A pressure injury wound assessment method and system based on deep learning

Abstract

The application discloses a pressure injury wound surface evaluation method and system based on deep learning, relates to the technical field of pressure injury wound surface data processing, and comprises the following steps: acquiring RGB image data and exudation data of a pressure injury wound surface; obtaining encoder shallow layer features and high-dimensional feature maps; obtaining a high-dimensional enhanced region feature vector to form a globally associated enhanced feature map; obtaining a fused feature map; obtaining epidermal tissue area, granulation tissue area, necrotic tissue area and necrotic tissue area; outputting area prediction scores, tissue prediction scores and exudation correction scores; outputting a trend determination result, wherein the trend determination result comprises a healing trend, a stable trend or a deterioration trend; the pressure injury wound surface evaluation method based on deep learning automatically realizes four types of tissue pixel-level segmentation and area statistics through a network, combines with exudation quantitative data to complete three-dimensional dimension scoring of a PUSH scale, and gives a reference basis for a development trend of the pressure injury wound surface.

Description

Technical Field

[0001] This invention relates to the field of pressure injury wound data processing technology, and in particular, to a method and system for assessing pressure injury wounds based on deep learning. Background Technology

[0002] Utilizing software models and artificial intelligence for intelligent processing and predictive analysis of medical data is currently the mainstream application scenario in medical auxiliary diagnosis. Pressure injuries (pressure ulcers / bedsores) are localized injuries to the skin and subcutaneous tissues caused by pressure, shear force, or friction. They are highly prevalent in patients who are bedridden for extended periods, the elderly, malnourished, and have limited limb mobility. They not only significantly increase patients' physical and mental suffering and the risk of infection but also greatly enhance the difficulty of clinical nursing work and the medical burden, making them a common complication that is a key focus of clinical nursing care. Early, accurate, standardized, and continuous assessment and dynamic monitoring of pressure injury wounds are crucial for improving wound prognosis, reducing the incidence of complications, and optimizing clinical nursing protocols. Quantitative assessment of wounds and analysis of healing trends are core components of the diagnosis and treatment assessment of pressure injuries.

[0003] Currently, the internationally accepted gold standard for standardized assessment of pressure injuries is the Pressure Ulcer Healing Scale (PUSH) developed by the National Pressure Ulcer Advisory Panel (NPUAP). This scale quantitatively scores the healing status of pressure ulcers from three core dimensions: wound area, exudate volume, and wound tissue type. The total score ranges from 0 to 17 points, with a maximum score of 10 points for wound area, a maximum of 3 points for exudate volume, and a maximum of 4 points for tissue type. The higher the score, the more severe the pressure injury. A dynamic decrease in the score indicates a healing trend, a dynamic increase indicates a deterioration trend, and a stable score indicates a stable wound condition. It can objectively and comprehensively reflect the real-time status and outcome of pressure injuries and is currently the most authoritative and effective quantitative evaluation tool for clinically assessing pressure ulcer healing effects and guiding nursing interventions.

[0004] Currently, clinical assessment of pressure injuries using the PUSH scoring system primarily relies on manual observation, experience-based judgment, and manual calculation by healthcare professionals. This approach suffers from significant technical limitations and application constraints. Traditional manual assessment methods are highly subjective and subject to significant individual differences in experience, leading to inconsistent scoring results and low accuracy among different healthcare workers. Furthermore, manually calculating irregular wound areas, determining exudate levels, and differentiating between various tissue types (necrotic / slough / granulation / epidermal) is challenging and labor-intensive, easily resulting in missed or incorrect assessments. This makes standardized and precise quantitative assessment across the three dimensions of PUSH impossible. Especially for pressure ulcers with complex morphology and mixed necrotic and granulation tissue, traditional manual methods struggle to accurately distinguish tissue distribution and quantify exudate levels, severely impacting the accuracy and clinical reference value of PUSH scores.

[0005] While existing AI-assisted diagnostic technologies have been preliminarily applied to pressure injury detection—for example, patent applications for training methods, staging methods, and staging systems for pressure ulcers (patent numbers 2020106996384 and 202010699638.4)—these current solutions cannot meet the standardized quantitative assessment requirements of the PUSH scale. AI-assisted diagnostic technologies primarily focus on simple classification of single-frame wound images, rough measurement of wound size, or coarse staging, achieving only basic visual assistance and failing to construct a dedicated quantitative assessment system around the three core scoring dimensions of the PUSH scale: area, exudate, and tissue type.

[0006] Therefore, there is an urgent need for a three-dimensional pressure injury assessment method that conforms to the PUSH scale standard, so as to realize dynamic analysis of pressure ulcer outcome trends. This would solve or at least alleviate the technical problems in the existing technology, such as low quantification accuracy due to single-dimensional quantification, lack of dynamic prediction ability due to static assessment, and low accuracy when pressure ulcer trend judgment results are used as a reference due to insufficient identification dimensions. Summary of the Invention

[0007] This invention provides a method and system for assessing pressure injury wounds based on deep learning, aiming to solve the technical problem that the accuracy of the pressure ulcer trend judgment results provided when identifying and providing data reference for pressure injury wounds is low due to insufficient identification dimensions.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for assessing pressure injury wounds includes the following steps: S10: Acquire RGB image data of pressure injury wounds; acquire exudate data based on the percentage difference or weight difference of dressing penetration area; normalize the pixel values ​​of RGB image data and exudate data respectively. S20, the normalized RGB image data is downsampled at multiple levels to obtain the encoder's shallow features and high-dimensional feature maps; S30, determine the global attention weight matrix based on the self-attention mechanism for all high-dimensional initial region feature vectors in the high-dimensional feature map, and modify the high-dimensional initial region feature vectors based on the global attention weight matrix to obtain high-dimensional enhanced region feature vectors, forming a global association enhanced feature map; S40 fuses shallow features from the encoder with globally associated enhanced feature maps, performs multi-level upsampling and convolution processing to output fused feature maps; S50 segments and identifies the fused feature map, outputting pixel-level segmentation masks for four types of tissues: epidermis, granulation tissue, putrefaction tissue, and necrotic tissue. The pixel-level segmentation masks for the four types of tissues are statistically analyzed to obtain the area of ​​epidermal tissue, granulation tissue, putrefaction tissue, and necrotic tissue. S60 takes the high-dimensional enhanced region feature vector, epidermal tissue area, granulation tissue area, putrefactive tissue area, and necrotic tissue area as input, and uses the fully connected network of the comprehensive scoring prediction head to perform regression prediction, outputting area prediction score and tissue prediction score; it takes the normalized exudate data as input, and uses the fully connected network of the exudate scoring prediction head to perform regression prediction and correction, outputting exudate correction score. S70: A weighted calculation is performed based on area prediction score, tissue prediction score, and exudate correction score to obtain the total PUSH score; a sequence of total PUSH scores corresponding to continuous time-series nodes within the prediction time period is constructed; logical reasoning is performed based on the trend of the total PUSH score to output a trend determination result, which includes a healing trend, a stable trend, or a deterioration trend; and / or, using the node data sequence within the prediction time period as the input sequence, prediction is performed through the time-series prediction module, and a trend determination result is output, which includes a healing trend, a stable trend, or a deterioration trend; wherein, the node data includes the global association enhancement feature map, epidermal tissue area, granulation tissue area, necrotic tissue area, and exudate data corresponding to the time-series nodes.

[0009] Furthermore, the self-attention enhancement backbone network and tissue segmentation head are pre-trained, specifically including: constructing a training set using historical normalized RGB image data and corresponding pixel-level labels, iteratively training until the loss converges, and obtaining the trained self-attention enhancement backbone network and tissue segmentation head; During the model inference stage, the normalized RGB image of the pressure injury is input into the trained self-attention enhanced backbone network. After encoding, attention enhancement, and decoding operations, a fused feature map is obtained. The segmentation head performs pixel-level recognition on the fused feature map and outputs pixel-level segmentation masks for four types of tissues.

[0010] Furthermore, the comprehensive score prediction head is pre-trained, specifically including: fixing the parameters of the self-attention enhancement backbone network after training; inputting the historical normalized RGB image into the self-attention enhancement backbone network to extract the high-dimensional enhancement region feature vector, and combining it with the four types of tissue areas obtained by mask statistics; constructing a training set based on the area sub-labels and tissue type sub-labels marked by clinical experts according to the PUSH scale; using the mean squared error as the loss function, iteratively training until the loss converges to obtain the trained comprehensive score prediction head.

[0011] Furthermore, the pre-training of the exudate scoring prediction head specifically includes: constructing a training set using historical normalized dressing exudate data and corresponding exudate sub-labels marked by clinical experts based on the PUSH scale, iteratively training it with the mean squared error loss function until the loss converges, and obtaining the trained exudate scoring prediction head.

[0012] Furthermore, when making predictions through the time-series prediction module, the time-series prediction module is pre-trained, specifically including: using historical time-series node data sequences of the historical prediction time period as input, the time-series node data includes global association enhancement feature maps, epidermal tissue area, granulation tissue area, necrotic tissue area, necrotic tissue area, and exudate data; using wound trend labels with corresponding preset intervals as supervised output to construct a training set, iteratively training until the loss converges, and obtaining the trained time-series prediction module, the wound trend labels include three categories of labels: healing, stable, and deterioration.

[0013] Furthermore, in step S20, each grid element in the high-dimensional feature map has a high-dimensional initial region feature vector; In step S30, linear mapping is performed on all the high-dimensional initial region feature vectors in the high-dimensional feature map to generate query vector Q, key vector K, and value vector V. The inner product similarity between each query vector Q and all key vectors K is calculated and normalized to obtain the global attention weight matrix. The value vector V is weighted and summed using the global attention weight matrix to obtain the high-dimensional enhanced region feature vector corresponding to each grid element in the high-dimensional feature map, forming a global association enhanced feature map.

[0014] Furthermore, in step S60, the normalized seepage data is used as input, and regression prediction is performed through the fully connected network of the seepage score prediction head to output the seepage prediction score. The effusion prediction score is corrected based on clinical physical constraints to obtain the corrected effusion score, which includes: Based on the area prediction score and tissue prediction score corresponding to the current time node and the previous time node, the gradient score of the overall wound structure evolution of adjacent time nodes is calculated. The gradient score of the overall wound structure evolution is input into a pre-trained pressure ulcer pathology model to derive the range of exudate changes corresponding to the current time-series node. The pressure ulcer pathology model is a piecewise mapping fitting model with monotonic constraints. Training the model includes: constructing a clinical time-series training dataset based on real pressure injury diagnosis and treatment time-series samples; coupling the area prediction difference and tissue prediction difference of adjacent time-series nodes to obtain the wound structure evolution gradient features as training input; and using the upper and lower boundary values ​​of exudate fluctuations labeled by clinical experts as supervision labels. The parameters are iteratively optimized using the boundary fitting mean square error loss function until the training loss converges to a preset threshold, thus obtaining the pressure ulcer pathology model. Determine whether the current seepage gradient score is within the seepage variation range; if the current seepage gradient score is within the seepage variation range, then use the current seepage prediction score as the seepage correction score; if the current seepage gradient score is greater than the maximum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the maximum value of the seepage variation range to obtain the current seepage correction score; if the current seepage gradient score is less than the minimum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the minimum value of the seepage variation range to obtain the current seepage correction score.

[0015] Furthermore, dressing penetration images are acquired at a preset acquisition frequency to obtain the original dressing image sequence; The original dressing image sequence was filtered and denoised to obtain a purified dressing keyframe sequence. The keyframe sequence of dressings is processed by frame extraction to construct the effective frame sequence of dressings. The dressing target image is obtained by matching the effective frame sequence of the dressing with the temporal nodes; Based on the dressing target image, regression prediction is performed through a fully connected network of the exudate scoring prediction head to output the exudate prediction score.

[0016] This invention also provides a deep learning-based system for assessing pressure injury wounds. It includes a data acquisition module, a feature segmentation module, a rating prediction module, a time series prediction module, and a storage module; The data acquisition module is used to collect RGB image data of pressure injury wounds and dressing exudate data, and to normalize the pixel values ​​of the RGB images and the exudate data respectively. The feature segmentation module includes a self-attention enhancement backbone network and a tissue segmentation head. The self-attention enhancement backbone network consists of an encoder, a self-attention layer, and a decoder. The encoder performs multi-level downsampling on the normalized RGB image data, outputting shallow features and high-dimensional feature maps. The self-attention layer generates a globally correlated enhanced feature map based on the high-dimensional feature map. The decoder fuses the shallow features from the encoder and the globally correlated enhanced feature map, outputting a fused feature map. The tissue segmentation head performs pixel-level recognition on the fused feature map, outputting four types of tissue segmentation masks, and statistically obtaining the area of ​​each type of tissue based on the segmentation masks. The self-attention enhancement backbone network and the tissue segmentation head are trained from historical RGB image data and pixel labels until the loss converges. The scoring prediction module includes a comprehensive scoring prediction head and an exudate scoring prediction head. The comprehensive scoring prediction head is trained and converged from historical high-dimensional enhanced region feature vectors, epidermal tissue area, granulation tissue area, necrotic tissue area, and corresponding scoring labels. The exudate scoring prediction head is trained and converged from historical exudate data and corresponding exudate scoring labels, and is used to output an exudate prediction score based on normalized exudate data. The scoring prediction module is also used to correct the seepage prediction score to obtain the seepage correction score; The scoring prediction module is also used to output a total PUSH score based on the area prediction score, tissue prediction score, and exudate correction score; The time series prediction module is trained and converged from historical multi-dimensional time series node data sequences and trend labels. It is used to output trend judgment results based on the global correlation enhancement feature map, epidermal tissue area, granulation tissue area, putrefactive tissue area, necrotic tissue area and exudate data of continuous time series. The storage module is used to store the model training parameters of the self-attention enhancement backbone network, feature segmentation module, rating prediction module, and temporal prediction module.

[0017] The present invention has the following beneficial effects: This invention provides a deep learning-based method for assessing pressure injuries. It acquires the percentage difference in the osmotic area or weight of the dressing as exudate data, and obtains RGB image data of the wound after dressing removal and before cleaning as the pressure injury wound image data. The RGB images and measured exudate data of the pressure injury wound are then normalized and preprocessed. Next, encoding is performed to obtain shallow features and high-dimensional feature maps from the encoder. A self-attention enhanced backbone network is used to extract and adaptively strengthen the spatial adjacency association features of multiple tissues in the wound, completing pixel-level segmentation and area quantification statistics of four types of wound tissues: epidermis, granulation tissue, necrotic tissue, and necrotic tissue. The method effectively solves the problems of inaccurate tissue differentiation and large measurement errors of irregular wound area in traditional identification methods by identifying the area of ​​skin tissue, granulation tissue, necrotic tissue, and decayed tissue. Next, using high-dimensional enhanced region feature vectors, epidermal tissue area, granulation tissue area, necrotic tissue area, and decayed tissue area as input, a comprehensive scoring prediction head is used to obtain area prediction scores and tissue prediction scores. An exudate scoring prediction head outputs an exudate prediction score, which is then constrained and corrected by combining temporal tissue area changes to obtain an exudate correction score. A quantitative mapping is performed based on three dimensions: tissue type, wound area of ​​each tissue type, and exudate change, and the results are summed to obtain the PUSH total score. Finally, based on the trend of the PUSH total score, logical reasoning is used to output a trend judgment result, or a node data sequence within the prediction time period is used as input, and a time-series prediction module is used to predict and output a trend judgment result. This allows for the determination of wound development trends based on reasoning about the PUSH total score or mining the long-term dynamic evolution patterns of the wound, thus providing a reference basis based on data processing. This invention presents a deep learning-based method for assessing pressure injuries. It combines deep learning with a self-attention mechanism to intelligently assess pressure injuries. First, it automatically performs pixel-level segmentation and area statistics for four types of tissues via a network. Then, it uses quantified exudate data to complete a three-dimensional scoring of the PUSH scale. The assessment process is objective and standardized, overcoming the drawbacks of manual assessment, such as strong subjectivity, large errors, and heavy workload. Second, it utilizes a global attention mechanism to enhance the spatial correlation features of wound tissues, significantly improving the recognition accuracy of complex wounds. Based on static quantitative scoring, it adds a multi-dimensional temporal trend prediction function, combining static assessment with dynamic monitoring. This comprehensively addresses the technical problems of existing artificial intelligence solutions, such as their single-dimensionality and inability to dynamically assess wound progression.

[0018] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a deep learning-based method for assessing pressure injuries in one embodiment of the present invention. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Through research and analysis, the Pressure Ulcer Healing Scale (PUSH) quantitatively and comprehensively scores the healing status of pressure ulcers from three core dimensions: wound area, exudate volume, and wound tissue type. The total score range of the Pressure Ulcer Healing Scale is 0-17 points, with the highest score in the wound area dimension being 10 points, the highest score in the exudate volume dimension being 3 points, and the highest score in the tissue type dimension being 4 points during expert evaluation. The proposed solution uses photographs of the actual wounds evaluated by experts as input data in the training set, and the normalized results of the expert evaluation scores as the label results in the training set. In this invention, on the one hand, improvements are made to address the shortcomings of existing technologies, which cannot accurately segment and quantitatively count the area of ​​four types of wound tissues: epidermis, granulation tissue, necrotic tissue, and exudate, making it difficult to support accurate scoring of the PUSH area and tissue type dimensions. On the other hand, improvements are made to address the shortcomings of existing solutions, which rely solely on image visual features to indirectly estimate exudate status, lack an independent exudate quantification and correction mechanism, and cannot achieve objective quantitative scoring of the PUSH exudate dimension. All three dimensions of scoring suffer from insufficient accuracy and poor clinical relevance. Finally, improvements are made to address the shortcomings of existing technologies, which only support real-time assessment of a single static image and lack the ability to continuously track and dynamically analyze multi-temporal data of pressure injury wounds. These technologies cannot mine the dynamic outcome patterns of wounds based on continuous temporal data (wound tissue evolution data, exudate change data), making it difficult to dynamically predict the trends of wound healing, stabilization, and deterioration.

[0023] This invention provides a system for processing and predicting data on pressure injury wounds, comprising the following steps: S10: Acquire RGB image data of pressure injury wounds; acquire exudate data based on the percentage difference or weight difference of dressing penetration area; normalize the pixel values ​​of RGB image data and exudate data respectively. S20, the normalized RGB image data is downsampled at multiple levels to obtain the encoder's shallow features and high-dimensional feature maps; S30, determine the global attention weight matrix based on the self-attention mechanism for all high-dimensional initial region feature vectors in the high-dimensional feature map, and modify the high-dimensional initial region feature vectors based on the global attention weight matrix to obtain high-dimensional enhanced region feature vectors, forming a global association enhanced feature map; S40 fuses shallow features from the encoder with globally associated enhanced feature maps, performs multi-level upsampling and convolution processing to output fused feature maps; S50 segments and identifies the fused feature map, outputting pixel-level segmentation masks for four types of tissues: epidermis, granulation tissue, putrefaction tissue, and necrotic tissue. The pixel-level segmentation masks for the four types of tissues are statistically analyzed to obtain the area of ​​epidermal tissue, granulation tissue, putrefaction tissue, and necrotic tissue. S60 takes the high-dimensional enhanced region feature vector, epidermal tissue area, granulation tissue area, putrefactive tissue area, and necrotic tissue area as input, and uses the fully connected network of the comprehensive scoring prediction head to perform regression prediction, outputting area prediction score and tissue prediction score; it takes the normalized exudate data as input, and uses the fully connected network of the exudate scoring prediction head to perform regression prediction and correction, outputting exudate correction score. S70: A weighted calculation of the total PUSH score is performed based on the area prediction score, tissue prediction score, and exudate correction score; a sequence of total PUSH scores corresponding to consecutive time nodes within the prediction period is constructed; logical reasoning is used to output a trend determination result based on the trend of the total PUSH score, which includes healing trend, stable trend, or deterioration trend; and / or The node data sequence within the prediction time period is used as the input sequence. The time series prediction module performs prediction and outputs the trend determination result, which includes healing trend, stable trend or deterioration trend. The node data includes the global association enhancement feature map, epidermal tissue area, granulation tissue area, necrotic tissue area, and exudate data corresponding to the time series nodes.

[0024] The present invention provides a deep learning-based method for assessing pressure injuries. It obtains the percentage difference or weight difference of the dressing's permeation area as exudate data, and acquires RGB image data of the wound after dressing removal and before cleaning as pressure injury wound data. The RGB images of the pressure injury wound and the measured exudate data are then normalized and preprocessed. Next, encoding is performed to obtain shallow features and high-dimensional feature maps from the encoder. A self-attention enhanced backbone network is used to extract and adaptively strengthen the spatial adjacency association features of multiple wound tissues, completing pixel-level segmentation and area quantification statistics for four types of wound tissues: epidermis, granulation tissue, necrotic tissue, and necrotic tissue. This effectively solves the problems of inaccurate tissue differentiation and large measurement errors for irregular wound areas in traditional identification methods. Then… Using high-dimensional enhanced regional feature vectors, epidermal tissue area, granulation tissue area, necrotic tissue area, and necrotic tissue area as inputs, the system obtains area prediction scores and tissue prediction scores through a comprehensive scoring prediction head regression. The exudate prediction head outputs an exudate prediction score, which is then constrained and corrected by combining temporal tissue area changes to obtain an exudate correction score. A quantitative mapping is performed based on three dimensions: tissue type, wound area for each tissue type, and exudate change, and the results are summed to obtain the total PUSH score. Finally, based on the trend of the PUSH total score, logical reasoning is used to output a trend determination result. Alternatively, a node data sequence within the prediction time period can be used as input, and a time-series prediction module can be used to predict and output a trend determination result. This allows for the determination of wound development trends based on inferring the PUSH total score or mining the long-term dynamic evolution patterns of the wound, providing a reference basis based on data processing. This invention presents a deep learning-based method for assessing pressure injuries. It combines deep learning with a self-attention mechanism to intelligently assess pressure injuries. First, it automatically performs pixel-level segmentation and area statistics for four types of tissues via a network. Then, it uses quantified exudate data to complete a three-dimensional scoring of the PUSH scale. The assessment process is objective and standardized, overcoming the drawbacks of manual assessment, such as strong subjectivity, large errors, and heavy workload. Second, it utilizes a global attention mechanism to enhance the spatial correlation features of wound tissues, significantly improving the recognition accuracy of complex wounds. Based on static quantitative scoring, it adds a multi-dimensional temporal trend prediction function, combining static assessment with dynamic monitoring. This comprehensively addresses the technical problems of existing artificial intelligence solutions, such as their single-dimensionality and inability to dynamically assess wound progression.

[0025] Understandably, in a specific embodiment of the present invention, in S60, the 32×32×512-dimensional high-dimensional enhanced region feature vector, the normalized epidermal tissue area, granulation tissue area, putrefactive tissue area, and necrotic tissue area are used as inputs, and regression prediction is performed using a fully connected network of the comprehensive scoring prediction head to output area prediction score and tissue prediction score.

[0026] Understandably, in the solution of this invention, the dressing can be continuously photographed before removal, and the percentage difference of dressing penetration area at adjacent time points can be obtained by processing the photos; alternatively, the dressing can be changed at each time point, and the weight difference at adjacent time points can be obtained based on the weight difference between the wet dressing and the original dry dressing after removal. Specifically, in the pressure ulcer healing scale, the maximum score for exudate is 3 points, the maximum score for tissue type is 4 points, and the maximum score for wound area is 10 points, with a minimum score of 0 points for each. If the tissue area of ​​a certain type of tissue is less than the preset effective area, it is determined that this type of tissue does not exist, thereby realizing the determination of the tissue type and damage area of ​​the damage composition based on the pixel-level segmentation mask of the four types of tissues.

[0027] Understandably, exudate data refers to quantitative data related to body fluids exuded from pressure injury wounds. This invention uses two methods to quantify this data: the percentage difference in the area infiltrated by the dressing or the difference in the weight of the dressing. By comparing the changes in the area infiltrated by the dressing and the overall weight of the dressing before and after dressing changes, the amount of exudate from the wound is objectively characterized, achieving automated data acquisition, replacing manual visual grading, and providing more accurate quantitative results.

[0028] This invention employs a U-Net backbone encoder structure, using multi-level downsampling—that is, progressively compressing image resolution and increasing feature dimensionality through convolution and pooling operations. The shallow features of the encoder retain low-level visual information such as wound texture, tissue edges, and detailed contours, while the high-dimensional feature map extracts high-level semantic information such as wound tissue category and overall morphology. Specifically, after dividing the high-dimensional feature map into grids, each grid corresponds to a high-dimensional initial region feature vector, used to characterize the semantic features of the local wound region, serving as the basic unit for self-attention calculation. The global attention (similarity) weight matrix is ​​used to quantify the features of any two grid regions of the wound. The correlation degree is defined by weight values, where higher values ​​indicate a stronger correlation between two tissue regions. A weight matrix is ​​used to weight and correct the feature vectors of the high-dimensional initial region, overcoming the limitation of traditional convolution which can only capture local features. This establishes a global spatial correlation of the wound, focusing on strengthening the feature connections between necrotic tissue and necrotic tissue, granulation tissue and epidermis, and other clinically inherent adjacent tissues, thus improving the recognition ability of complex mixed tissue regions. The global correlation enhancement feature map integrates local details and global tissue correlation information, significantly improving feature representation capabilities. Upsampling is used to restore image resolution, fusing shallow detail features from the stitching encoder with high-level semantic features enhanced by attention.

[0029] In the solution of this invention, a multi-class mask image is used to achieve pixel-level segmentation mask. Each pixel is labeled as one of four types of tissue: epidermis, granulation tissue, putrefactive tissue, and necrotic tissue, thereby achieving accurate tissue division. The number of pixels of each type of tissue is counted based on the mask, and the actual tissue area is obtained by combining the physical scale of the image.

[0030] In one specific embodiment of the present invention, the area prediction score, tissue prediction score, and exudate correction score are added together to obtain the PUSH total score. The PUSH total score is a clinical static quantitative result, and the higher the score, the more severe the injury.

[0031] Furthermore, the self-attention enhancement backbone network and tissue segmentation head are pre-trained, specifically including: constructing a training set using historical normalized RGB image data and corresponding pixel-level labels, iteratively training until the loss converges, and obtaining the trained self-attention enhancement backbone network and tissue segmentation head; During the model inference phase, normalized RGB images of pressure injury wounds are input into the trained self-attention enhanced backbone network. After encoding, attention enhancement, and decoding operations, a fused feature map is obtained. The segmentation head performs pixel-level recognition on the fused feature map and outputs pixel-level segmentation masks for four types of tissues. Specifically, a training dataset is constructed based on historically labeled RGB images of pressure injury wounds and their corresponding pixel-level tissue labels. A pixel-level cross-entropy loss function is used, with normalized RGB images as input and pixel labels as ground truth for end-to-end joint training of the self-attention enhanced backbone network and the tissue segmentation head. Backpropagation iteratively updates the network weights, continuously reducing the segmentation loss until the validation set loss stabilizes and converges to a preset threshold, resulting in a trained and parameter-fixed self-attention enhanced backbone network and tissue segmentation head. Understandably, the input to the self-attention enhanced backbone network is normalized RGB image data, and the output includes shallow features from the encoder, high-dimensional feature maps, global association enhanced feature maps, and fused feature maps. The tissue segmentation head is used for pixel-level tissue classification. Through supervised learning, the network autonomously learns the visual features and spatial distribution patterns of four types of tissues to achieve the segmentation of four types of wound tissues. By setting the self-attention enhanced backbone network in conjunction with the tissue segmentation head, and simultaneously learning the underlying texture of the wound and global tissue association features, the segmentation accuracy and edge detail integrity of the four types of tissues are effectively improved, providing accurate and reliable basic data support for subsequent tissue area statistics and quantitative scoring.

[0032] Furthermore, the comprehensive score prediction head is pre-trained, specifically including: fixing the parameters of the self-attention enhancement backbone network after training; inputting the historical normalized RGB image into the self-attention enhancement backbone network to extract the high-dimensional enhancement region feature vector, and combining it with the four types of tissue areas obtained by mask statistics; constructing a training set based on the area sub-labels and tissue type sub-labels marked by clinical experts according to the PUSH scale; using the mean squared error as the loss function, iteratively training until the loss converges to obtain the trained comprehensive score prediction head. Specifically, during the pre-training of the comprehensive scoring prediction head, all parameters of the trained self-attention-enhanced backbone network and tissue segmentation head are fixed. Historical normalized RGB images are input into the trained backbone network and tissue segmentation head, outputting multi-class pixel-level segmentation masks for four tissue types: epidermis, granulation tissue, putrefaction, and necrosis. The number of pixels for each tissue type is counted based on the segmentation masks, and the true area of ​​epidermal tissue, granulation tissue, putrefaction tissue, and necrosis tissue is calculated by combining the physical scale of image acquisition. Simultaneously, a 32×32×512-dimensional high-dimensional enhanced region feature vector is extracted from the backbone network. The high-dimensional enhanced region feature vector is fused with the true areas of the four tissue types, and a training set is constructed using area sub-labels and tissue type sub-labels provided by clinical experts based on the PUSH scale. The training is iteratively trained using mean squared error as the loss function until the loss converges, resulting in the trained comprehensive scoring prediction head. Optionally, the high-dimensional enhanced region feature vector (e.g., 32×32×512-dimensional) is concatenated with the areas of the four tissue types for feature fusion.

[0033] Furthermore, the pre-training of the exudate scoring prediction head specifically includes: constructing a training set using historical normalized dressing exudate data and corresponding exudate sub-labels marked by clinical experts based on the PUSH scale, iteratively training it with the mean squared error loss function until the loss converges, and obtaining the trained exudate scoring prediction head.

[0034] This implementation method separately trains models for the comprehensive scoring prediction head and the exudate scoring prediction head. The comprehensive scoring prediction head is a two-branch fully connected regression network, and all parameters of the self-attention enhanced backbone network and the tissue segmentation head are fixed during training. Historical normalized RGB images are input into the backbone network to extract 32×32-dimensional high-dimensional enhanced region feature vectors. At the same time, the tissue segmentation head outputs pixel-level segmentation masks for four types of tissues. Based on the number of pixels in the masks and combined with the physical scale conversion of the image, the true areas of epidermis, granulation tissue, necrotic tissue, and decayed tissue are obtained. The high-dimensional enhanced region feature vectors are fused with the true areas of the four types of tissues, and a training set is constructed by combining the area sub-labels and tissue type sub-labels marked by clinical experts according to the PUSH scale. Iterative training is performed using mean squared error as the loss function, so that the network learns the scoring mapping relationship after the fusion of global wound features and tissue quantitative information until the loss converges, ensuring that the output score conforms to international clinical assessment standards and avoiding subjective bias caused by manual scoring. The exudate scoring prediction head is an independent fully connected regression branch that relies on historical normalized dressing exudate quantification data and corresponding PUSH standard exudate score labels to complete iterative training. It accurately establishes the mapping relationship between the exudate quantification value and the initial exudate score until the loss converges, resulting in a fully trained exudate scoring prediction head. Its output initial exudate score has a clinical reference basis, and after being corrected by clinical physical constraints based on time-series changes, the accuracy of the exudate score can be further improved.

[0035] Furthermore, when making predictions through the time-series prediction module, the time-series prediction module is pre-trained, specifically including: using historical time-series node data sequences of the historical prediction time period as input, the time-series node data includes global association enhancement feature maps, epidermal tissue area, granulation tissue area, necrotic tissue area, necrotic tissue area, and exudate data; using wound trend labels with corresponding preset intervals as supervised output to construct a training set, iteratively training until the loss converges, and obtaining the trained time-series prediction module, the wound trend labels include three categories of labels: healing, stable, and deterioration. When training the time-series prediction module, the node data and wound trend labels in the training set are first normalized. After integrating multi-dimensional time-series dynamic data and fully capturing the long-term evolution pattern of pressure injury wounds, dynamic prediction of wound healing, stabilization, and deterioration trends is achieved. The prediction time period can be several hours, one day, or several days, depending on the actual situation. The interval between time-series nodes can be hours or days, and a prediction time period must include at least two time-series nodes. The preset interval can be several hours, one day, or several days, as long as the result labels corresponding to the preset interval are used for training.

[0036] Further, in step S20, each grid element in the high-dimensional feature map has a high-dimensional initial region feature vector; in step S30, all high-dimensional initial region feature vectors in the high-dimensional feature map are linearly mapped to generate query vector Q, key vector K, and value vector V; the inner product similarity between each query vector Q and all key vectors K is calculated, and the global attention (similarity) weight matrix is ​​obtained by Softmax normalization. The value vector V is weighted and summed using the global attention (similarity) weight matrix to obtain the high-dimensional enhanced region feature vector corresponding to each grid element in the high-dimensional feature map, forming the high-dimensional enhanced region feature vector corresponding to each grid element in the globally associated enhanced feature map, thus forming the globally associated enhanced feature map. The solution of this invention establishes the global association relationship of grid features across the entire image through a self-attention mechanism, breaking through the limitations of the local receptive field of traditional convolution, fully exploring the spatial distribution features of different tissues in the wound, strengthening feature expression capabilities, and significantly improving the robustness of subsequent segmentation and scoring tasks.

[0037] Furthermore, the global attention weight matrix satisfies the following: the attention weight of the feature grid corresponding to necrotic tissue to the feature grid of neighboring slough tissue is higher than the attention weight to the feature grid of other tissues; the attention weight of the feature grid corresponding to granulation tissue to the feature grid of neighboring epidermal tissue is higher than the attention weight to the feature grid of other tissues, thereby adaptively strengthening the spatial adjacency association features between different tissues in pressure injury wounds. In the scheme of this invention, based on the tissue distribution law that necrotic tissue is usually accompanied by slough tissue around pressure injury in clinical practice, and that newly formed granulation tissue gradually transforms into epidermal tissue, the attention weights are set differently to conform to prior pathological knowledge, guide the model to focus on adjacent tissue areas with high clinical correlation, further improve the accuracy of tissue boundary recognition and mixed tissue differentiation, and make the feature expression more in line with objective medical laws.

[0038] Furthermore, when mapping the percentage difference of the dressing's permeation area to the exudate dimension for exudate scoring, the exudate scoring has significant limitations due to the inconsistency in the adsorption capacity of different dressings or the inconsistency in the adsorption capacity of the same dressing at different wetting stages. Therefore, in step S60, the normalized exudate data is used as input, and regression prediction is performed through the fully connected network of the exudate scoring prediction head to output the exudate prediction score. The effusion prediction score is corrected based on clinical physical constraints to obtain the corrected effusion score, which includes: Based on the area prediction score and tissue prediction score corresponding to the current time node and the previous time node, the gradient score of the overall wound structure evolution of adjacent time nodes is calculated. The gradient score of the overall wound structure evolution is input into a pre-trained pressure ulcer pathology model to derive the range of exudate changes corresponding to the current time-series node. The pressure ulcer pathology model is a piecewise mapping fitting model with monotonic constraints. Training the model includes: constructing a clinical time-series training dataset based on real pressure injury diagnosis and treatment time-series samples; coupling the area prediction difference and tissue prediction difference of adjacent time-series nodes to obtain the wound structure evolution gradient features as training input; and using the upper and lower boundary values ​​of exudate fluctuations labeled by clinical experts as supervision labels. The parameters are iteratively optimized using the boundary fitting mean square error loss function until the training loss converges to a preset threshold, thus obtaining the pressure ulcer pathology model. Determine whether the current seepage gradient score is within the seepage variation range; if the current seepage gradient score is within the seepage variation range, then use the current seepage prediction score as the seepage correction score; if the current seepage gradient score is greater than the maximum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the maximum value of the seepage variation range to obtain the current seepage correction score; if the current seepage gradient score is less than the minimum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the minimum value of the seepage variation range to obtain the current seepage correction score.

[0039] Specifically, the pressure ulcer pathology relationship model is a piecewise mapping fitting model with monotonic constraints. The training of the pressure ulcer pathology relationship model includes: constructing a clinical time-series training dataset based on real pressure injury diagnosis and treatment time-series samples; coupling the area prediction score difference and tissue prediction score difference of adjacent time-series nodes to obtain the wound structure evolution gradient features as training input; and using the reasonable upper and lower boundary values ​​of exudate fluctuation marked by clinical experts according to pathological healing rules as supervision labels. The training process introduces pathological prior monotonicity constraint regularization, which limits the model to satisfy the following: as the wound deterioration gradient increases, the exudate fluctuation range widens; as the wound repair gradient increases, the exudate fluctuation range narrows. The parameters are iteratively optimized using the boundary fitting mean square error loss function until the training loss converges to a preset threshold to obtain the pressure ulcer pathology relationship model. During model inference, the area prediction score and tissue prediction score of the current and previous time-series nodes are extracted, the time-series difference is calculated and coupled to obtain the overall wound structure evolution gradient score, which is input into the pressure ulcer pathology relationship model with fixed parameters, and the reasonable dynamic change range of exudate corresponding to the current time-series is output.

[0040] Furthermore, dressing penetration images are acquired at a preset acquisition frequency to obtain the original dressing image sequence; The original dressing image sequence was filtered and denoised to obtain a purified dressing keyframe sequence. The keyframe sequence of dressings is processed by frame extraction to construct the effective frame sequence of dressings. The dressing target image is obtained by matching the effective frame sequence of the dressing with the temporal nodes; Based on the dressing target image, regression prediction is performed through a fully connected network of the exudate scoring prediction head to output the exudate prediction score.

[0041] The present invention also provides a data processing and prediction system for pressure injury wounds, including a data acquisition module, a feature segmentation module, a scoring prediction module, a time series prediction module, and a storage module; The data acquisition module is used to collect RGB image data of pressure injury wounds and dressing exudate data, and to normalize the pixel values ​​of the RGB images and the exudate data respectively. The feature segmentation module includes a self-attention enhancement backbone network and a tissue segmentation head. The self-attention enhancement backbone network consists of an encoder, a self-attention layer, and a decoder. The encoder performs multi-level downsampling on the normalized RGB image data, outputting shallow features and a high-dimensional feature map. The self-attention layer generates a globally correlated enhanced feature map based on the high-dimensional feature map. The decoder fuses the shallow features from the encoder and the globally correlated enhanced feature map, outputting a fused feature map. The tissue segmentation head performs pixel-level recognition on the fused feature map, outputting four types of tissue segmentation masks. Based on the segmentation masks, the number of pixels is counted, and combined with the physical scale conversion of the image, the area of ​​the epidermal tissue is obtained. The system calculates the area of ​​granulation tissue, necrotic tissue, and necrotic tissue. The self-attention-enhanced backbone network and tissue segmentation head are trained on historical RGB image data and pixel labels until the loss converges. The scoring prediction module includes a comprehensive scoring prediction head and an exudate scoring prediction head. The comprehensive scoring prediction head is trained on historical 32×32×512-dimensional high-dimensional enhanced region feature vectors and the corresponding four types of tissue areas until convergence. It takes the high-dimensional enhanced region feature vectors and the four types of tissue areas as input and outputs area prediction scores and tissue prediction scores. The exudate scoring prediction head is trained on historical exudate data and corresponding exudate scoring labels and is used to output exudate prediction scores based on normalized exudate data. The scoring prediction module is also used to correct the seepage prediction score to obtain the seepage correction score; The scoring prediction module is also used to output a total PUSH score based on the area prediction score, tissue prediction score, and exudate correction score; The time series prediction module is trained and converged from historical multi-dimensional time series node data sequences and trend labels. It is used to output trend judgment results based on the global correlation enhancement feature map, epidermal tissue area, granulation tissue area, putrefactive tissue area, necrotic tissue area and exudate data of continuous time series. The storage module is used to store the model training parameters of the self-attention enhancement backbone network, feature segmentation module, rating prediction module, and temporal prediction module.

[0042] This invention provides a specific method for assessing pressure injury wounds based on deep learning. It is implemented through a deep learning-based pressure injury wound assessment system. The feature segmentation module of this system includes a self-attention enhancement backbone network. This network encodes the normalized RGB image data (RGB image of the pressure injury wound) to obtain a 32×32×512 high-dimensional feature map. The high-dimensional feature map is then enhanced using a self-attention enhancement mechanism to obtain a 32×32×512 high-dimensional enhanced region feature vector, i.e., a globally correlated enhanced feature map. The method includes the following steps: Step 1: Collect RGB images of pressure injury wounds and scale them to 512×512×3. Obtain exudate data by the percentage difference in dressing penetration area or weight difference. Normalize the pixel values ​​of the RGB images to [0,1] and normalize the exudate data to [0,1]. Step 2: A self-attention enhanced backbone network is used to perform four-layer convolutional downsampling on the RGB image of the pressure injury wound, progressively compressing the resolution and increasing the channel dimension. The output is a shallow encoder feature that preserves the low-level details of the wound edge, texture, and contour, as well as a 32×32×512 high-dimensional feature map containing high-level semantic features of the wound. High-dimensional initial region feature vectors are extracted grid by grid from the 32×32×512 high-dimensional feature map. A global attention weight matrix is ​​calculated through a self-attention mechanism to weight and correct the feature vectors, resulting in a 32×32×512 high-dimensional enhanced region feature vector, and a global association enhanced feature map is generated. The shallow encoder features and the global association enhanced feature map are fused, and the resolution is restored through multi-level upsampling and convolution to output a fused feature map. The tissue segmentation head performs pixel-level classification on the fused feature map, outputting four types of tissue segmentation masks: epidermis, granulation tissue, necrotic tissue, and necrotic tissue. Based on the number of pixels in the masks and combined with the physical scale conversion of the image, the area of ​​epidermal tissue, granulation tissue, necrotic tissue, and necrotic tissue are obtained. Step 3: Based on all parameters of the fixed self-attention enhancement (U-Net) backbone network, mapping is performed on different fully connected layers; specifically, the 32×32×512-dimensional enhancement region feature vector and the areas of epidermal tissue, granulation tissue, necrotic tissue, and putrefactive tissue are used as fusion feature inputs, and regression prediction is performed using the dual-branch fully connected network of the comprehensive scoring prediction head. The output values ​​from 0 to 10 are used as area prediction scores, and the output values ​​from 0 to 4 are used as tissue prediction scores. Using normalized exudate data as input, the exudate prediction score is output from 0 to 3 through a fully connected network regression of the exudate score prediction head; the range constraint correction of the exudate prediction score is obtained based on the pressure ulcer pathology relationship model that combines the temporal evolution gradient of the wound with the pathological pattern. Step 4: Sum the area prediction score (0-10), tissue prediction score (0-4), and exudate correction score (0-3) to obtain the PUSH total score, which serves as the static quantitative assessment result of the wound. Step 5: Based on the trend of the total PUSH score, perform logical reasoning to output the trend determination result, and / or, perform prediction through the time series prediction module to output the trend determination result. The node data includes the global association enhancement feature map, epidermal tissue area, granulation tissue area, necrotic tissue area, and exudate data corresponding to the time series node; the trend determination result includes healing trend, stable trend, or deterioration trend.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing pressure injury wounds based on deep learning, characterized in that, Including the following steps: S10, acquire RGB image data of the pressure injury wound; acquire exudate data based on the difference in percentage of dressing penetration area or weight difference; normalize the pixel values ​​of the RGB image data and the exudate data respectively; S20, the normalized RGB image data is downsampled at multiple levels to obtain shallow features and high-dimensional feature maps of the encoder; S30, determine the global attention weight matrix based on the self-attention mechanism for all high-dimensional initial region feature vectors in the high-dimensional feature map, and modify the high-dimensional initial region feature vectors based on the global attention weight matrix to obtain high-dimensional enhanced region feature vectors, forming a global association enhanced feature map; S40, fuse the shallow features of the encoder with the global association enhancement feature map, perform multi-level upsampling and convolution processing to output the fused feature map; S50, the fused feature map is segmented and identified to output pixel-level segmentation masks for four types of tissues: epidermis, granulation tissue, necrotic tissue, and sludge. The pixel-level segmentation masks of the four types of tissues were statistically analyzed to obtain the area of ​​epidermal tissue, granulation tissue, putrefactive tissue, and necrotic tissue. S60, using the high-dimensional enhanced region feature vector, the area of ​​epidermal tissue, the area of ​​granulation tissue, the area of ​​putrefied tissue, and the area of ​​necrotic tissue as input, regression prediction is performed using the fully connected network of the comprehensive scoring prediction head, and the area prediction score and tissue prediction score are output; using the normalized exudate data as input, regression prediction is performed and corrected through the fully connected network of the exudate scoring prediction head, and the exudate correction score is output. S70: The total PUSH score is obtained by weighted calculation based on the area prediction score, tissue prediction score, and exudate correction score; the total PUSH score sequence corresponding to the continuous time nodes within the prediction time period is constructed; the trend judgment result is output by logical reasoning based on the trend of the total PUSH score change, and the trend judgment result includes healing trend, stable trend, or deterioration trend. and / or Using the node data sequence within the prediction time period as the input sequence, the time series prediction module performs prediction and outputs a trend determination result, which includes a healing trend, a stable trend, or a deterioration trend; wherein, the node data includes the global association enhancement feature map, epidermal tissue area, granulation tissue area, necrotic tissue area, and exudate data corresponding to the time series node.

2. The method for assessing pressure injury wounds based on deep learning according to claim 1, characterized in that, Pre-training the self-attention enhancement backbone network and tissue segmentation head specifically includes: constructing a training set using historical normalized RGB image data and corresponding pixel-level labels, iteratively training until the loss converges, and obtaining the trained self-attention enhancement backbone network and tissue segmentation head; During the model inference stage, the normalized RGB image of the pressure injury wound is input into the trained self-attention enhanced backbone network. After encoding, attention enhancement, and decoding operations, a fused feature map is obtained. The segmentation head performs pixel-level recognition on the fused feature map and outputs pixel-level segmentation masks for four types of tissues.

3. The method for assessing pressure injury wounds based on deep learning according to claim 2, characterized in that, The pre-training of the comprehensive score prediction head specifically includes: fixing the parameters of the self-attention enhancement backbone network after training; inputting the historical normalized RGB image into the self-attention enhancement backbone network to extract the high-dimensional enhancement region feature vector, and combining it with the four types of tissue areas obtained by mask statistics; constructing a training set based on the area sub-labels and tissue type sub-labels marked by clinical experts according to the PUSH scale; using the mean squared error as the loss function, iteratively training until the loss converges to obtain the trained comprehensive score prediction head.

4. The method for assessing pressure injury wounds based on deep learning according to claim 1, characterized in that, The pre-training of the exudate scoring prediction head specifically includes: constructing a training set using historical normalized dressing exudate data and corresponding exudate sub-labels marked by clinical experts based on the PUSH scale; iteratively training the head using the mean squared error loss function until the loss converges; and obtaining the trained exudate scoring prediction head.

5. The method for assessing pressure injury wounds based on deep learning according to claim 1, characterized in that, When making predictions using the time-series prediction module, the time-series prediction module is pre-trained. Specifically, it includes: using historical time-series node data sequences of the prediction time period as input, the time-series node data including global association enhancement feature maps, epidermal tissue area, granulation tissue area, necrotic tissue area, and exudate data; constructing a training set using wound trend labels with corresponding preset intervals as supervised output, iterating training until the loss converges, and obtaining the trained time-series prediction module. The wound trend labels include three categories: healing, stable, and deteriorating.

6. A method for assessing pressure injury wounds based on deep learning according to any one of claims 1 to 5, characterized in that, In step S20, each grid element in the high-dimensional feature map has a high-dimensional initial region feature vector; In step S30, linear mapping is performed on all the high-dimensional initial region feature vectors in the high-dimensional feature map to generate query vector Q, key vector K, and value vector V. The inner product similarity between each query vector Q and all key vectors K is calculated and normalized to obtain a global attention weight matrix. The value vector V is weighted and summed using the global attention weight matrix to obtain the high-dimensional enhanced region feature vector corresponding to each grid element in the high-dimensional feature map, forming a global association enhanced feature map.

7. A method for assessing pressure injury wounds based on deep learning according to any one of claims 1 to 5, characterized in that, In step S60, the normalized seepage data is used as input, and regression prediction is performed through the fully connected network of the seepage score prediction head to output the seepage prediction score. The effusion prediction score is corrected based on clinical physical constraints to obtain the corrected effusion score, which includes: Based on the area prediction score and tissue prediction score corresponding to the current time node and the previous time node, the gradient score of the overall wound structure evolution of adjacent time nodes is calculated. The gradient score of the overall wound structure evolution is input into a pre-trained pressure ulcer pathology relationship model to derive the range of exudate changes corresponding to the current time node. The pressure ulcer pathology relationship model is a piecewise mapping fitting model with monotonic constraints. The training of the pressure ulcer pathology relationship model includes: constructing a clinical time-series training dataset based on real pressure injury diagnosis and treatment time-series samples; coupling the area prediction difference and tissue prediction difference of adjacent time-series nodes to obtain the wound structure evolution gradient features as training input; and using the upper and lower boundary values ​​of exudate fluctuations labeled by clinical experts as supervision labels; iteratively optimizing the parameters using the boundary fitting mean square error loss function until the training loss converges to a preset threshold to obtain the pressure ulcer pathology relationship model. Determine whether the current seepage gradient score is within the seepage variation range; if the current seepage gradient score is within the seepage variation range, then use the current seepage prediction score as the seepage correction score; if the current seepage gradient score is greater than the maximum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the maximum value of the seepage variation range to obtain the current seepage correction score; if the current seepage gradient score is less than the minimum value of the seepage variation range, then combine the seepage correction score of the previous time series node with the minimum value of the seepage variation range to obtain the current seepage correction score.

8. The method for assessing pressure injury wounds based on deep learning according to claim 7, characterized in that, Specifically, obtaining exudate data based on the difference in the percentage of dressing penetration area includes: Acquire dressing penetration images at a preset acquisition frequency to obtain the original dressing image sequence; The original dressing image sequence is filtered and denoised to obtain a purified dressing keyframe sequence. The keyframe sequence of the dressing is subjected to frame extraction processing to construct an effective frame sequence of the dressing; The dressing target image is obtained by matching the effective frame sequence of the dressing with the time sequence nodes; Based on the target image of the dressing, regression prediction is performed through a fully connected network of the exudate scoring prediction head to output an exudate prediction score.

9. A deep learning-based system for assessing pressure-related wounds, characterized in that, It includes a data acquisition module, a feature segmentation module, a rating prediction module, a time series prediction module, and a storage module; The data acquisition module is used to collect RGB image data of pressure injury wounds and dressing exudate data, and to normalize the pixel values ​​of the RGB images and the exudate data respectively. The feature segmentation module includes a self-attention enhancement backbone network and a tissue segmentation head; the self-attention enhancement backbone network consists of an encoder, a self-attention layer, and a decoder; the encoder performs multi-level downsampling on the normalized RGB image data and outputs shallow features and high-dimensional feature maps of the encoder. The self-attention layer generates a global association enhancement feature map based on the high-dimensional feature map; the decoder fuses the shallow features of the encoder and the global association enhancement feature map to output a fused feature map. The tissue segmentation head performs pixel-level recognition on the fused feature map, outputs four types of tissue segmentation masks, and calculates the area of ​​each type of tissue based on the segmentation masks. The self-attention enhancement backbone network and tissue segmentation head are trained on historical RGB image data and pixel labels until the loss converges; the scoring prediction module includes a comprehensive scoring prediction head and an exudate scoring prediction head; the comprehensive scoring prediction head is trained and converged on historical high-dimensional enhanced region feature vectors, epidermal tissue area, granulation tissue area, necrotic tissue area, and corresponding scoring labels; the exudate scoring prediction head is trained and converged on historical exudate data and corresponding exudate scoring labels, and is used to output an exudate prediction score based on normalized exudate data. The scoring prediction module is also used to correct the seepage prediction score to obtain a seepage correction score; The scoring prediction module is also used to output a total PUSH score based on the area prediction score, tissue prediction score, and exudate correction score. The time series prediction module is trained and converged by historical multi-dimensional time series node data sequences and trend labels. It is used to output trend judgment results based on the global correlation enhancement feature map, epidermal tissue area, granulation tissue area, putrefactive tissue area, necrotic tissue area and exudate data of continuous time series. The storage module is used to store the model training parameters of the self-attention enhancement backbone network, feature segmentation module, scoring prediction module, and temporal prediction module.

Images