Skin tumor classification system based on multi-scale decoupled representation and adaptive learning

Abstract

The application provides a skin tumor classification system based on multi-scale decoupling representation and adaptive learning, and relates to the technical field of skin tumor classification.The system comprises a multi-scale feature module, a self-encoding decoupling module, a training sample module, a feature purity module, a severity module, a deterioration risk module and a tumor classification module.The multi-scale feature module is used for acquiring a multi-scale feature map.The self-encoding decoupling module is used for determining pathological features and interference features.The training sample module is used for acquiring training sample features.The feature purity module is used for determining a pathological feature purity index.The severity module is used for determining a tumor severity coefficient according to the multi-scale feature map and the pathological features.The deterioration risk module is used for determining a tumor deterioration risk coefficient according to the pathological features and the training sample features.The tumor classification module is used for determining a skin tumor classification according to the pathological feature purity index, the tumor severity coefficient and the tumor deterioration risk coefficient.The application can improve the accuracy of the skin tumor classification result based on multi-scale decoupling representation and adaptive learning.

Description

Technical Field

[0001] This invention relates to the field of skin tumor classification technology, and in particular to a skin tumor classification system based on multi-scale decoupled representation and adaptive learning. Background Technology

[0002] In related technologies, existing technologies typically use a single scale or a simple fusion of multi-scale features for skin tumor classification. These technologies fail to fully model and utilize the complementarity and hierarchical relationships between features of different scales, and cannot comprehensively capture the complex characteristics of lesions. They only output a single benign or malignant classification result, lacking quantitative assessment of tumor severity and dynamic prediction of future deterioration risk. In other words, these technologies are difficult to support the in-depth application of skin tumor classification systems in clinical decision-making and are difficult to improve the accuracy of skin tumor classification results based on multi-scale decoupled representation and adaptive learning.

[0003] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a skin tumor classification system based on multi-scale decoupled representation and adaptive learning, which can solve the technical problems that related technologies are difficult to support the in-depth application of skin tumor classification systems in clinical decision-making and difficult to improve the accuracy of skin tumor classification results based on multi-scale decoupled representation and adaptive learning.

[0005] According to a first aspect of the present invention, a skin tumor classification system based on multi-scale decoupled representation and adaptive learning is provided, comprising:

[0006] The multi-scale feature module is used to construct a multi-scale deep feature extraction network to obtain multi-scale feature maps;

[0007] An autoencoder decoupling module is used to construct a variational autoencoder decoupling module, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features;

[0008] The training sample module is used to construct an adaptive learning training set and obtain training sample features;

[0009] The feature purity module is used to determine the pathological feature purity index based on the multi-scale feature map, the pathological feature, and the interference feature;

[0010] The severity module is used to determine the tumor severity coefficient based on the multi-scale feature map and the pathological features;

[0011] The deterioration risk module is used to determine the tumor deterioration risk coefficient based on the pathological features and the training sample features;

[0012] The tumor classification module is used to determine the classification of skin tumors based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient.

[0013] According to the present invention, determining the pathological feature purity index based on the multi-scale feature map, the pathological feature, and the interference feature includes:

[0014] Based on the pathological features and the interference features, the pathological factors and interference factors are determined;

[0015] Based on the multi-scale feature map, the pathological first scale, the pathological second scale, the interference first scale, and the interference second scale are determined.

[0016] The pathological feature purity index is determined based on the first pathological scale, the second pathological scale, the first interference scale, the second interference scale, the pathological factor, and the interference factor.

[0017] According to the present invention, determining a tumor severity coefficient based on the multi-scale feature map and the pathological features includes:

[0018] Based on the multi-scale feature map, determine the number of pathological feature scales and the significance of the pathological feature in the pathological scale.

[0019] Based on the pathological features, a pathological feature vector is determined, wherein the pathological feature vector includes: a single-scale pathological feature vector, a pathological feature edge region vector, and a pathological feature center region vector.

[0020] The tumor severity coefficient is determined based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector.

[0021] According to the present invention, determining a tumor severity coefficient based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector includes: determining the tumor severity coefficient according to the formula... ,

[0022] Determining the tumor severity coefficient Where d is the number of pathological feature scales. The significance of the pathological feature at the i-th pathological scale. Let be the single-scale pathological feature vector of the i-th pathological scale. The vector represents the edge region of pathological features. The vector represents the central region of the pathological features. , and For preset weights, and , Let i be a very small positive value, i ≤ d, and both i and d are positive integers.

[0023] According to the present invention, determining the tumor malignancy risk coefficient based on the pathological features and the training sample features includes:

[0024] Based on the pathological characteristics, determine the pathological feature vector, local sample feature vector variance, and global sample feature vector variance of the input sample;

[0025] Based on the characteristics of the training samples, determine the pathological feature vectors of benign samples and malignant samples;

[0026] The tumor malignancy risk coefficient is determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample.

[0027] According to the present invention, determining the tumor malignancy risk coefficient based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample includes: determining the tumor malignancy risk coefficient according to the formula ,

[0028] Determine the risk factor for tumor progression ,in, The pathological feature vector of the input sample. This represents the pathological feature vector of a benign sample. This is the pathological feature vector of malignant samples. The variance of the local sample feature vector. The variance of the global sample feature vector. For a very small positive value, , and For preset weights, and .

[0029] According to the present invention, the classification of skin tumors is determined based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient, including:

[0030] The pathological classification of skin tumors is determined based on the purity index of the pathological features.

[0031] Based on the tumor severity coefficient and the tumor malignancy risk coefficient, a threshold for the tumor severity coefficient and a threshold for the tumor malignancy risk coefficient are determined.

[0032] Based on the tumor severity coefficient threshold and the tumor malignancy risk coefficient threshold, a tumor benign or malignant risk assessment matrix is ​​constructed.

[0033] The skin tumor classification is determined based on the aforementioned pathological classification of skin tumors and the aforementioned benign / malignant tumor risk assessment matrix.

[0034] According to a second aspect of the present invention, a skin tumor classification method based on multi-scale decoupled representation and adaptive learning is provided, comprising:

[0035] Construct a multi-scale deep feature extraction network to obtain multi-scale feature maps;

[0036] A variational autoencoder decoupling module is constructed, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features;

[0037] Construct an adaptive learning training set and obtain the features of the training samples;

[0038] The pathological feature purity index is determined based on the multi-scale feature map, the pathological features, and the interference features.

[0039] The tumor severity coefficient is determined based on the multi-scale feature map and the pathological features.

[0040] The tumor malignancy risk coefficient is determined based on the pathological features and the training sample features.

[0041] The skin tumor classification is determined based on the purity index of the pathological features, the tumor severity coefficient, and the tumor malignancy risk coefficient.

[0042] Technical Effects: According to the present invention, multi-scale feature maps can be obtained based on the constructed multi-scale deep feature extraction network. Pathological features and interference features are determined based on the constructed variational autoencoder decoupling module. Training sample features are obtained based on the constructed adaptive learning training set. Then, based on the multi-scale feature maps, pathological features, and interference features, a pathological feature purity index is determined. Based on the multi-scale feature maps and pathological features, a tumor severity coefficient is determined. Based on the pathological features and training sample features, a tumor malignancy risk coefficient is determined. Furthermore, based on the pathological feature purity index, tumor severity coefficient, and tumor malignancy risk coefficient, a skin tumor classification is determined, improving the accuracy and comprehensiveness of skin tumor classification. When determining the tumor severity coefficient, it can be determined based on the number of pathological feature scales, the saliency of the pathological scales, and the pathological feature vector. During the calculation process, the contribution of the information entropy term, the feature distribution discrete term, and the edge-center feature contrast term to the tumor severity coefficient can be accurately evaluated, improving the accuracy of the tumor severity coefficient. When determining the risk coefficient of tumor malignancy, the risk coefficient can be determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample. During the calculation process, the contribution of the relative distance term, the change direction term, and the local stability term to the risk coefficient of tumor malignancy can be accurately assessed, thereby improving the accuracy of the risk coefficient of tumor malignancy.

[0043] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0045] Figure 1 An exemplary block diagram of a skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to an embodiment of the present invention is shown;

[0046] Figure 2 An exemplary flowchart of a skin tumor classification method based on multi-scale decoupled representation and adaptive learning according to an embodiment of the present invention is shown. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0048] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0049] Figure 1 An exemplary block diagram of a skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to an embodiment of the present invention is shown, the system comprising:

[0050] The multi-scale feature module is used to construct a multi-scale deep feature extraction network to obtain multi-scale feature maps;

[0051] An autoencoder decoupling module is used to construct a variational autoencoder decoupling module, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features;

[0052] The training sample module is used to construct an adaptive learning training set and obtain training sample features;

[0053] The feature purity module is used to determine the pathological feature purity index based on the multi-scale feature map, the pathological feature, and the interference feature;

[0054] The severity module is used to determine the tumor severity coefficient based on the multi-scale feature map and the pathological features;

[0055] The deterioration risk module is used to determine the tumor deterioration risk coefficient based on the pathological features and the training sample features;

[0056] The tumor classification module is used to determine the classification of skin tumors based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient.

[0057] The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to embodiments of the present invention can obtain multi-scale feature maps based on a constructed multi-scale deep feature extraction network, determine pathological features and interference features based on a constructed variational autoencoder decoupling module, obtain training sample features based on a constructed adaptive learning training set, determine a pathological feature purity index based on the multi-scale feature maps, pathological features, and interference features, determine a tumor severity coefficient based on the multi-scale feature maps and pathological features, determine a tumor malignancy risk coefficient based on the pathological features and training sample features, and further determine the skin tumor classification based on the pathological feature purity index, tumor severity coefficient, and tumor malignancy risk coefficient, thereby improving the accuracy and comprehensiveness of skin tumor classification.

[0058] According to one embodiment of the present invention, a multi-scale deep feature extraction network is constructed in the multi-scale feature module to obtain multi-scale feature maps.

[0059] For example, the Swin-Transformer is a deep learning model designed specifically for visual tasks. It can be used to build multi-scale deep feature extraction networks. Compared to traditional convolutional neural networks, it can construct hierarchical feature maps, naturally supports multiple scales, and achieves efficient computation through a shift window mechanism. It also captures both local and global dependencies. The multi-scale deep feature extraction network divides the input raw dermoscopy image into multiple non-overlapping patches (e.g., 4×4 pixels), extracts features layer by layer through multiple shift windows, and each layer outputs a feature map of different resolution, forming a pyramid structure (e.g., S...). Stage 1: 128×128 pixels, Stage 2: 64×64 pixels, Stage 3: 32×32 pixels, Stage 4: 16×16 pixels. The deeper the stage (e.g., Stage 1), the lower the resolution and the larger the receptive field, used to capture semantics, overall morphology, and pathological patterns (e.g., irregular boundaries, asymmetrical structures). The shallower the stage (e.g., Stage 4), the higher the resolution and the smaller the receptive field, used to capture edges, texture, and local structures (e.g., pigment spots, vascular morphology). The output feature maps of each stage are stitched together to form a multi-scale feature map.

[0060] According to one embodiment of the present invention, in the autoencoder decoupling module, a variational autoencoder decoupling module is constructed, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features.

[0061] For example, the variational autoencoder decoupling module is composed of a VAE variational autoencoder. A VAE variational autoencoder is a generative model that can map data to a latent space and learn the data distribution through the encoder structure. Pathological features refer to low-dimensional or medium-dimensional descriptive labels with clear pathological meaning obtained by filtering and aggregating them according to the pathological scale from the multi-scale feature map. The remaining descriptive labels are interference features. By inputting the multi-scale feature map into the encoder, multiple latent space vectors are decoupled. The latent space vectors are input into the decoder to filter the latent space vectors. The scales of the latent space vectors that are related to pathology (e.g., cell atypia, structural disorder) are segmented into pathological features, and the scales of other interferences (e.g., hair, light, skin color) of the latent space vectors are segmented into interference features. The reconstructed pathological scale feature map and interference scale feature map are output, that is, pathological features and interference features.

[0062] According to one embodiment of the present invention, in the training sample module, an adaptive learning training set is constructed to obtain training sample features.

[0063] For example, the adaptive learning training set refers to the set of labeled dermoscopic image data and related information used to train the system's learning module parameters during the system model development phase. By inputting a large number of standardized dermoscopic images into the adaptive learning training set, each image has a reliable diagnostic pathological label determined by pathological biopsy results or consensus of multiple senior dermatologists. The label content includes information describing specific pathological morphologies (such as uneven pigmentation and blurred boundaries) and information about interfering attributes such as hair, lighting, skin color, and bubbles, based on image segmentation or feature clustering, thus obtaining training sample features.

[0064] According to one embodiment of the present invention, in the feature purity module, a pathological feature purity index is determined based on the multi-scale feature map, the pathological feature, and the interference feature.

[0065] According to an embodiment of the present invention, determining a pathological feature purity index based on the multi-scale feature map, the pathological feature, and the interference feature includes:

[0066] Based on the pathological features and the interference features, the pathological factors and interference factors are determined;

[0067] Based on the multi-scale feature map, the pathological first scale, the pathological second scale, the interference first scale, and the interference second scale are determined.

[0068] The pathological feature purity index is determined based on the first pathological scale, the second pathological scale, the first interference scale, the second interference scale, the pathological factor, and the interference factor.

[0069] For example, pathological factors refer to pathological structure labels annotated by doctors, while interference factors refer to automatically extracted interference labels such as light or hair. Both are obtained through actual diagnosis. Pathological factors and interference factors are real factors, representing the real pathological and interference information of the original dermoscopy image. The first pathological scale refers to the set of features with the largest mutual information with pathological factors among all candidate scales (candidate scales refer to a predefined and extracted set of feature levels or data versions that may contain pathological information). Pathological factor mutual information is used to measure how much effective information related to disease diagnosis is contained in the current candidate scale; pathological factor mutual information can be extracted through image features. This is the principal feature scale that contributes the most to diagnosis / classification. The second pathological scale refers to the set of features with the largest mutual information with pathological factors among all remaining candidate scales. The feature set with the second largest mutual information is the auxiliary feature scale that has a supplementary role in diagnosis but contributes slightly less. The first interference scale refers to the feature scale with the smallest mutual information with pathological factors but the largest mutual information with background / technical factors (background / technical factor mutual information is used to measure how much technical noise or non-pathological information related to experimental operation, scanning equipment, environment, etc., is contained in the current candidate scale, and can also be extracted through image features). It is the level that carries the most non-pathological information. The second interference scale refers to the feature scale with the second smallest mutual information with pathological factors (or the second largest mutual information with background / technical factors) among all remaining candidate scales. Based on the first pathological scale, the second pathological scale, the first interference scale, the second interference scale, the pathological factors, and the interference factors, it can be specifically defined by the formula: Determine the purity index of pathological features ,in, For mutual information functions, As the primary standard of pathology, The second standard of pathology Pathological factors To interfere with the first scale, To interfere with the second scale, As an interference factor, This refers to the mutual information between the first-scale pathological marker and pathological factors, measuring the degree of information sharing between the first-scale pathological features and the true pathological information; that is, how much pathological information is contained in the first-scale pathological features. This refers to the mutual information between the second-scale pathological features and pathological factors, measuring the degree of information sharing between the second-scale pathological features and the true pathological information; that is, how much pathological information is contained in the second-scale pathological features. The item is a pathology-related item used to determine the degree to which the first pathology scale shares more true pathological information compared to the second pathology scale. The larger the difference, the more pathological information the first pathological scale contains compared to the second pathological scale, and the more representative the pathological features are. To measure the mutual information between the first-scale interference and the interference factor, the degree of information sharing between the first-scale interference features and the true interference information (i.e., non-pathological information) is assessed; that is, how much interference information is contained in the first-scale interference features. To measure the mutual information between the second-scale interference and the interference factor, we need to assess the degree of information sharing between the second-scale interference features and the real interference information; that is, how much interference information is contained in the second-scale interference features. The term is an interference-related term, used to determine how much richer the sharing of real interference information is between the first-scale interference and the second-scale interference. The larger the difference, the more interference information the first scale contains compared to the second scale, and the more representative the interference features are. The larger the pathological related items and interference related items are, the purer the pathological features. The higher the value, the richer the pathological information is in the first pathological scale compared to other scales in the multi-scale feature map. The clearer the pathological feature structure in the reconstructed pathological scale feature map, the cleaner the interference information is removed, making the whole system more scientific and credible in medical evaluation.

[0070] According to one embodiment of the present invention, in the severity module, a tumor severity coefficient is determined based on the multi-scale feature map and the pathological features.

[0071] According to one embodiment of the present invention, determining a tumor severity coefficient based on the multi-scale feature map and the pathological features includes:

[0072] Based on the multi-scale feature map, determine the number of pathological feature scales and the significance of the pathological feature in the pathological scale.

[0073] Based on the pathological features, a pathological feature vector is determined, wherein the pathological feature vector includes: a single-scale pathological feature vector, a pathological feature edge region vector, and a pathological feature center region vector.

[0074] The tumor severity coefficient is determined based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector.

[0075] For example, the number of pathological feature scales refers to the number of levels of a tumor at different pathological grades or morphological features. It is the number of scales that can reflect the actual pathological significance of the tumor. For example, it can be divided into morphological scale, molecular marker scale, and tissue structure scale, in which case the number of pathological feature scales is 3. The pathological scale significance of a pathological feature within a pathological scale refers to the prominence of the pathological feature at that pathological scale. It can be based on an adaptive learning training set, statistically analyzing the correlation between each pathological feature and the degree of tumor malignancy (e.g., Pearson correlation coefficient, chi-square test), and standardizing the correlation as the significance. For example, the quantitative value of the degree of tissue structure disorder under the tissue structure change scale. The range of pathological scale significance values ​​for pathological features within a pathological scale. Furthermore, the sum of the pathological scale significance of pathological features across all pathological scales must be 1. A single-scale pathological feature vector refers to a vector formed by splicing or aggregating pathological features at a certain scale in a multi-scale feature map. This is obtained by globally averaging and pooling the pathological features to compress them into a one-dimensional vector, such as boundary ambiguity, pigment heterogeneity, and asymmetry index. The pathological feature edge region vector refers to the vectorization of pathological features in the tumor edge region (e.g., cell infiltration at the edge, tissue structure destruction, etc.), with the vectorization method consistent with the single-scale pathological feature vector. The pathological feature center region vector refers to the vectorization of pathological features in the tumor center region (e.g., central necrosis, densely packed cell areas, etc.), with the vectorization method consistent with the single-scale pathological feature vector. Based on the number of pathological scales, the pathological scale significance of pathological features within the pathological scales, and the pathological feature vector, the complex characteristics exhibited by the tumor in pathological features are comprehensively evaluated to determine the tumor severity coefficient.

[0076] According to an embodiment of the present invention, determining a tumor severity coefficient based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector includes: determining the tumor severity coefficient according to formula (1). , (1),

[0077] Where d represents the number of pathological feature scales. The significance of the pathological feature at the i-th pathological scale. Let be the single-scale pathological feature vector of the i-th pathological scale. The vector represents the edge region of pathological features. The vector represents the central region of the pathological features. , and For preset weights, and , Let i be a very small positive value, i ≤ d, and both i and d are positive integers.

[0078] According to one embodiment of the present invention, Here, d represents the information entropy term, and d represents the number of pathological feature scales. The significance of the pathological feature at the i-th pathological scale. The term represents the information entropy of pathological features across all d pathological feature scales. The term represents the maximum possible entropy of all d pathological feature scales. Item Use Item pair Normalization is performed; the negative sign is used to cancel out the difference. The negative value of the term ensures that the entropy value is positive. To avoid the denominator being 0, which would render the formula meaningless, the value should be a very small positive value. For example, the greater the significance of a pathological feature on a certain or a few pathological scales, the closer it is to 1. The closer a term is to 0, the smaller the significance of the pathological feature across other pathological scales; the closer it is to 0, the less significant the pathological feature is. The closer the term is to negative infinity, the smaller the resulting information entropy term. Pathological features are highly concentrated on a few scales (e.g., a pigmented nevus, mainly showing pigmentation). Benign tumors usually have regular and uniform pathological feature scales. The smaller the information entropy term, the less severe the tumor. If the pathological features are evenly distributed across most pathological scales, and the pathological features are uniformly distributed across multiple scales (e.g., simultaneously showing irregular boundaries, cell deformation, and atypical hyperplasia), it indicates that the internal pathological structure of the current lesion is more chaotic and atypical, a strong indicator of malignancy. The closer the term is to the maximum possible entropy The larger the information entropy term, the greater the severity of the tumor.

[0079] According to one embodiment of the present invention, The term is a discrete term with a characteristic distribution. Let be the single-scale pathological feature vector of the i-th pathological scale. This is the mean vector of pathological feature vectors across all scales. The degree of dispersion of a single-scale pathological feature vector at each scale relative to the mean vector of pathological feature vectors at all scales indicates the degree of difference between the single-scale pathological feature vectors at each pathological scale. The greater the difference, the higher the degree of dispersion. The larger the value, the more contradictory the pathological features of the tumor are at multiple scales, indicating greater tumor severity. This aligns with the understanding that malignant tumors exhibit significant differences in pathological manifestations at different scales due to heterogeneity. Conversely, smaller differences indicate lower dispersion. The smaller the value, the more consistent the pathological manifestations of the tumor are across different scales, indicating a lower tumor severity and a more consistent pathological pattern across all scales characteristic of benign tumors.

[0080] According to one embodiment of the present invention, For edge-center feature contrast term, The vector represents the edge region of pathological features. The vector represents the central region of the pathological features. The term represents the Euclidean distance between the vectors of the edge regions and the central regions of the pathological features, indicating the magnitude of the difference between the pathological feature vectors of the edge and central regions. The term represents the magnitude of the vectors representing the edge regions and the center regions of the pathological features. It should be a very small positive value to avoid the denominator being zero, which would render the formula meaningless. Item Use Item pair The terms are normalized to represent the degree of difference in pathological features between the tumor periphery and the central region. The larger the value, the greater the difference in pathological features between the tumor margin and the central region, and the greater the severity of the tumor. This aligns with the characteristics of malignant tumors, which include invasive growth (invasion of the periphery into the central tissue) and strong heterogeneity (more active and atypic cells at the periphery, while the central region may exhibit distinct features due to necrosis). The smaller the value, the less difference there is between the pathological features of the tumor margin and the central region, and the less severe the tumor is. This is consistent with the characteristics of benign tumors, which usually have clear boundaries, grow slowly, and have relatively consistent pathological features (such as cellular atypia, tissue structure, and mitotic density) between the margin and the center (the center may be slightly more mature due to slow growth, but the difference is much smaller than that of malignant tumors).

[0081] According to one embodiment of the present invention, , and To preset weights, , and These represent the contributions of the information entropy term, the feature distribution discreteness term, and the edge-center feature contrast term to the tumor severity coefficient, respectively. The information entropy term controls the scale centrality of pathological features and, as the most fundamental and robust information theory indicator for measuring pathological disorder, is given the highest weight. It directly corresponds to the pathological variability, a core characteristic of tumor malignancy, such as... It can be set to 0.5. The discrete term of the feature distribution is used to control the spatial distribution uniformity of pathological features. It has a good evaluation ability for tumor types with uniform manifestation of various pathological features, but its general applicability is slightly weaker. For example, It can be set to 0.15. The edge-center feature contrast term is used to control the edge-center difference of pathological features. As the most spatially morphologically innovative indicator, it directly quantifies the key features observed by clinicians (such as boundaries and uniformity). It can be set to 0.35. The term represents the information entropy term. Discrete terms of characteristic distribution and edge-center feature contrast term Under the combined influence of various factors, the higher the value of each factor, the greater the tumor severity coefficient. The larger the tumor, the more malignant it is.

[0082] In this way, the tumor severity coefficient can be determined based on the number of pathological feature scales, the saliency of the pathological scales, and the pathological feature vector. During the calculation process, the contribution of the information entropy term, the feature distribution discrete term, and the edge-center feature contrast term to the tumor severity coefficient can be accurately evaluated, thus improving the accuracy of the tumor severity coefficient.

[0083] According to one embodiment of the present invention, in the deterioration risk module, a tumor deterioration risk coefficient is determined based on the pathological features and the training sample features.

[0084] According to one embodiment of the present invention, determining a tumor malignancy risk coefficient based on the pathological features and the training sample features includes:

[0085] Based on the pathological characteristics, determine the pathological feature vector, local sample feature vector variance, and global sample feature vector variance of the input sample;

[0086] Based on the characteristics of the training samples, determine the pathological feature vectors of benign samples and malignant samples;

[0087] The tumor malignancy risk coefficient is determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample.

[0088] For example, the pathological feature vector of the input sample refers to the pathological feature information of the sample to be evaluated (covering pathological information at different tissue levels). It is formed by splicing or aggregating pathological features from the multi-scale feature map of the input sample, and the acquisition method is consistent with that of the single-scale pathological feature vector. The local sample feature vector variance refers to the degree of fluctuation of pathological features at shallow scales (e.g., sub-regions or cell subpopulations within the tumor microenvironment) in the multi-scale feature map. It is calculated by vectorizing the pathological features at shallow scales and using the weighted average variance among the pathological feature vectors as the local sample feature vector variance. The global sample feature vector variance refers to the degree of fluctuation of pathological features at deep scales (e.g., the entire tumor tissue or the whole genome) in the multi-scale feature map. By vectorizing pathological features at a deep scale, the weighted average variance among the pathological feature vectors is calculated as the global sample feature vector variance. The benign sample pathological feature vector refers to the arithmetic mean of the pathological feature vectors of all samples labeled as benign by experts in the training set, representing the position of the benign pathological features of the tumor in the feature space. The malignant sample pathological feature vector refers to the arithmetic mean of the pathological feature vectors of all samples labeled as malignant by experts in the training set, representing the position of the malignant pathological features of the tumor in the feature space. Based on the pathological feature vectors of the input samples, the variance of the local sample feature vectors, the variance of the global sample feature vectors, the pathological feature vectors of benign samples, and the pathological feature vectors of malignant samples, the complex characteristics of the tumor in its development trend are comprehensively evaluated, and the tumor malignancy risk coefficient is determined.

[0089] According to an embodiment of the present invention, determining the tumor malignancy risk coefficient based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample includes: determining the tumor malignancy risk coefficient according to formula (2). , (2),

[0090] in, The pathological feature vector of the input sample. This represents the pathological feature vector of a benign sample. This is the pathological feature vector of malignant samples. The variance of the local sample feature vector. The variance of the global sample feature vector. For a very small positive value, , and For preset weights, and .

[0091] According to one embodiment of the present invention, The term is a relative distance term. The pathological feature vector of the input sample. This represents the pathological feature vector of a benign sample. This is the pathological feature vector of malignant samples. This is the Euclidean distance between the pathological feature vector of the input sample and the pathological feature vector of the benign sample, used to measure the degree to which the pathological features of the input sample are biased towards benign characteristics. The smaller the value, the more similar the pathological features of the input sample are to those of the benign sample. The larger the value, the more the pathological characteristics of the input sample deviate from benign. The Euclidean distance between the pathological feature vectors of the input sample and the pathological feature vectors of malignant samples is used to measure the degree to which the pathological features of the input sample are biased towards malignancy. The smaller the value, the more similar the pathological features of the input sample are to those of the malignant sample. The larger the value, the more the pathological characteristics of the input sample deviate from malignancy. The term represents the sum of the degree to which the pathological features of the input sample are biased towards benign and the degree to which the pathological features of the input sample are biased towards malignant. Item Use Item pair Normalization is performed to determine the position of the feature vector of the input sample on the benign-malignant pathological feature continuum. This is a very small positive value, to prevent the denominator from being zero and rendering the formula meaningless. The smaller the value of the term, the closer it is to 0, indicating that the pathological features of the input sample are closer to benign pathology in the feature space, much smaller than the distance to malignant pathology, and the lower the risk of future tumor malignancy. The larger the value of the term, the closer it is to 1, indicating that the pathological features of the input sample are closer to malignant pathology in the feature space than to benign pathology, and the higher the risk of future tumor deterioration.

[0092] According to one embodiment of the present invention, The term is the term indicating the direction of change. This is the vector representing the change in pathological features of the input sample relative to benign pathological features, indicating the direction in which the sample's pathology deviates from benign pathology. This is the standard evolutionary direction vector from benign pathological features to malignant pathological features, representing the direction of pathological standard deterioration. The term is used to measure the directional similarity between two vectors. The cosine similarity between the direction of pathological deviation from a benign state and the direction of pathological deterioration is obtained. ,in The angle between the direction in which the pathology deviates from the benign state and the direction in which the pathological standard indicates deterioration. When the value approaches 0°, it indicates that the direction of the pathological deviation from the benign state is becoming consistent with the direction of deterioration according to pathological standards. Tend to 1, It tends towards 1, that is, The higher the value of a number, the greater the risk of the tumor worsening in the future. When the angle approaches 180°, it indicates that the pathological deviation from a benign state has completely deviated from the pathological standard of deterioration, and the pathological features are developing towards benign pathological features. It tends towards -1. It tends towards 0, that is, The closer the value is to 0, the lower the risk of the tumor worsening in the future.

[0093] According to one embodiment of the present invention, The term is a locally stable term. The variance of the local sample feature vector represents the fluctuation of local pathological features. The larger the value, the higher the instability of features in localized areas of the surface. The variance of the global sample feature vector represents the fluctuation of global pathological features. This is a very small positive value, to prevent the denominator from being zero and rendering the formula meaningless. The term represents the ratio of the variance of the local sample feature vector to the variance of the global sample feature vector, used to measure the severity of fluctuations in local pathological features relative to fluctuations in global pathological features. The larger the value, the closer it is to 1, indicating that the local area containing the current pathological features is highly dispersed, possibly containing a mixture of benign and malignant pathological features, or located at the boundary of different pathological patterns. This suggests that the pathological state of the input sample is highly unstable, and the likelihood of future tumor malignancy is higher. The smaller the value, the closer it is to 0, indicating that the current pathological features are in a well-defined region with clustered features, the pathological state of the input sample is very stable, the future development trend of the tumor is clear, and the possibility of controlling the tumor to prevent further deterioration is higher.

[0094] According to one embodiment of the present invention, , and The preset weights represent the relative distance term, the direction of change term, and the local stability term, respectively, as factors affecting the risk coefficient of tumor progression. The degree of contribution, such as the relative distance term, is a core factor in judging the risk of tumor malignancy. The distance of a sample from malignant pathological features in the pathological feature space can directly reflect the malignancy status of the tumor. It can be set to 0.5. The direction of change is an important factor in judging the risk of tumor malignancy. In the early stages of tumor development, the sample may not be close to malignant pathological features, but it tends to become malignant, thus playing an early warning role. It can be set to 0.3. The local stability term is a general factor in assessing the risk of tumor progression. It is used to determine whether the pathological features of a sample are stable. Heterogeneous pathological features can amplify the risk of tumor progression. For example, It can be set to 0.2. Item in relative distance item Items with changing direction and local stability terms Under the combined influence of various factors, the higher the value of each factor, the greater the risk of tumor progression. The greater the risk of the input sample becoming a tumor in the future, the higher the risk of the tumor becoming a malignant tumor.

[0095] In this way, the tumor malignancy risk coefficient can be determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of benign samples, and the pathological feature vector of malignant samples. During the calculation process, the contribution of the relative distance term, the change direction term, and the local stability term to the tumor malignancy risk coefficient can be accurately assessed, thus improving the accuracy of the tumor malignancy risk coefficient.

[0096] According to one embodiment of the present invention, in the tumor classification module, the skin tumor classification is determined based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient.

[0097] According to one embodiment of the present invention, determining the classification of skin tumors based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient includes:

[0098] The pathological classification of skin tumors is determined based on the purity index of the pathological features.

[0099] Based on the tumor severity coefficient and the tumor malignancy risk coefficient, a threshold for the tumor severity coefficient and a threshold for the tumor malignancy risk coefficient are determined.

[0100] Based on the tumor severity coefficient threshold and the tumor malignancy risk coefficient threshold, a tumor benign or malignant risk assessment matrix is ​​constructed.

[0101] The skin tumor classification is determined based on the aforementioned pathological classification of skin tumors and the aforementioned benign / malignant tumor risk assessment matrix.

[0102] For example, a pathological feature purity index threshold can be set, and the pathological classification of skin tumors can be determined by comparing the pathological feature purity index of a sample with the threshold. When the pathological feature purity index is high, it indicates that the pathological information contained in the sample is clear and has little interference, making it suitable for tumor classification, grading, or prognosis. At that time, the purity index of the pathological features was medium purity, indicating that the pathological features of the sample could be used as candidate features, but they needed to be combined with other features (such as deep learning features and morphological features) for comprehensive judgment. When the pathological feature purity index is low, it indicates that the pathological features of the sample are severely interfered with (e.g., folded sections, uneven staining, large background ratio), making it unsuitable for direct classification. It is recommended to remove or enhance preprocessing. Set thresholds for tumor severity coefficient and tumor malignancy risk coefficient; for example, set 0.6 as the high-risk threshold for tumor severity coefficient and 0.5 as the high-risk threshold for tumor malignancy risk coefficient. A tumor benign / malignancy risk assessment matrix can be constructed based on these thresholds. and At this time, the area is benign, and the tumor is determined to be benign. The patient only needs to undergo routine annual checkups. and At this stage, the intervention zone is considered, and the tumor is highly suspected of being malignant. The pathological features are already highly abnormal, but the progression is relatively gradual. A biopsy is recommended to confirm the pathological diagnosis. and At this time, it is the follow-up area. The tumor is currently benign, but the trend of deterioration is quite obvious. Frequent follow-up is recommended (e.g., a check-up every 3-6 months). and At this time, it is a high-risk area, and the tumor is highly likely to be malignant, with a very obvious trend of deterioration. It is recommended to perform a biopsy immediately and actively treat the tumor. The specific type of the tumor is determined according to the pathological classification of skin tumors, and then the benign or malignant nature of the patient's tumor is determined according to the tumor benign or malignant risk assessment matrix. Specific decisions are made based on the benign or malignant nature of the tumor.

[0103] The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to embodiments of the present invention can obtain multi-scale feature maps based on a constructed multi-scale deep feature extraction network, determine pathological features and interference features based on a constructed variational autoencoder decoupling module, obtain training sample features based on a constructed adaptive learning training set, then determine a pathological feature purity index based on the multi-scale feature maps, pathological features, and interference features, determine a tumor severity coefficient based on the multi-scale feature maps and pathological features, determine a tumor malignancy risk coefficient based on the pathological features and training sample features, and further determine the skin tumor classification based on the pathological feature purity index, tumor severity coefficient, and tumor malignancy risk coefficient, thereby improving the accuracy and comprehensiveness of skin tumor classification. When determining the tumor severity coefficient, the tumor severity coefficient can be determined based on the number of pathological feature scales, the saliency of the pathological scales, and the pathological feature vector. During the calculation process, the contribution of the information entropy term, the feature distribution discrete term, and the edge-center feature contrast term to the tumor severity coefficient can be accurately evaluated, improving the accuracy of the tumor severity coefficient. When determining the risk coefficient of tumor malignancy, the risk coefficient can be determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample. During the calculation process, the contribution of the relative distance term, the change direction term, and the local stability term to the risk coefficient of tumor malignancy can be accurately assessed, thereby improving the accuracy of the risk coefficient of tumor malignancy.

[0104] Figure 2 An exemplary flowchart of a skin tumor classification method based on multi-scale decoupled representation and adaptive learning according to an embodiment of the present invention is shown, the method comprising:

[0105] Step S1: Construct a multi-scale deep feature extraction network to obtain multi-scale feature maps;

[0106] Step S2: Construct a variational autoencoder decoupling module and substitute the multi-scale feature map into the variational autoencoder decoupling module to determine pathological features and interference features;

[0107] Step S3: Construct an adaptive learning training set and obtain training sample features;

[0108] Step S4: Determine the pathological feature purity index based on the multi-scale feature map, the pathological features, and the interference features;

[0109] Step S5: Determine the tumor severity coefficient based on the multi-scale feature map and the pathological features;

[0110] Step S6: Determine the tumor malignancy risk coefficient based on the pathological features and the training sample features;

[0111] Step S7: Determine the skin tumor classification based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient.

[0112] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0113] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

Claims

1. A skin tumor classification system based on multi-scale decoupled representation and adaptive learning, characterized in that, include: The multi-scale feature module is used to construct a multi-scale deep feature extraction network to obtain multi-scale feature maps; An autoencoder decoupling module is used to construct a variational autoencoder decoupling module, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features; The training sample module is used to construct an adaptive learning training set and obtain training sample features; The feature purity module is used to determine the pathological feature purity index based on the multi-scale feature map, the pathological feature, and the interference feature; The severity module is used to determine the tumor severity coefficient based on the multi-scale feature map and the pathological features; The deterioration risk module is used to determine the tumor deterioration risk coefficient based on the pathological features and the training sample features; The tumor classification module is used to determine the classification of skin tumors based on the pathological feature purity index, the tumor severity coefficient, and the tumor malignancy risk coefficient.

2. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 1, characterized in that, Based on the multi-scale feature map, the pathological features, and the interference features, a pathological feature purity index is determined, including: Based on the pathological features and the interference features, the pathological factors and interference factors are determined; Based on the multi-scale feature map, the pathological first scale, the pathological second scale, the interference first scale, and the interference second scale are determined. The pathological feature purity index is determined based on the first pathological scale, the second pathological scale, the first interference scale, the second interference scale, the pathological factor, and the interference factor.

3. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 1, characterized in that, Based on the multi-scale feature map and the pathological features, the tumor severity coefficient is determined, including: Based on the multi-scale feature map, determine the number of pathological feature scales and the significance of the pathological feature in the pathological scale. Based on the pathological features, a pathological feature vector is determined, wherein the pathological feature vector includes: a single-scale pathological feature vector, a pathological feature edge region vector, and a pathological feature center region vector. The tumor severity coefficient is determined based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector.

4. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 3, characterized in that, The tumor severity coefficient is determined based on the number of pathological feature scales, the significance of the pathological scales, and the pathological feature vector, including: according to the formula ， Determining the tumor severity coefficient Where d is the number of pathological feature scales. The significance of the pathological feature at the i-th pathological scale. Let be the single-scale pathological feature vector of the i-th pathological scale. The vector represents the edge region of pathological features. The vector represents the central region of the pathological features. , and For preset weights, and , Let i be a very small positive value, i ≤ d, and both i and d are positive integers.

5. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 1, characterized in that, Based on the pathological features and the training sample features, the tumor malignancy risk coefficient is determined, including: Based on the pathological characteristics, determine the pathological feature vector, local sample feature vector variance, and global sample feature vector variance of the input sample; Based on the characteristics of the training samples, determine the pathological feature vectors of benign samples and malignant samples; The tumor malignancy risk coefficient is determined based on the pathological feature vector of the input sample, the variance of the local sample feature vector, the variance of the global sample feature vector, the pathological feature vector of the benign sample, and the pathological feature vector of the malignant sample.

6. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 5, characterized in that, Based on the pathological feature vectors of the input samples, the variance of the local sample feature vectors, the variance of the global sample feature vectors, the pathological feature vectors of the benign samples, and the pathological feature vectors of the malignant samples, the tumor malignancy risk coefficient is determined, including: according to the formula ， Determine the risk factor for tumor progression ,in, The pathological feature vector of the input sample. This represents the pathological feature vector of a benign sample. This is the pathological feature vector of malignant samples. The variance of the local sample feature vector. The variance of the global sample feature vector. For a very small positive value, , and For preset weights, and .

7. The skin tumor classification system based on multi-scale decoupled representation and adaptive learning according to claim 1, characterized in that, Based on the purity index of the pathological features, the tumor severity coefficient, and the tumor malignancy risk coefficient, the classification of skin tumors is determined, including: The pathological classification of skin tumors is determined based on the purity index of the pathological features. Based on the tumor severity coefficient and the tumor malignancy risk coefficient, a threshold for the tumor severity coefficient and a threshold for the tumor malignancy risk coefficient are determined. Based on the tumor severity coefficient threshold and the tumor malignancy risk coefficient threshold, a tumor benign or malignant risk assessment matrix is ​​constructed. The skin tumor classification is determined based on the aforementioned pathological classification of skin tumors and the aforementioned benign / malignant tumor risk assessment matrix.

8. A skin tumor classification method based on multi-scale decoupled representation and adaptive learning, characterized in that, The method is used in any one of the skin tumor classification systems based on multi-scale decoupled representation and adaptive learning as described in claims 1-7, comprising: Construct a multi-scale deep feature extraction network to obtain multi-scale feature maps; A variational autoencoder decoupling module is constructed, and the multi-scale feature map is substituted into the variational autoencoder decoupling module to determine pathological features and interference features; Construct an adaptive learning training set and obtain the features of the training samples; The pathological feature purity index is determined based on the multi-scale feature map, the pathological features, and the interference features. The tumor severity coefficient is determined based on the multi-scale feature map and the pathological features. The tumor malignancy risk coefficient is determined based on the pathological features and the training sample features. The skin tumor classification is determined based on the purity index of the pathological features, the tumor severity coefficient, and the tumor malignancy risk coefficient.

Images