heterogeneity, comprehensive noise eigenvalue, and noise intensity classification identifier of block A, respectively, K. A 0, 1, and 2 represent that block A is a low-noise-intensity block, a medium-noise-intensity block, and a high-noise-intensity block, respectively; D A The center coordinates of block A Coordinates of the center of the skin region image The distance weights are given by R, where R is the diagonal length of the skin region image. and These are the maximum pixel variance and the maximum LBP value variance in all blocks, respectively. T1 is the classification threshold for low noise intensity blocks and medium noise intensity blocks, and T2 is the classification threshold for medium noise intensity blocks and high noise intensity blocks.

[0045] An adaptive filtering and denoising unit addresses the issue that a single filter cannot adequately handle the uneven noise in the damaged area. It dynamically adjusts the filter weights based on noise intensity classification, combining the advantages of two filters to remove noise while preserving key damage features to the greatest extent possible, thus retaining valuable information for subsequent feature extraction. For each coordinate location in the skin region image... The adaptive filtering weights are calculated based on the noise intensity of the block in which the image is located. The results of bilateral filtering and Gaussian filtering are then linearly weighted and fused using these adaptive filtering weights to obtain the denoised skin region image. The formula used is as follows:

[0046] ;

[0047] ;

[0048] In the formula, It is the location coordinates of the skin region image. Adaptive filter weights, It is the location coordinates of the skin region image. The noise intensity classification label of the block to which the location belongs. It is the location coordinates of the denoised image of the skin region. Pixel value at that location, and These are the location coordinates of the skin region images after bilateral filtering and Gaussian filtering, respectively. Pixel value at;

[0049] Enhancement Unit: Early pressure lesions often lack clear features, or lesion details are weakened after denoising, making it difficult for the model to capture stage differences. Enhancement factors are generated using local entropy and Sobel gradients to weight the difference between the denoised image and the initial denoised image, strengthening lesion details and highlighting low-contrast lesions, making subtle differences between different stages easier for the model to recognize. In denoised skin region images, enhancement factors are generated based on the coordinates of each location. A 5×5 pixel neighborhood is selected centered on the image. The local entropy and Sobel gradient magnitude of this neighborhood are calculated. The local entropy and Sobel gradient magnitude are combined to generate a detail enhancement factor. By weighting the difference between the initial denoised image and the denoised skin region image, a skin region detail enhancement image is obtained. Then, a contrast-limited adaptive histogram equalization algorithm is used to enhance the local contrast of the skin region detail enhancement image, resulting in the processed skin region image. The formulas used are as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] In the formula, and These are respectively based on position coordinates Let i be the local entropy value of the neighborhood centered at the pixel and the magnitude of the Sobel gradient. It represents the probability of pixel value i appearing in the neighborhood. and These are respectively based on position coordinates The horizontal and vertical gradients of the neighborhood centered on the center, E max and These are the maximum local entropy and the maximum Sobel gradient magnitude in the denoised skin region image, respectively. It is the location coordinates of the denoised image of the skin region. Detail enhancement factor, and These are the location coordinates of the skin region detail enhancement image and the preliminary noise reduction image, respectively. The pixel value at that location.

[0055] By performing the above operations, this solution addresses the problems in existing pressure injury classification systems, such as unstable image acquisition quality, severe background interference, uneven lighting conditions, large differences in noise distribution, unclear injury details, and insufficient contrast, which make image feature extraction difficult and affect the accuracy of pressure injury classification. This solution segments the pressure injury image to eliminate background interference, improving the targeting and accuracy of feature extraction. Based on pixel variance, LBP value variance, and distance weights, a comprehensive noise feature value is obtained to evaluate noise intensity, achieving quantitative evaluation of noise intensity and avoiding information loss or residual noise caused by uniform denoising, thus improving the consistency and controllability of image quality. Adaptive filtering weights are calculated, and bilateral filtering and Gaussian filtering are fused for denoising, enhancing the usability and feature distinguishability of the image. A detail enhancement factor is generated by combining local entropy and Sobel gradient magnitude to enhance details and improve the model's ability to identify injuries at different stages. Local contrast is also enhanced to obtain processed skin region images, significantly improving image quality and thus increasing the accuracy of pressure injury staging.

[0056] Example 4, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, includes a multi-scale feature extraction unit, a fusion unit, and an attention feature fusion network unit in the module for constructing a pressure injury classification model, specifically including the following:

[0057] Multi-scale feature extraction unit: Pressure lesions are multi-scale, and single-scale features cannot encompass the key information of different stages. ResNet50 is used to extract multi-scale features from the bottom to the top layer. The residual structure of ResNet50 avoids gradient vanishing and effectively captures features of different stages, providing a comprehensive feature foundation for subsequent fusion. ResNet50 is used to extract multi-scale features from the skin region processing image. Based on the output of the Conv2_x, Conv3_x, Conv4_x, and Conv5_x residual block groups, bottom-layer feature B1, middle-layer feature B2, high-layer feature B3, and top-layer feature B4 are obtained. A 1×1 convolution is used to unify the number of channels for all features to 256. The size of the skin region processing image is H×W, and the sizes of B1, B2, B3, and B4 are respectively... , , and ;

[0058] Fusion Unit: High- and low-level features exist in isolation and cannot reflect the local details and overall structural relationships of damage. Fusion allows coarse-grained features to contain more details and fine-grained features to contain more global information, improving the ability of features to represent different stages. After upsampling the upper-level features, they are added to the lower-level features of the same size. The discontinuous edges of the upsampling are repaired by 3×3 convolution, resulting in coarse-grained feature M1, medium-grained feature M2, and fine-grained feature M3. Then, sampling allows features from different levels to interact at the same size. Combined with BN, ReLU, and 3×3 convolution optimization, corresponding optimized features are obtained, including coarse-grained optimized features. Medium-grained optimization features and fine-grained optimization features The formula used is as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, It uses a bilinear interpolation algorithm to upsample the input features to twice the original size. It is a 3×3 convolution. It is an element-wise addition operator, and the dimensions of M1, M2, and M3 are respectively... , and , It uses a 3×3 convolution with a stride of 2 to downsample the input features to twice their original size. It uses a bilinear interpolation algorithm to upsample the input features to four times the original size. It is a modified linear unit activation function. It is a batch of normalization, , and The dimensions are respectively , and ;

[0066] Attention feature fusion network unit; includes the following:

[0067] Spatial-channel attention mechanism: Inconsistent feature sizes and channel numbers at different granularities prevent direct calculation of the attention mechanism, leading to fusion bias. To address this, a unified feature size and channel number ensure the attention mechanism applies evenly to all granularities, preventing the neglect of certain features due to format differences. Furthermore, irrelevant information within features can interfere with the model's focus on core damage features. Spatial attention highlights key damage locations, while channel attention emphasizes key feature channels, merging them into global weights. This allows the model to focus on crucial classification information and reduces interference from irrelevant information. Upsampled 2 times Downsampling by 2 times Size, and through 1×1 convolution , and The number of channels was uniformly set to 64, resulting in coarse-grained uniform features F1, medium-grained uniform features F2, and fine-grained uniform features F3. These features were then concatenated to generate multi-grained uniform features. Spatial pooling and channel pooling are performed on the multi-granularity unified feature F to obtain spatial attention weights and channel attention weights. After expanding the dimensions of both to match the size of the multi-granularity unified feature F, they are multiplied element-wise and fused to obtain the global weights. The formula used is as follows:

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, W s and W c These are spatial attention weights and channel attention weights, W b It is the global weight. and These are spatial max pooling and spatial average pooling, respectively. and These are channel max pooling and channel average pooling, respectively. It is the Sigmoid activation function. It is a 3×3 depthwise separable convolution. It's channel splicing. It is a multilayer perceptron. It is the element-wise multiplication operator. It is a dimension expansion function, where the dimensions of F1, F2, F3, and F are all 1. ;

[0072] Global and local enhancement: High-importance features are not enhanced, and auxiliary information of low-importance features is ignored, resulting in insufficient feature completeness. Based on a global threshold, high-importance and low-importance features are divided, high-importance features are enhanced, and low-importance features are weighted. This highlights the staging characteristics while retaining auxiliary information, improving the feature's ability to distinguish complex damage. Furthermore, the importance of different channels within the same feature granularity is uneven, with some key channels being submerged, affecting the discrimination of local details. After grouping features, key channels within a group are enhanced by channel weighting, improving the discriminative power of local features and enhancing the model's ability to identify subtle staging. A global threshold W is obtained by average pooling the global weights. th According to W th Element-level thresholding is applied to F1, F2, and F3 to obtain their respective high-importance and low-importance features. The high-importance features are concatenated, enhanced by depthwise separable convolution, and multiplied by global weights. The low-importance features are concatenated and weighted using Softmax reconstruction. The enhanced high-importance features are then element-wise added to the weighted low-importance features to obtain the globally enhanced fusion feature F. ef And based on the multi-granularity unified feature F, the local enhancement feature F is obtained. e , will F e The channels are evenly divided into 8 groups. Each group generates a weight distribution matching its own number of channels through a 1×1 convolution. After Softmax normalization, this weight is used as the channel weight for that group. The channel weights of each group are then used to weight the features of that group element by element. The concatenation results in the locally enhanced fusion feature F. wf The formula used is as follows:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula, , and These are the highly important features in F1, F2, and F3, respectively. , and These are the low-importance features in F1, F2, and F3, respectively; F h It is a highly important feature after enhancement. β represents the weighted low-importance features, and β is the reconstructed weight of the low-importance features. It is global average pooling. and They are F e The first and eighth group features, It is a normalized exponential function;

[0079] Classify and output the features; global and local enhancement features each have their own emphasis, and using them alone cannot fully represent the damage features. The comprehensive feature includes both the overall morphology and local details of the damage, providing a more comprehensive basis for classification; the global enhancement fusion feature and the local enhancement fusion feature are added element-wise to obtain the comprehensive feature F. all ;right , , and F all Global average pooling is performed separately to obtain four feature vectors. After concatenation, the vectors are processed through two fully connected layers and Softmax normalization to obtain the probability of each damage category. The damage category with the highest probability is output as the image label.

[0080] By performing the above operations, this scheme addresses the problem in existing pressure injury classification systems where pressure injuries exhibit multi-scale, multi-morphological, blurred boundary, and subtle feature differences across different stages. Traditional single-scale feature extraction methods struggle to fully capture these complex features, leading to insufficient recognition capabilities for pressure injuries at different stages. This solution extracts and fuses multi-scale features to comprehensively cover the details and structural features of the injury, enhancing information interaction between features at different scales. A spatial-channel attention mechanism is introduced to obtain global weights, strengthening high-importance features and weighting low-importance features to obtain globally enhanced fusion features. This strengthens features that contribute significantly to classification while retaining auxiliary information from low-importance features. Locally enhanced fusion features are obtained through channel grouping and weighting, enhancing the model's ability to discriminate injury details. Finally, comprehensive features are generated through fusion, classified, and image labels are output, further enhancing the model's comprehensive discrimination capability for injuries at different stages and significantly improving the recognition ability for pressure injuries at different stages.

[0081] Example 5, see Figure 1 This embodiment is based on the above embodiment. In the pressure injury classification module, real-time pressure injury images of the patient's skin are acquired, pre-processed, and then input into the pressure injury classification model. Based on the output image labels, the corresponding injury category is obtained, thus completing the real-time pressure injury classification of the patient's skin.

[0082] Example 6, see Figure 1This embodiment is based on the above embodiment and uses the "Pressure Ulcer Image Dataset" released by Brno University of Technology in 2021. This dataset contains standardized pressure injury images certified by the International Wound Care Association (IWCA), totaling 1860 images covering six injury stages: ecchymosis stage (320 images), inflammatory infiltration stage (350 images), superficial ulcer stage (380 images), necrotic ulcer stage (310 images), unstageable stage (280 images), and deep tissue injury stage (220 images). After preprocessing the dataset, it was divided into a training set:validation set:test set ratio of 7:1:2 to construct a pressure injury classification model. The classification ability of the pressure injury classification model was evaluated using the test set, with an accuracy rate of 95.7%. To verify the clinical applicability of the model, 200 real clinical real-time pressure injury images were collected from Zhejiang Cancer Hospital and Hangzhou Gongshu District People's Hospital of Integrated Traditional Chinese and Western Medicine, with an accuracy rate of 92.5%.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0085] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A pressure injury classification system based on image processing, characterized in that: It includes a pressure injury image acquisition module, a pressure injury image preprocessing module, a pressure injury classification model construction module, and a pressure injury classification module; The pressure injury image acquisition module acquires historical images of pressure injuries to the patient's skin and the corresponding injury categories, and uses the injury categories as image labels to annotate the pressure injury images. The pressure injury image preprocessing module segments the pressure injury image, obtains a comprehensive noise feature value based on pixel variance, LBP value variance and distance weight, evaluates the noise intensity, calculates adaptive filtering weights, and fuses bilateral filtering and Gaussian filtering for noise reduction. It also generates a detail enhancement factor by combining local entropy value and Sobel gradient magnitude to enhance details and local contrast, thus obtaining a processed image of the skin region. The module for constructing a pressure injury classification model extracts and fuses multi-scale features, introduces a spatial-channel attention mechanism to obtain global weights, strengthens high-importance features and weights low-importance features to obtain globally enhanced fusion features, obtains locally enhanced fusion features through channel grouping and weighting, fuses to generate comprehensive features, classifies and outputs image labels; The pressure injury classification module acquires real-time pressure injury images of the patient's skin, which are then preprocessed and input into the pressure injury classification model. Based on the output image labels, the corresponding injury category is obtained. The pressure injury image preprocessing module includes a noise intensity evaluation unit. It performs preliminary filtering on the pressure injury image to obtain a preliminary denoised image, and segments it into skin region images and background region images according to an optimal segmentation threshold. The skin region image is divided into non-overlapping initial blocks. The pixel variance and LBP value variance of each initial block are calculated to obtain the dual-feature heterogeneity. A splitting threshold is set; if the dual-feature heterogeneity of an initial block is greater than the splitting threshold, the initial block is split into four sub-blocks. For each block, the pixel variance and LBP value variance are normalized, and then combined with the distance weight between the block center coordinates and the skin region image center coordinates to obtain the comprehensive noise feature value of each block. The noise intensity is divided into three categories: low noise intensity blocks, medium noise intensity blocks, and high noise intensity blocks. The formula used is as follows: ； ； ； ； In the formula, and These are the pixel variance and LBP value variance of block A, respectively, and C. A L A and K A These are the dual-feature heterogeneity, comprehensive noise eigenvalue, and noise intensity classification identifier of block A, respectively, K. A 0, 1, and 2 respectively indicate that block A is a low noise intensity block, a medium noise intensity block, and a high noise intensity block; D A The center coordinates of block A Coordinates of the center of the skin region image The distance weights are given by R, where R is the diagonal length of the skin region image. and These are the maximum pixel variance and the maximum LBP value variance in all blocks, respectively. T1 is the classification threshold for low noise intensity blocks and medium noise intensity blocks, and T2 is the classification threshold for medium noise intensity blocks and high noise intensity blocks.

2. The image processing-based pressure injury classification system according to claim 1, characterized in that: The pressure injury image preprocessing module includes a noise intensity assessment unit, an adaptive filtering and denoising unit, and an enhancement unit, specifically comprising the following: Noise intensity assessment unit; Adaptive filtering denoising unit: For each location coordinate in the skin region image, the adaptive filtering weight is calculated based on the noise intensity of the block in which it is located. The results of bilateral filtering and Gaussian filtering are linearly weighted and fused using the adaptive filtering weight to obtain the denoised skin region image. Enhancement unit.

3. The image processing-based pressure injury classification system according to claim 2, characterized in that: The enhancement unit takes a 5×5 pixel neighborhood centered on each location coordinate in the denoised skin region image, calculates the local entropy value and Sobel gradient magnitude of the neighborhood, and then combines it with the generated detail enhancement factor. By weighting the difference between the initial denoised image and the denoised skin region image, a skin region detail enhancement image is obtained, and local contrast enhancement is performed to obtain the skin region processed image.

4. The image processing-based pressure injury classification system according to claim 3, characterized in that: The module for constructing the pressure injury classification model includes a multi-scale feature extraction unit, a fusion unit, and an attention feature fusion network unit, specifically comprising the following: Multi-scale feature extraction unit: Extracts multi-scale features from skin region processed images, obtains four-layer features, and unifies the number of channels for all features; The fusion unit upsamples the upper-layer features and adds them to the lower-layer features of the same size. It then repairs discontinuous edges through convolution to obtain coarse-grained, medium-grained, and fine-grained features. Finally, it uses sampling to allow features from different levels to interact at the same size. Combined with BN, ReLU, and 3×3 convolution optimization, it obtains the corresponding optimized features. Attention feature fusion network unit.

5. The image processing-based pressure injury classification system according to claim 4, characterized in that: The attention feature fusion network unit includes the following: Spatial-channel attention mechanism; uniformly optimize the size and number of channels of features to obtain coarse-grained unified features F1, medium-grained unified features F2 and fine-grained unified features F3, and concatenate the channels to generate multi-grained unified features; Spatial pooling and channel pooling are performed on the multi-granularity unified features respectively to obtain spatial attention weights and channel attention weights. After the size of both is matched with the multi-granularity unified features, they are fused to obtain the global weights. Global and local enhancements; Classify and output.

6. The image processing-based pressure injury classification system according to claim 5, characterized in that: The global and local enhancements involve average pooling the global weights to obtain a global threshold. Based on this global threshold, F1, F2, and F3 are divided into element-wise thresholds to obtain their respective high-importance and low-importance features. The high-importance features are concatenated, enhanced by depthwise separable convolution, and multiplied by the global weights. The low-importance features are concatenated and weighted using Softmax reconstruction. The enhanced high-importance features and the weighted low-importance features are then added element-wise to obtain the globally enhanced fusion feature F. ef And based on multi-granularity unified features, local enhancement features F are obtained. e The local enhancement features are evenly divided into 8 groups according to channels. Each group is generated with a weight distribution matching its own number of channels through 1×1 convolution. After Softmax normalization, the weights are used as the channel weights of the group. The channel weights of each group are then used to weight the features of that group element by element. The resulting concatenation yields the local enhancement fusion feature F. wf The formula used is as follows: ； ； ； ； ； In the formula, , and These are the highly important features in F1, F2, and F3, respectively. , and These are the low-importance features in F1, F2, and F3, respectively. F h It is a highly important feature after enhancement. β represents the weighted low-importance features, and β is the reconstructed weight of the low-importance features. It is global average pooling. and They are F e The first and eighth group features, It is a normalized exponential function, W b It is the global weight. It is a 3×3 depthwise separable convolution. It's channel splicing. It is a 1×1 convolution. It is the Sigmoid activation function. It is an element-wise addition operator. It is the element-wise multiplication operator, and F is a multi-granularity unified feature.

7. The image processing-based pressure injury classification system according to claim 6, characterized in that: The classification and output process involves adding the global enhanced fusion features and the local enhanced fusion features element by element to obtain the comprehensive features; Global average pooling is performed on the optimized features and the comprehensive features respectively to obtain four feature vectors. After concatenation, the vectors are normalized by two fully connected layers and Softmax, and the damage category with the highest probability is output as the image label.

Images