Wound type detection method based on image segmentation and storage medium

Abstract

The present application relates to the technical field of wound image intelligent identification, and discloses a wound type detection method based on image segmentation and a storage medium.The method comprises the following steps: synchronously acquiring a wound time sequence image and multivariate physiological data, generating a fusion feature atlas, and generating a preliminary region mask through a dynamic path selection network according to the fusion feature atlas.Next, a high attention region image is obtained by using the mask to cut, and the high attention region image is input into a cascade segmentation network after being enhanced.The network first outputs a refined wound contour by the first level, and then classifies and segments the internal organization type under the constraint of the contour by the second level.According to the segmentation result, tissue composition and spatial distribution features are extracted, and a predefined knowledge base is matched to output a type result.The method fuses multivariate data and a cascade segmentation architecture, improves the accuracy and robustness of wound segmentation under complex conditions, and thus realizes more accurate automatic judgment of the wound type.

Description

Technical Field

[0001] This invention relates to the field of intelligent wound image recognition technology, specifically to a wound type detection method and storage medium based on image segmentation. Background Technology

[0002] In current wound care and diagnostic practices, automated and accurate wound type identification primarily relies on computer vision analysis of wound appearance images. Existing solutions typically use a single wound image as input, training a deep neural network model to directly perform semantic segmentation or classification of the wound region. These methods treat wound contour localization and internal tissue type identification as a hybrid task for model learning, or process them in parallel using a multi-task learning framework. However, wound healing is a dynamic process, and its appearance is influenced by the local microenvironment; a single image cannot fully reflect this complex physiological state.

[0003] Single-modal image analysis ignores the physiological context of wound healing, resulting in insufficient robustness of the model when dealing with variations in image quality or complex wound presentations. Coupled contour segmentation and tissue classification in a single step, the model is prone to confusion in regions with blurred boundaries. Misclassification of internal tissues can erode contour accuracy, while contour deviations can further amplify internal classification errors. This mutual interference limits the segmentation results to the clinically required precision, making it difficult to consistently output reliable tissue composition and spatial distribution features for type interpretation. Summary of the Invention

[0004] The purpose of this invention is to provide a wound type detection method and storage medium based on image segmentation to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a wound type detection method based on image segmentation, the method comprising:

[0006] Acquire temporal image sequences of the target wound and synchronously recorded multivariate physiological monitoring data to generate a fused feature map;

[0007] Based on the fused feature map, a preliminary wound region mask is generated through a dynamic path selection network.

[0008] The images in the temporal image sequence are cropped using the preliminary wound region mask to obtain a high-resolution wound region of interest image.

[0009] The high-resolution wound region of interest image is subjected to texture enhancement and boundary sharpening processing, and the processed image is input into a cascaded segmentation network;

[0010] The first-level network of the cascaded segmentation network outputs a refined wound contour segmentation result, and the second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and performs tissue type classification segmentation on the internal region of the wound under its constraints.

[0011] Based on the results of the tissue type classification and segmentation, the tissue composition ratio and spatial distribution characteristics of the wound area are extracted.

[0012] Combining the tissue composition ratio features and spatial distribution features, the corresponding wound type labels are matched from the predefined wound type knowledge base, and the final wound type detection result is output.

[0013] Preferably, the step of acquiring the time-series image sequence of the target wound and simultaneously recording multivariate physiological monitoring data to generate a fused feature map includes:

[0014] Multi-scale feature extraction is performed on each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and the multivariate physiological monitoring data is encoded into a numerical feature vector.

[0015] The pixel-level multi-scale feature map and the numerical feature vector are fused across modalities in the feature space to generate a fused feature map.

[0016] The step of extracting multi-scale features from each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and encoding the multivariate physiological monitoring data into a numerical feature vector, includes:

[0017] Convolutional layers with different kernel sizes are used to process single-frame images in the temporal image sequence in parallel and extract image features under different receptive fields.

[0018] Image features from different receptive fields are concatenated along the channel dimension, and an attention weight generation layer is used to assign weights to the features of different channels to generate a weighted pixel-level multi-scale feature map.

[0019] Simultaneously, multi-dimensional physiological monitoring data, including temperature, humidity, and pH, are input into a fully connected encoder;

[0020] The fully connected encoder maps the multivariate physiological monitoring data to a high-dimensional space through multi-layer nonlinear transformation, and outputs a numerical feature vector of fixed dimension.

[0021] Preferably, the step of fusing the pixel-level multi-scale feature map and the numerical feature vector across modalities in the feature space to generate a fused feature map includes:

[0022] The numerical feature vector is expanded into a feature map with the same spatial size as the pixel-level multi-scale feature map through a spatial broadcasting operation;

[0023] Calculate the channel cross-correlation matrix between the pixel-level multi-scale feature map and the expanded numerical feature map;

[0024] Based on the channel cross-correlation matrix, an affine transformation parameter matrix for modulating image features is generated;

[0025] The affine transformation parameter matrix is ​​used to perform channel-by-channel feature transformation and fusion on the pixel-level multi-scale feature map to generate a fused feature map.

[0026] Preferably, based on the fused feature map, a preliminary wound region mask is generated through a dynamic path selection network, including:

[0027] The dynamic path selection network contains multiple parallel and structurally heterogeneous image segmentation sub-paths;

[0028] The fused feature map is input into a routing decision layer, and the routing decision layer outputs a routing probability distribution vector, which corresponds to each image segmentation sub-path in the dynamic path selection network.

[0029] Based on the routing probability distribution vector, the image segmentation sub-path with the highest probability is selected as the active path, while the image segmentation sub-path with the second highest probability is retained as the auxiliary path.

[0030] The fused feature map is simultaneously input into the activation path and the auxiliary path;

[0031] The activation path outputs a primary segmented activation graph, and the auxiliary path outputs an auxiliary segmented activation graph.

[0032] The primary segmentation activation map and the auxiliary segmentation activation map are weighted and summed, and a thresholding operation is performed to generate a preliminary binary wound region mask.

[0033] Preferably, the high-resolution wound region of interest image is subjected to texture enhancement and boundary sharpening processing, and the processed image is input into a cascaded segmentation network, including:

[0034] A family of local binary mode operators is used to perform multi-directional texture feature calculation on the high-resolution wound region of interest image to generate a texture intensity map.

[0035] The texture intensity map is superimposed on the original image to enhance the texture contrast of the wound tissue;

[0036] A gradient-based anisotropic diffusion filter is used to process the texture-enhanced image, which smooths image noise while sharpening the boundary between wounds and healthy skin.

[0037] The image, after texture enhancement and boundary sharpening, is input into the cascaded segmentation network.

[0038] Preferably, the first-level network of the cascaded segmentation network outputs a refined wound contour segmentation result, including:

[0039] The first-level network adopts an encoder-decoder structure, in which the encoder part progressively downsamples to extract deep semantic features, and the decoder part progressively upsamples and combines the skip connection features in the encoding process.

[0040] At the end of the decoder of the first-level network, a contour point sequence prediction head is used to predict the positional offset of each point on the wound contour in a dense manner.

[0041] The predicted positional offset is applied to an initial contour mesh, and a closed and continuous refined wound contour segmentation result is generated through iterative deformation.

[0042] Preferably, the second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and, under its constraints, performs tissue type classification segmentation of the internal region of the wound, including:

[0043] The refined wound contour segmentation result is used as a spatial mask to crop out the precise internal region image of the wound from the image after texture enhancement and boundary sharpening.

[0044] The image of the internal region of the wound is input into the second-level network, which consists of multiple parallel lightweight convolutional branches, each branch focusing on extracting features of a preset tissue type;

[0045] Each lightweight convolutional branch outputs a tissue type confidence map, where each pixel value represents the probability of belonging to the tissue type corresponding to the lightweight convolutional branch.

[0046] By using a cross-channel competition mechanism, the organization type label with the highest confidence level is assigned to each pixel to generate an organization type classification segmentation map.

[0047] Preferably, based on the results of the tissue type classification and segmentation, the tissue composition ratio characteristics and spatial distribution characteristics of the wound area are extracted, including:

[0048] The total number of pixels occupied by each tissue type label in the tissue type classification segmentation map is counted, and the ratio of the total number of pixels to the total number of pixels in the wound area is calculated to obtain the tissue composition ratio feature.

[0049] For each tissue type in the tissue type classification segmentation diagram, calculate the geometric moments of its connected regions, including the centroid position, principal axis direction, and spatial scatter matrix.

[0050] Based on the centroid location, principal axis direction, and spatial scattering matrix, a spatial distribution feature vector describing the relative position and orientation of different tissue types within the wound area is generated.

[0051] Preferably, combining the tissue composition ratio characteristics and spatial distribution characteristics, the corresponding wound type label is matched from a predefined wound type knowledge base, including:

[0052] The predefined wound type knowledge base stores feature templates for various standard wound types. Each feature template consists of a standard tissue composition ratio range and a standard spatial distribution pattern.

[0053] Calculate the degree of conformity between the tissue composition ratio feature and the standard tissue composition ratio range in each standard wound type feature template;

[0054] Calculate the structural similarity between the spatial distribution feature vector and the standard spatial distribution pattern in each standard wound type feature template;

[0055] The conformity and structural similarity are weighted and fused to generate a comprehensive matching score for each standard wound type label;

[0056] The standard wound type label with the highest comprehensive matching score is selected as the final wound type detection result.

[0057] Preferably, the present invention further includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wound type detection method based on image segmentation as described above.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] By simultaneously acquiring temporal image sequences and multivariate physiological monitoring data of the wound, and generating a fused feature map, the model used for initial localization can perceive physiological dynamic information beyond visual appearance. The dynamic path selection network adaptively adjusts the processing path based on this fused information, invoking more suitable feature extraction and inference logic when image quality is poor or the physiological state is special. This makes the generation of the initial wound region mask no longer solely dependent on image texture and color contrast, enhancing its resistance to interference such as changes in lighting and exudate reflection, thus improving the stability and accuracy of the initial localization stage in complex real-world scenarios.

[0060] A cascaded segmentation network is employed, with the refined wound contour segmentation results output by the first-level network serving as a hard constraint input to the second-level network. This structure forcibly decouples and serializes the two tasks of "overall region localization" and "internal component analysis." The first-level network can focus on learning precise boundary features, unaffected by the complexity of internal textures; the second-level network, within known and reliable contour boundaries, specializes in identifying different tissue types, preventing classification results from unconstrainedly spilling over into normal skin areas. This process ensures that tissue classification and segmentation have clear spatial limitations, making the extracted tissue composition proportions and spatial distribution features more realistic and reliable, thus improving the quality of the logical foundation for subsequent type matching. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the working principle of the wound type detection method based on image segmentation described in this invention.

[0062] Figure 2 A flowchart for cross-modal fusion;

[0063] Figure 3 A flowchart for texture enhancement and boundary sharpening processing;

[0064] Figure 4 The convergence analysis diagram is shown for the iterative deformation of the wound contour.

[0065] Figure 5 A heatmap showing the combined matching scores of wound samples and standard wound types. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see Figure 1This invention provides a wound type detection method based on image segmentation. The method includes: acquiring a temporal image sequence of a target wound and synchronously recorded multivariate physiological monitoring data to generate a fused feature map. Based on the fused feature map, a preliminary wound region mask is generated using a dynamic path selection network. The preliminary wound region mask is used to crop regions in the temporal image sequence to obtain a high-resolution wound region of interest image. Texture enhancement and boundary sharpening are performed on the high-resolution wound region of interest image, and the processed image is input into a cascaded segmentation network. The first-level network of the cascaded segmentation network outputs a refined wound contour segmentation result, and the second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and performs tissue type classification segmentation of the internal region of the wound under its constraints. Based on the tissue type classification segmentation result, the tissue composition ratio features and spatial distribution features of the wound region are extracted. Combining the tissue composition ratio features and spatial distribution features, corresponding wound type labels are matched from a predefined wound type knowledge base to output the final wound type detection result.

[0068] Example 1: See Figure 2 Multi-scale feature extraction is performed on each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and the multivariate physiological monitoring data is encoded into a numerical feature vector. Convolutional layers with different kernel sizes are used to process single frames of the time-series image sequence in parallel, extracting image features from different receptive fields. The image features from different receptive fields are concatenated along their channel dimensions, and an attention weight generation layer is used to assign weights to the features of different channels, generating a weighted pixel-level multi-scale feature map. Simultaneously, multivariate physiological monitoring data including temperature, humidity, and pH are input into a fully connected encoder. The fully connected encoder maps the multivariate physiological monitoring data to a high-dimensional space through multi-layer nonlinear transformations, outputting a fixed-dimensional numerical feature vector. The pixel-level multi-scale feature map and the numerical feature vector are fused across modally in the feature space to generate a fused feature map. The numerical feature vector is expanded into a feature map with the same spatial size as the pixel-level multi-scale feature map through a spatial broadcast operation. The channel cross-correlation matrix between the pixel-level multi-scale feature map and the expanded numerical feature map is calculated. Based on the channel cross-correlation matrix, an affine transformation parameter matrix for modulating image features is generated. The affine transformation parameter matrix is ​​then used to perform channel-by-channel feature transformation and fusion on the pixel-level multi-scale feature map to generate a fused feature map.

[0069] In the specific implementation, multi-scale feature extraction is performed on each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and the multivariate physiological monitoring data is encoded into a numerical feature vector. Convolutional layers with different kernel sizes are used in parallel to process single frames in the time-series image sequence, thereby extracting image features from different receptive fields. The image features from different receptive fields are concatenated along the channel dimension. The concatenated features are then passed through an attention weight generation layer, which dynamically assigns weights to the features of different channels, generating a weighted pixel-level multi-scale feature map. Simultaneously, multivariate physiological monitoring data, including temperature, humidity, and pH data, are input into a fully connected encoder. The fully connected encoder maps the multivariate physiological monitoring data to a high-dimensional space through multiple nonlinear transformations, outputting a fixed-dimensional numerical feature vector.

[0070] In some embodiments, pixel-level multi-scale feature maps and numerical feature vectors are fused across modalities in the feature space to generate a fused feature map. The numerical feature vectors are expanded into a feature map with the same height and width as the pixel-level multi-scale feature map through a spatial broadcast operation. The channel cross-correlation matrix between the pixel-level multi-scale feature map and the expanded numerical feature map is calculated. Based on the channel cross-correlation matrix, an affine transformation parameter matrix for modulating image features is generated. It can be understood that the generation of the affine transformation parameter matrix can be achieved through a nonlinear mapping. In one example, this mapping relationship can be expressed as:

[0071]

[0072] in: This represents the calculated channel cross-correlation matrix. This represents the compact feature vector obtained after global average pooling of the expanded numerical feature map. This represents the sigmoid activation function. and These represent the weights and biases in the generated affine transformation parameter matrix, respectively. The affine transformation parameter matrix is ​​used to perform channel-by-channel feature transformation and fusion on the pixel-level multi-scale feature map, generating the final fused feature map.

[0073] In practice, feature transformation and fusion operations are performed pixel-by-pixel. For each spatial location on the pixel-level multi-scale feature map, its corresponding feature vector undergoes an affine transformation with the corresponding weights and biases in the affine transformation parameter matrix. Optionally, this transformation can be represented as scaling and translation operations on the original feature vector. The transformed feature vector contains multi-scale visual information from the image modality and numerical physiological information from the physiological monitoring data modality. The spatial dimension of the fused feature map remains consistent with the original pixel-level multi-scale feature map, but the feature representation at each location is modulated by the multivariate physiological monitoring data.

[0074] Example 2: See Figure 3 The dynamic path selection network comprises multiple parallel and structurally heterogeneous image segmentation sub-paths. The fused feature map is input into a routing decision layer, which outputs a routing probability distribution vector corresponding to each image segmentation sub-path in the dynamic path selection network. Based on the routing probability distribution vector, the image segmentation sub-path with the highest probability is selected as the activation path, while the second-highest probability image segmentation sub-path is retained as an auxiliary path. The fused feature map is simultaneously input into the activation path and the auxiliary path. The activation path outputs a primary segmentation activation map, and the auxiliary path outputs an auxiliary segmentation activation map. The primary segmentation activation map and the auxiliary segmentation activation map are weighted and summed, and a thresholding operation is performed to generate a binarized preliminary wound region mask.

[0075] In practice, the process of generating a preliminary wound region mask based on the fused feature map is accomplished by a dynamic path selection network. This network comprises multiple parallel and heterogeneous image segmentation sub-paths, each with different network depths, convolutional kernel configurations, or attention mechanism designs. The fused feature map is input into a routing decision layer, which performs global feature analysis on the map and outputs a routing probability distribution vector. Each scalar element in this vector corresponds to the probability that a specific image segmentation sub-path in the dynamic path selection network is selected. In some embodiments, the routing decision layer consists of a global average pooling layer, a multilayer perceptron with two fully connected layers, and a softmax activation function.

[0076] The image segmentation sub-path with the highest probability is selected as the activation path based on the routing probability distribution vector, while the second-highest probability image segmentation sub-path is retained as an auxiliary path. The fused feature map is simultaneously input to both the selected activation path and the retained auxiliary path. The activation path performs a series of forward propagation calculations on the input fused feature map, outputting a primary segmentation activation map. Each pixel value in the primary segmentation activation map represents the confidence that the location belongs to the wound region. The auxiliary path processes the same fused feature map, outputting an auxiliary segmentation activation map, which provides supplementary information for wound region prediction. In a specific implementation, the primary segmentation activation map and the auxiliary segmentation activation map are weighted and summed. The weighted summation operation involves a learnable weighting coefficient. Optionally, the weighted summation calculation can be expressed as:

[0077]

[0078] in: This represents the main segmentation activation map generated by the activation path. This represents the auxiliary segmentation activation map generated by the auxiliary path. It is a learnable weighting coefficient between 0 and 1, whose initial value can be determined by the ratio of the highest probability value to the second highest probability value in the routing probability distribution vector. This represents the fused segmented activation map. The fused segmented activation map is processed by a preset fixed threshold or an adaptive thresholding operation to generate a preliminary binarized wound region mask. In some embodiments, the adaptive thresholding operation can automatically determine the optimal threshold for segmenting the foreground and background using Otsu's method. Otsu's method analyzes the grayscale value distribution of the fused segmented activation map, calculates the inter-class variance between the foreground and background classes corresponding to each possible threshold, and selects the threshold that maximizes the inter-class variance as the optimal segmentation threshold. This threshold is compared with each pixel value in the fused segmented activation map; regions with pixel values ​​greater than or equal to the threshold are classified as foreground (wound region), and regions with pixel values ​​less than the threshold are classified as background (non-wound region), thus generating a preliminary binarized wound region mask.

[0079] Example 3: A family of local binary mode operators is used to calculate multi-directional texture features on the high-resolution wound region of interest image to generate a texture intensity map. The texture intensity map is then superimposed on the original image to enhance the texture contrast of the wound tissue. A gradient-based anisotropic diffusion filter is used to process the texture-enhanced image, smoothing image noise while sharpening the boundary between the wound and healthy skin. The image after texture enhancement and boundary sharpening is input into the cascaded segmentation network. The first-stage network employs an encoder-decoder structure, where the encoder progressively downsamples to extract deep semantic features, and the decoder progressively upsamples and incorporates skip connection features from the encoding process. At the decoder end of the first-stage network, a contour point sequence prediction head is used to densely predict the positional offset of each point on the wound contour. The predicted positional offsets are applied to an initial contour grid, and closed and continuous refined wound contour segmentation results are generated through iterative deformation.

[0080] In practice, texture enhancement and boundary sharpening are performed on the high-resolution wound region of interest image. A family of local binary mode operators is used to calculate multi-directional texture features in the high-resolution wound region of interest image. This family of local binary mode operators includes a set of local binary mode operators with different principal direction parameters. These operators encode the local neighborhood of each pixel in the image using binary mode from multiple predefined directions, generating texture encoding maps in multiple directions. These texture encoding maps are then superimposed and normalized to generate a comprehensive texture intensity map. The texture intensity map is then superimposed on the original high-resolution wound region of interest image. The superposition method involves linearly combining the texture intensity map with the original image at pixel values ​​using preset weighting coefficients to enhance the texture contrast between the wound area and the healthy skin area, making the differences between areas with different texture characteristics, such as granulation tissue and necrotic tissue, more obvious. A gradient-based anisotropic diffusion filter is used to process the texture-enhanced image. The gradient-based anisotropic diffusion filter controls the diffusion coefficient through the local gradient information of the image. It performs strong smoothing in flat areas with small image gradients to suppress noise, and performs weak or even suppressed diffusion in edge areas with large image gradients to preserve and sharpen the boundaries.

[0081] In some embodiments, the first-stage network of the cascaded segmentation network employs an encoder-decoder structure to achieve refined wound contour segmentation. The encoder portion of the first-stage network consists of multiple alternating convolutional and downsampling layers, progressively downsampling to extract deep semantic features of the image. The decoder portion of the first-stage network consists of multiple alternating upsampling and convolutional layers, progressively upsampling to restore the spatial resolution of the feature map. During the upsampling process, the decoder fuses features from corresponding levels of the encoder through skip connections to combine shallow positional information with deep semantic information. At the end of the decoder of the first-stage network, a contour point sequence prediction head is used. The contour point sequence prediction head is a convolutional layer with two output channels, predicting the positional offsets of contour points along the row and column directions of the image, respectively. The contour point sequence prediction head operates in a dense prediction manner, outputting a two-dimensional offset vector for each spatial location on the feature map. This offset vector represents the displacement required to reach the true wound contour if the location were on the initial contour grid. It can be understood that the contour point sequence prediction head learns the vector field from the initial grid points to the true boundary.

[0082] In practice, the predicted positional offset is applied to an initial contour grid, which is a predefined sequence of simple polygonal points covering the central region of the image. A closed and continuous refined wound contour segmentation result is generated through iterative deformation; the deformation process is iterative. In one example, in each iteration, each point on the initial contour grid undergoes bilinear interpolation in the offset field output by the contour point sequence prediction head based on its current position to obtain the corresponding predicted offset vector, and its position is updated accordingly. Optionally, single-point updates can be described as follows:

[0083]

[0084] in: Indicates the first The contour points at the iteration number Two-dimensional coordinates of time, Indicates coordinates The predicted offset vector obtained by interpolation at that point. It is a scaling factor that controls the step size. This represents the updated coordinates. After several iterations, the point sequence of the initial contour mesh converges and closely fits the true boundary of the wound region. Connecting these ordered points forms a closed and continuous contour polygon, and the region inside this polygon is defined as the refined wound contour segmentation result. In some embodiments, the iteration process can be set to a fixed number of iterations, or it can automatically terminate when the average movement distance of all contour points is less than a threshold.

[0085] See Figure 4This is a composite analysis chart illustrating the iterative optimization process of the wound contour. It combines a line graph with the contour change trajectory, fully presenting the characteristics of the algorithm's iterative convergence. In the early stages of iteration (0-10 times), the average movement distance decreases rapidly, from approximately 1.3 pixels to below 1.0 pixel, indicating that the contour points are quickly converging towards the real wound boundary. After 10 iterations, the movement distance stabilizes and fluctuates slightly, indicating that the contour has basically converged and only minor adjustments are needed to fit the real boundary. The red line directly quantifies the average movement distance of the contour points, clearly demonstrating that the algorithm can converge within 10 iterations, proving the efficiency of the contour iteration algorithm and providing crucial performance evidence for its engineering implementation. The multi-colored contour trajectory visually presents the morphological evolution process from the initial contour to the final contour, verifying that the algorithm does not experience contour divergence or morphological distortion during iteration, ensuring the stability of the segmentation results.

[0086] Example 4: Using the refined wound contour segmentation result as a spatial mask, a precise image of the internal wound region is cropped from the image after texture enhancement and boundary sharpening. This image is input into the second-level network, which consists of multiple parallel lightweight convolutional branches, each focused on extracting features of a preset tissue type. Each lightweight convolutional branch outputs a tissue type confidence map, where each pixel value represents the probability of belonging to the corresponding tissue type. Through a cross-channel competition mechanism, the tissue type label with the highest confidence is assigned to each pixel, generating a tissue type classification segmentation map. The total number of pixels occupied by each tissue type label in the tissue type classification segmentation map is counted, and the ratio of this to the total number of pixels in the wound region is calculated to obtain the tissue composition ratio feature. For each tissue type in the tissue type classification segmentation map, the geometric moments of its connected regions are calculated, including the centroid position, principal axis direction, and spatial scatter matrix. Based on the centroid position, principal axis direction, and spatial scatter matrix, a spatial distribution feature vector describing the relative position and orientation of different tissue types within the wound region is generated.

[0087] In the specific implementation, the second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and performs tissue type classification segmentation of the internal region of the wound under the constraint of the refined wound contour segmentation result. The refined wound contour segmentation result is used as a spatial mask. The refined wound contour segmentation result is a binary image where pixels inside the wound contour have a value of 1, and pixels outside have a value of 0. This spatial mask is used to perform pixel-by-pixel multiplication on the image after texture enhancement and boundary sharpening, thereby cropping out an accurate image of the internal wound region. Only pixel information within the wound contour is retained in the image of the internal wound region, while the background region is set to zero. The image of the internal wound region is input into the second-level network, which consists of multiple parallel lightweight convolutional branches. The number of lightweight convolutional branches is the same as the number of preset tissue type categories. Each lightweight convolutional branch focuses on extracting features of a preset tissue type, which may include epidermis, dermis, subcutaneous tissue, granulation tissue, necrotic tissue, and eschar. Each lightweight convolutional branch outputs a tissue type confidence map, which has the same spatial dimensions as the input wound interior region image. The value of each pixel in the map represents the probability that the corresponding location in the input image belongs to the tissue type corresponding to that lightweight convolutional branch, with probability values ​​ranging from 0 to 1. A cross-channel competition mechanism is used to assign tissue type labels to each pixel. For each pixel location in the image, the probability values ​​of the tissue type confidence maps output by all lightweight convolutional branches at that location are compared, and the tissue type label with the highest probability value is assigned to that pixel, generating the final tissue type classification segmentation map. The tissue type classification segmentation map is a single-channel image, where the integer value of each pixel represents the assigned tissue type label number (see Table 1).

[0088] Table 1: Preset Organization Types and Feature Descriptions

[0089] Organization type label number Organization type name Typical visual feature description 1 Healthy epidermis The surface is smooth, the texture is uniform, and the color is close to normal skin tone. 2 Granulation tissue It has a bright red granular or downy surface. 3 Necrotic tissue It appears yellow, gray, or black, and its texture may be ointment-like or leather-like. 4 eschar Dry, hard brown or black covering

[0090] In some embodiments, tissue composition ratio features and spatial distribution features of the wound region are extracted based on the tissue type classification and segmentation results. The total number of pixels occupied by each tissue type label in the tissue type classification and segmentation image is counted, and the ratio of the total number of pixels for each tissue type to the total number of pixels in the wound region is calculated. The total number of pixels in the wound region is the sum of all non-background pixels in the tissue type classification and segmentation image, resulting in a set of numerical vectors as the tissue composition ratio features. It can be understood that the tissue composition ratio features quantitatively describe the percentage of area occupied by each type of tissue in the wound region. For each tissue type in the tissue type classification and segmentation image, the geometric moments of all its connected regions are calculated. The geometric moments are calculated based on a binary image, which consists of pixels in the tissue type classification and segmentation image equal to the label number of a specific tissue type. The calculated geometric moments include the centroid position, principal axis direction, and spatial scatter matrix. The centroid position describes the average center coordinates of the connected regions of that tissue type in the image plane. The principal axis direction describes the main extension direction of the connected regions of that tissue type. The spatial scatter matrix describes the two-dimensional distribution of pixels of that tissue type around its centroid. The calculation formula is:

[0091]

[0092] in: Represents the spatial scattering matrix. Representing the Types of organizations The representative belongs to the organizational type The total number of pixels, It is a two-dimensional column vector representing the organization type. The The coordinates of each pixel, in the form of: = ,in Represents the x-coordinate of that pixel. Represents the ordinate of that pixel. It is a two-dimensional column vector representing the organization type. The centroid coordinates, superscript This represents the matrix transpose operation. Based on the centroid location, principal axis direction, and spatial scatter matrix, a spatial distribution feature vector describing the relative position and orientation of different tissue types within the wound region is generated. In specific implementations, the spatial distribution feature vector can include various derived metrics. Optionally, derived metrics include the Euclidean distance between the centroids of different tissue types, the angle difference between the principal axis directions of different tissue types, and the ratio of eigenvalues ​​of the spatial scatter matrices of different tissue types.

[0093] Example 5: The predefined wound type knowledge base stores feature templates for various standard wound types. Each feature template consists of a standard tissue composition ratio range and a standard spatial distribution pattern. The conformity between the tissue composition ratio feature and the standard tissue composition ratio range in each standard wound type feature template is calculated. The structural similarity between the spatial distribution feature vector and the standard spatial distribution pattern in each standard wound type feature template is calculated. The conformity and structural similarity are weighted and fused to generate a comprehensive matching score for each standard wound type label. The standard wound type label with the highest comprehensive matching score is selected as the final wound type detection result.

[0094] In practice, the system combines tissue composition ratio characteristics and spatial distribution characteristics to match corresponding wound type labels from a predefined wound type knowledge base. This knowledge base stores feature templates for various standard wound types. Each standard wound type's feature template consists of a standard tissue composition ratio range and a standard spatial distribution pattern. The standard tissue composition ratio range defines the theoretically permissible range for the percentage area of ​​various tissues within that standard wound type. The standard spatial distribution pattern defines the typical relative positional relationships and geometric layout constraints between different tissue types within that standard wound type, such as whether a certain tissue is typically located in the center of the wound, or whether different tissues exhibit a layered distribution.

[0095] The calculation determines the degree of conformity between the tissue composition proportion feature and the standard tissue composition proportion range in each standard wound type feature template. The conformity calculation requires comparing the proportion value of each tissue type in the extracted tissue composition proportion feature of the current wound to see if it falls within the standard proportion range for that tissue type in the corresponding standard wound type feature template. In an example scenario, assuming a suspected pressure injury is being treated, the extracted tissue composition proportion feature shows "granulation tissue 65%, necrotic tissue 30%, eschar 5%", while the standard tissue composition proportion range for the "stage III pressure injury" template in the predefined wound type knowledge base is defined as "granulation tissue: 55%-75%, necrotic tissue: 20%-40%, eschar: 0%-10%". If the current feature value falls completely within this template range, the conformity is high; if it does not conform to the "stage II pressure injury" template's range of "epidermal loss, dermal exposure, no obvious necrotic tissue", the conformity is low.

[0096] In some embodiments, compliance This can be quantified as the normalized distance between the current feature vector and the center point of the template range, expressed by the formula:

[0097]

[0098] in: Indicates the first A standard wound type label, This indicates the total number of organization types. Indicates the current wound number The actual composition ratio of similar organizations Indicates the first The first standard wound type template The center value of the standard proportion range for similar organizations. Indicates the first The first standard wound type template The width of the standard proportion range for class organizations, That is, with the first The degree of conformity of the tissue composition ratio for a standard wound type is indicated by a value closer to 1, which represents a higher degree of conformity.

[0099] Calculate the structural similarity between the spatial distribution feature vector and the standard spatial distribution pattern in each standard wound type feature template. Structural similarity assesses the consistency between the spatial layout of various tissues within the current wound and the typical layout defined by the template. Continuing with the comparison using the pressure injury example scenario above, if the current wound spatial distribution features show granulation tissue widely distributed at the bottom of the wound bed and necrotic tissue scattered on the surface, while the standard spatial distribution pattern of the "stage III pressure injury" template is described as "granulation tissue mainly located at the base of the wound bed, with necrotic tissue attached to the surface or mixed within it," then the structural similarity is high. If there is a significant difference from the pattern of the "stage IV pressure injury" template, which states "bones and tendons may be exposed, and necrotic tissue may be located in deeper layers," then the structural similarity is low.

[0100] The conformity and structural similarity are weighted and fused to generate a comprehensive matching score for each standard wound type label. The weighted fusion assigns preset weight coefficients to conformity and structural similarity. It is understood that the weight coefficients can be set based on prior knowledge; for example, if the proportion of tissue composition is considered more discriminative in distinguishing certain wound types, then it is given a higher weight. Optionally, the weighted fusion uses a linear weighted summation method. The comprehensive matching score reflects the overall degree of matching between the current wound feature and a standard wound type template in the knowledge base. The standard wound type label with the highest comprehensive matching score is selected as the final wound type detection result. In some embodiments, if the highest comprehensive matching score is lower than a preset confidence threshold, "unknown type" is output or manual review is suggested. It is understood that the predefined wound type knowledge base can be expanded and updated based on medical consensus and new clinical evidence to cover a wider range of wound types.

[0101] 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.

[0102] See Figure 5 This is a heatmap showing the comprehensive matching scores between wound samples and standard wound types. It displays the degree of matching between different wound samples and various standard wound types, and is one of the core outputs of the wound type detection process. Matching distribution: High-matching samples are concentrated in different areas for different wound types (e.g., high-matching samples for stage I pressure injuries are concentrated in numbers 0-10, while those for stage II pressure injuries are concentrated in numbers 15-25), reflecting the correspondence between samples and standard types. High-score characteristics: The yellow area (matching score ≥ 0.7) appears only at the intersection of a few samples and their corresponding types, indicating that the tissue characteristics and spatial distribution of these samples are highly consistent with the corresponding standard wound type. Low-score characteristics: The dark purple area (matching score ≤ 0.2) accounts for a high proportion, indicating that most samples have a low matching degree with non-corresponding types, and the detection results have strong discriminative power. This heatmap can quickly locate the optimal matching wound type for each sample, while intuitively displaying the confidence level of the matching results, providing data support for clinical wound classification and treatment plan selection.

[0103] 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, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wound type detection method based on image segmentation, characterized in that, The method includes: Acquire temporal image sequences of the target wound and synchronously recorded multivariate physiological monitoring data to generate a fused feature map; Based on the fused feature map, a preliminary wound region mask is generated through a dynamic path selection network, including: The dynamic path selection network contains multiple parallel and structurally heterogeneous image segmentation sub-paths; The fused feature map is input into a routing decision layer, and the routing decision layer outputs a routing probability distribution vector, which corresponds to each image segmentation sub-path in the dynamic path selection network. Based on the routing probability distribution vector, the image segmentation sub-path with the highest probability is selected as the active path, while the image segmentation sub-path with the second highest probability is retained as the auxiliary path. The fused feature map is simultaneously input into the activation path and the auxiliary path; The activation path outputs a primary segmentation activation map, where each pixel value represents the confidence level that the location belongs to the wound region; the auxiliary path outputs an auxiliary segmentation activation map, which provides supplementary information for wound region prediction. The main segmentation activation map and the auxiliary segmentation activation map are weighted and summed, and a thresholding operation is performed to generate a preliminary binary wound region mask. The images in the temporal image sequence are cropped using the preliminary wound region mask to obtain a high-resolution wound region of interest image. The high-resolution wound region of interest image is subjected to texture enhancement and boundary sharpening processing, and the processed image is input into a cascaded segmentation network; The first-level network of the cascaded segmentation network outputs a refined wound contour segmentation result, and the second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and performs tissue type classification segmentation on the internal region of the wound under its constraints. Based on the results of the tissue type classification and segmentation, the tissue composition ratio and spatial distribution characteristics of the wound area are extracted. Combining the tissue composition ratio features and spatial distribution features, the corresponding wound type labels are matched from the predefined wound type knowledge base, and the final wound type detection result is output. The second-level network of the cascaded segmentation network receives the refined wound contour segmentation result and, under its constraints, performs tissue type classification segmentation of the internal region of the wound, including: The refined wound contour segmentation result is used as a spatial mask to crop out the precise internal region image of the wound from the image after texture enhancement and boundary sharpening. The image of the internal region of the wound is input into the second-level network, which consists of multiple parallel lightweight convolutional branches, each branch focusing on extracting features of a preset tissue type. Each lightweight convolutional branch outputs a tissue type confidence map, where each pixel value represents the probability of belonging to the tissue type corresponding to the lightweight convolutional branch. By using a cross-channel competition mechanism, the organization type label with the highest confidence level is assigned to each pixel to generate an organization type classification segmentation map.

2. The wound type detection method based on image segmentation according to claim 1, characterized in that, The process of acquiring a time-series image sequence of the target wound and simultaneously recording multivariate physiological monitoring data to generate a fused feature map includes: Multi-scale feature extraction is performed on each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and the multivariate physiological monitoring data is encoded into a numerical feature vector. The pixel-level multi-scale feature map and the numerical feature vector are fused across modalities in the feature space to generate a fused feature map. The step of extracting multi-scale features from each frame of the time-series image sequence to generate a pixel-level multi-scale feature map, and encoding the multivariate physiological monitoring data into a numerical feature vector, includes: Convolutional layers with different kernel sizes are used to process single-frame images in the temporal image sequence in parallel and extract image features under different receptive fields. Image features from different receptive fields are concatenated along the channel dimension, and an attention weight generation layer is used to assign weights to the features of different channels to generate a weighted pixel-level multi-scale feature map. Simultaneously, multi-dimensional physiological monitoring data, including temperature, humidity, and pH, are input into a fully connected encoder; The fully connected encoder maps the multivariate physiological monitoring data to a high-dimensional space through multi-layer nonlinear transformation, and outputs a numerical feature vector of fixed dimension.

3. The wound type detection method based on image segmentation according to claim 2, characterized in that, The step of fusing the pixel-level multi-scale feature map and the numerical feature vector across modalities in the feature space to generate a fused feature map includes: The numerical feature vector is expanded into a feature map with the same spatial size as the pixel-level multi-scale feature map through a spatial broadcasting operation. Calculate the channel cross-correlation matrix between the pixel-level multi-scale feature map and the expanded numerical feature map; Based on the channel cross-correlation matrix, an affine transformation parameter matrix for modulating image features is generated; The affine transformation parameter matrix is ​​used to perform channel-by-channel feature transformation and fusion on the pixel-level multi-scale feature map to generate a fused feature map.

4. The wound type detection method based on image segmentation according to claim 1, characterized in that, The high-resolution wound region of interest image is subjected to texture enhancement and boundary sharpening processing, and the processed image is input into a cascaded segmentation network, including: A family of local binary mode operators is used to perform multi-directional texture feature calculation on the high-resolution wound region of interest image to generate a texture intensity map. The texture intensity map is superimposed on the original image to enhance the texture contrast of the wound tissue; A gradient-based anisotropic diffusion filter is used to process the texture-enhanced image, which smooths the image noise while sharpening the boundary between the wound and healthy skin. The image, after texture enhancement and boundary sharpening, is input into the cascaded segmentation network.

5. The wound type detection method based on image segmentation according to claim 1, characterized in that, The first-level network of the cascaded segmentation network outputs refined wound contour segmentation results, including: The first-level network adopts an encoder-decoder structure, in which the encoder part progressively downsamples to extract deep semantic features, and the decoder part progressively upsamples and combines the skip connection features in the encoding process. At the end of the decoder of the first-level network, a contour point sequence prediction head is used to predict the positional offset of each point on the wound contour in a dense manner. The predicted positional offset is applied to an initial contour mesh, and a closed and continuous refined wound contour segmentation result is generated through iterative deformation.

6. The wound type detection method based on image segmentation according to claim 5, characterized in that, Based on the results of the tissue type classification and segmentation, the tissue composition ratio and spatial distribution characteristics of the wound area are extracted, including: The total number of pixels occupied by each tissue type label in the tissue type classification segmentation map is counted, and the ratio of the total number of pixels to the total number of pixels in the wound area is calculated to obtain the tissue composition ratio feature. For each tissue type in the tissue type classification segmentation diagram, calculate the geometric moments of its connected regions, including the centroid position, principal axis direction, and spatial scatter matrix. Based on the centroid location, principal axis direction, and spatial scattering matrix, a spatial distribution feature vector describing the relative position and orientation of different tissue types within the wound area is generated.

7. The wound type detection method based on image segmentation according to claim 6, characterized in that, Combining the aforementioned tissue composition ratio characteristics and spatial distribution characteristics, corresponding wound type labels are matched from a predefined wound type knowledge base, including: The predefined wound type knowledge base stores feature templates for various standard wound types. Each feature template consists of a standard tissue composition ratio range and a standard spatial distribution pattern. Calculate the degree of conformity between the tissue composition ratio feature and the standard tissue composition ratio range in each standard wound type feature template; Calculate the structural similarity between the spatial distribution feature vector and the standard spatial distribution pattern in each standard wound type feature template; The conformity and structural similarity are weighted and fused to generate a comprehensive matching score for each standard wound type label; The standard wound type label with the highest comprehensive matching score is selected as the final wound type detection result.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wound type detection method based on image segmentation as described in any one of claims 1 to 7.

Images