Image recognition system for skin disease analysis
By introducing OFA structure reuse and extension mechanism and fractal enhancement graph model, the problems of insufficient adaptability and robustness of traditional skin disease analysis system are solved, and accurate identification and analysis of complex skin lesions are achieved.
Patent Information
- Application Number
- CN202511511345.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2025-11-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing skin disease analysis systems, traditional systems lack adaptability, robustness, and cross-population recognition capabilities when dealing with complex and variable skin images, making it difficult to effectively capture the potential evolution patterns of lesion areas.
By introducing OFA's structural reuse and extension mechanism, three-objective evaluation method, and distillation training strategy, and combining multifractal structural features and contrastive learning mechanism, an OFA three-objective extended distillation recognition model and a fractal-enhanced contrastive graph neural model are constructed to improve the system's adaptability and recognition accuracy.
It significantly improves the system's recognition ability and robustness in complex environments, accurately captures subtle structural differences in lesion areas, and enables efficient and interpretable skin lesion analysis.
Smart Images
Figure CN120976227A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to an image recognition system for skin disease analysis. Background Technology
[0002] With the rapid development of artificial intelligence technology, deep learning, image processing, and big data analysis are increasingly being applied in the field of medical image recognition and analysis, especially in areas such as high-dimensional image understanding, target recognition, and feature extraction, where significant progress has been made. However, traditional image recognition systems still have many shortcomings: First, traditional systems are trained using fixed architectures (such as ResNet, DenseNet, and EfficientNet), lacking adaptive design and optimization of the network structure. Faced with complex and varied skin lesion types, these fixed structures are difficult to flexibly adjust between capacity, receptive field, and parameter efficiency, leading to system image... First, the system has poor adaptability. Second, most existing systems prioritize recognition accuracy as the sole optimization objective, neglecting performance requirements in terms of structural complexity (such as the number of parameters and inference speed) and cross-group fairness (balanced recognition across different skin colors, ages, and genders). This single-objective optimization strategy easily leads to overfitting of the system to the training data and insufficient adaptability to complex scenarios in real-world environments. Finally, traditional systems model skin lesion areas as graph structures, but these are usually based solely on static graphs, lacking modeling of deep structural relationships between lesion areas and robustness to structural perturbations, making it difficult to effectively capture the potential evolutionary patterns of lesion areas. Summary of the Invention
[0003] This invention provides an image recognition system for skin disease analysis, aiming to solve the problems of poor recognition stability and limited analytical dimensions in existing systems when dealing with complex, variable, and high-dimensional skin images. It enhances the system's analytical capabilities and adaptability in complex application environments. The system constructs an OFA three-objective extended distillation recognition model by introducing OFA structural reuse and extension mechanisms, a three-objective evaluation method, and a distillation training strategy. This model processes high-dimensional skin image datasets, maintaining computational efficiency while enhancing structural adaptation and performance balance control for diverse image data, achieving stable recognition results under multi-source image data. Furthermore, the system is based on… Multifractal structural features, fractal dimensional differences, and contrastive learning mechanisms guide graph neural networks to learn graph representations with stronger structural expressive power and discriminative power. A fractal-enhanced contrastive graph neural model is constructed to process skin lesion region maps, improving the system's ability to model structural relationships between complex regions and its analytical consistency. In summary, this invention improves the adaptability and accuracy of image recognition systems in complex scenarios through structural reuse and extension mechanisms and fractal-enhanced graph modeling strategies. Combined with an image processing module, image annotation, segmentation, and heatmap generation are achieved based on recognition results and graph representations, providing an effective solution for building efficient and interpretable intelligent image analysis systems.
[0004] This invention provides an image recognition system for skin disease analysis, the system comprising a data acquisition module, an image recognition module, an image analysis module, and an image processing module;
[0005] The data acquisition module collects close-up images of skin lesions, multi-angle and multi-time point images, and comparison images of lesion and healthy areas, extracts image metadata, performs image denoising and image enhancement, and generates a high-dimensional skin image dataset.
[0006] The image recognition module constructs an OFA three-objective extended distillation recognition model through OFA's structure reuse and extension mechanism, three-objective evaluation method, and distillation training. This model processes high-dimensional skin image datasets to generate skin disease type recognition results. The OFA three-objective extended distillation recognition model includes a structure generation unit, a data partitioning unit, a benchmark training and evaluation unit, a structure extension control unit, and a distillation training unit. The structure generation unit is electrically connected to the data partitioning unit, the data partitioning unit is electrically connected to the benchmark training and evaluation unit, the benchmark training and evaluation unit is electrically connected to the structure extension control unit, and the structure extension control unit is electrically connected to the distillation training unit.
[0007] The image analysis module uses U-Net to extract features from a high-dimensional skin image dataset, followed by superpixel segmentation to construct a map of skin lesions. Through multifractal structure features, fractal dimension differences, and a contrastive learning mechanism, it guides the graph neural network to learn more robust and discriminative graph representations, constructing a fractal-enhanced contrastive graph neural model. This model processes the skin lesion map to generate discriminative graph representations. The fractal-enhanced contrastive graph neural model includes a lesion generation unit, a graph enhancement unit, a graph representation unit, and a training unit. The lesion generation unit is electrically connected to the graph enhancement unit, the graph enhancement unit is electrically connected to the graph representation unit, and the graph representation unit is electrically connected to the training unit.
[0008] The image processing module performs image annotation, image segmentation, and heatmap generation based on image recognition results and discriminative graph representation.
[0009] Furthermore, the image recognition module, in the process of generating image recognition results, specifically includes the following:
[0010] The structure generation unit constructs a constrained candidate network structure subspace through an OFA-based structure weight sharing mechanism. It uses its pre-trained structure and weights to achieve parameter sharing, enabling all candidate networks to efficiently reuse existing model capabilities in the same space, significantly reducing search and training costs, and providing good compatibility and transferability for subsequent structure expansion, thus obtaining a set of candidate network structures.
[0011] The data partitioning unit divides the high-dimensional skin image dataset into temporal data subsets through a time-aware data reconstruction strategy, constructs an incremental training data stream, and obtains an incremental training data subset sequence.
[0012] The benchmark training and evaluation unit performs phased training on the candidate network structure set through an incremental training data subset sequence, serving as a benchmark model; it uses a three-objective evaluation function to evaluate the recognition accuracy, structural robustness, and cross-population fairness of each structure in the candidate network structure set, generating an initial performance profile;
[0013] Based on the initial performance profile, the structural expansion control unit uses a three-bit binary control vector to perform structural expansion control on the models in the candidate network structure set in three dimensions: depth, width, and convolutional kernel size, to generate an expanded network structure.
[0014] The distillation training unit uses the output of the baseline model as a soft objective and employs the cross-distillation loss function to train the extended network structure. It integrates cross-entropy and KL divergence loss to achieve knowledge transfer and fine-tuning without freezing old weights, thereby effectively mitigating catastrophic forgetting and generating an adaptive distillation model. Based on the adaptive distillation model, image recognition results are generated.
[0015] Furthermore, the benchmark training and evaluation unit specifically includes: performing phased training on the candidate network structure set through an incremental training data subset sequence to serve as a benchmark model; evaluating the recognition accuracy, structural robustness, and cross-population fairness of each structure in the candidate network structure set using a three-objective evaluation function to generate an initial performance profile; the three-objective evaluation function includes a precision-complexity weighted function, a robustness function, and a fairness fitness function.
[0016] Furthermore, the image analysis module, in the process of generating discriminative map representations, specifically includes the following:
[0017] The skin lesion region map in the lesion generation unit includes nodes and edges. Nodes represent lesion regions, and edges represent the spatial relationships between regions. The skin lesion region map is normalized to unify the node feature dimensions, construct the initial feature representation of the nodes, and perform message propagation and embedding initialization on the nodes to generate a preliminary lesion map structure.
[0018] The graph enhancement unit performs subgraph sampling, edge weight perturbation, and node masking on the preliminary lesion graph structure to generate a multifractal enhanced graph while keeping the fractal structure of the skin lesion region graph unchanged.
[0019] The graph representation unit performs node-level message propagation on the multifractal enhanced graph to extract higher-order semantics; it converges to obtain the global graph representation vector and generates graph embedding pairs.
[0020] The training unit constructs a fractal-contrast learning loss function, combines it with graph embedding pairs, trains a graph neural network, obtains a trained graph neural network, and generates a discriminative graph representation through the trained graph neural network.
[0021] Furthermore, the training unit specifically includes: introducing a fractal dimension weighting method and a contrastive learning loss function to construct a fractal-contrast learning loss function. This loss function dynamically adjusts the weights of positive and negative sample pairs according to the fractal dimension differences between the enhanced graphs, thereby achieving accurate modeling and optimization of structural consistency; and training a graph neural network by combining graph embedding pairs to obtain the trained graph neural network.
[0022] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:
[0023] This invention constructs an OFA three-objective extended distillation recognition model, achieving stable processing of high-dimensional skin image data in a dermatology analysis system and significantly improving the system's recognition capability in complex environments. The algorithm effectively handles multi-source image inputs from different devices, time points, and shooting angles, especially ensuring consistent recognition and stable output of dermatology images when image quality is unstable and the sources are complex. Through structural reuse and extension mechanisms, the system can dynamically adjust the model structure according to the diversity of input images, improving the adaptability of image processing, reducing recognition inconsistencies caused by differences in image quality, and enhancing the system's robustness in complex scenarios with multi-source and variable dermatology images. Simultaneously, through a distillation training strategy, the system not only improves the stability of recognition results but also significantly reduces computational resource consumption, making the system more efficient and economical in practical deployments.
[0024] In processing the structural analysis of skin lesion areas, this invention introduces a fractal enhanced contrast map neural model, effectively improving the system's ability to analyze the microstructure and spatial relationships of lesion areas, especially suitable for processing complex skin lesion areas. In practical applications, skin lesions often exhibit characteristics such as blurred boundaries and irregular shapes, making it difficult for traditional methods to accurately capture these subtle differences. This invention, by introducing multifractal structural features and fractal dimension difference mechanisms, can more accurately capture the subtle structural differences between lesion areas. Combined with a contrastive learning mechanism, the system can adaptively analyze lesion areas, improving its adaptability to heterogeneous lesion areas. This innovative technology significantly solves the shortcomings of existing skin disease analysis systems in processing complex and variable lesions, improving the system's stability and consistency.
[0025] In summary, this invention effectively improves the adaptability, recognition accuracy, and semantic analysis depth of image recognition systems in complex environments through structural reuse and extended recognition mechanisms, as well as fractal structure-enhanced graph modeling strategies. The system employs multi-dimensional structure search and distillation optimization methods, combined with graph enhancement and contrastive learning techniques, enabling precise modeling and discriminative representation of complex image structures while maintaining a lightweight model. Furthermore, by integrating an image processing module, based on recognition results and discriminative graph representation, it achieves accurate image annotation, region segmentation, and heatmap visualization, providing technical support for constructing intelligent, robust, and interpretable image analysis systems. Attached Figure Description
[0026] Figure 1 This invention provides a schematic diagram of an image recognition system module for skin disease analysis.
[0027] Figure 2 The figures show a comparison of the performance of the fractal-contrast learning loss function and the contrastive learning loss function training graph neural network in Examples 5 and 6. Detailed Implementation
[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0029] Example 1, according to Figure 1 The present invention provides an image recognition system for skin disease analysis, the system comprising a data acquisition module, an image recognition module, an image analysis module and an image processing module;
[0030] The data acquisition module collects close-up images of skin lesions, multi-angle and multi-time point images, and comparison images of lesion and healthy areas, extracts image metadata, performs image denoising and image enhancement, and generates a high-dimensional skin image dataset.
[0031] The image recognition module constructs an OFA three-objective extended distillation recognition model through OFA's structure reuse and extension mechanism, three-objective evaluation method, and distillation training. This model processes high-dimensional skin image datasets to generate skin disease type recognition results. The OFA three-objective extended distillation recognition model includes a structure generation unit, a data partitioning unit, a benchmark training and evaluation unit, a structure extension control unit, and a distillation training unit. The structure generation unit is electrically connected to the data partitioning unit, the data partitioning unit is electrically connected to the benchmark training and evaluation unit, the benchmark training and evaluation unit is electrically connected to the structure extension control unit, and the structure extension control unit is electrically connected to the distillation training unit.
[0032] The image analysis module uses U-Net to extract features from a high-dimensional skin image dataset, followed by superpixel segmentation to construct a map of skin lesions. Through multifractal structure features, fractal dimension differences, and a contrastive learning mechanism, it guides the graph neural network to learn more robust and discriminative graph representations, constructing a fractal-enhanced contrastive graph neural model. This model processes the skin lesion map to generate discriminative graph representations. The fractal-enhanced contrastive graph neural model includes a lesion generation unit, a graph enhancement unit, a graph representation unit, and a training unit. The lesion generation unit is electrically connected to the graph enhancement unit, the graph enhancement unit is electrically connected to the graph representation unit, and the graph representation unit is electrically connected to the training unit.
[0033] The image processing module performs image annotation, image segmentation, and heatmap generation based on image recognition results and discriminative graph representation.
[0034] Example 2, based on Example 1, describes the process of generating image recognition results using the image recognition module, specifically including the following:
[0035] The structure generation unit constructs a constrained candidate network structure subspace through an OFA-based structure weight sharing mechanism. It uses its pre-trained structure and weights to achieve parameter sharing, enabling all candidate networks to efficiently reuse existing model capabilities in the same space, significantly reducing search and training costs, and providing good compatibility and transferability for subsequent structure expansion, thus obtaining a set of candidate network structures.
[0036] The data partitioning unit divides the high-dimensional skin image dataset into temporal data subsets through a time-aware data reconstruction strategy, constructs an incremental training data stream, and obtains an incremental training data subset sequence.
[0037] The benchmark training and evaluation unit performs phased training on the candidate network structure set through an incremental training data subset sequence, serving as a benchmark model; it uses a three-objective evaluation function to evaluate the recognition accuracy, structural robustness, and cross-population fairness of each structure in the candidate network structure set, generating an initial performance profile;
[0038] Based on the initial performance profile, the structure expansion control unit uses a three-bit binary control vector to perform structure expansion control on the models in the candidate network structure set in three dimensions: depth, width, and convolutional kernel size, generating expanded network structures to build candidate models with stronger structural capacity. The expansion process follows the compatibility constraints of the OFA network architecture search space, searching for insertable structural modules layer by layer from the end of the network forward, prioritizing the maximization of parameter reuse and structural compatibility, thereby avoiding redundant training and structural oscillations.
[0039] The distillation training unit uses the output of the baseline model as a soft objective and employs the cross-distillation loss function to train the extended network structure. It integrates cross-entropy and KL divergence loss to achieve knowledge transfer and fine-tuning without freezing old weights, thereby effectively mitigating catastrophic forgetting and generating an adaptive distillation model. Based on the adaptive distillation model, image recognition results are generated.
[0040] Example 3, based on Example 2, specifically includes the following steps for the benchmark training and evaluation unit: performing phased training on a candidate network structure set using an incremental training data subset sequence to serve as a benchmark model; evaluating the recognition accuracy, structural robustness, and cross-population fairness of each structure in the candidate network structure set using a three-objective evaluation function to generate an initial performance profile; the three-objective evaluation function includes a precision-complexity weighted function, a robustness function, and a fairness fitness function, and the formulas used are as follows:
[0041] Precision-complexity weighted function:
[0042] ;
[0043] in, This represents the network structure feature representation of the candidate network structures in the candidate network structure set. This indicates the parameter configuration of the network structure in the candidate network structure. Indicates the weighting factor. This represents the precision-complexity weighted function value. Indicating in the evaluation set The recognition accuracy on Indicates the network structure complexity in the candidate network structure;
[0044] Robustness function formula:
[0045] ;
[0046] in, Indicates the disturbance intensity coefficient. Represents the robustness function value. Indicates the number of noise disturbance samples. Indicates the noise disturbance sampling index. Indicating in the evaluation set The weight parameters on, Indicating in the evaluation set The change in accuracy after the perturbation; Indicates the first The standard normal noise vector of the second perturbation; This represents the weighted noise disturbance term;
[0047] Fairness fitness function:
[0048] ;
[0049] in, This represents the value of the fairness fitness function. Indicates the number of people in a group. Represents a group index. This indicates the specific model comprised of the candidate network structures currently being evaluated in the population. The recognition accuracy on This represents the variance function.
[0050] Example 4, based on Example 2, specifically includes the following: performing phased training on the candidate network structure set using an incremental training data subset sequence as a benchmark model; evaluating the recognition accuracy and structural robustness of each structure in the candidate network structure set using a bi-objective evaluation function to generate an initial performance profile; the bi-objective evaluation function includes a precision-complexity weighted function and a robustness function.
[0051] Example 5, according to Figure 2 This embodiment is based on Embodiment 3. In this embodiment, the process of generating discriminative image representation by the image analysis module specifically includes the following:
[0052] The skin lesion region map in the lesion generation unit includes nodes and edges. Nodes represent lesion regions, and edges represent the spatial relationships between regions. The skin lesion region map is normalized to unify the node feature dimensions, construct the initial feature representation of the nodes, and perform message propagation and embedding initialization on the nodes to generate a preliminary lesion map structure.
[0053] The graph enhancement unit performs subgraph sampling, edge weight perturbation, and node masking on the preliminary lesion graph structure to generate a multifractal enhanced graph while keeping the fractal structure of the skin lesion region graph unchanged.
[0054] The graph representation unit performs node-level message propagation on the multifractal enhanced graph to extract higher-order semantics; it converges to obtain the global graph representation vector and generates graph embedding pairs.
[0055] The training unit constructs a fractal-contrast learning loss function, combines it with graph embedding pairs, trains a graph neural network, and obtains a trained graph neural network; through the trained graph neural network, a discriminative graph representation is generated.
[0056] Example 6, according to Figure 2 This embodiment is based on Embodiment 3. In this embodiment, the process of generating discriminative image representation by the image analysis module specifically includes the following:
[0057] The skin lesion region map in the lesion generation unit includes nodes and edges. Nodes represent lesion regions, and edges represent the spatial relationships between regions. The skin lesion region map is normalized to unify the node feature dimensions, construct the initial feature representation of the nodes, and perform message propagation and embedding initialization on the nodes to generate a preliminary lesion map structure.
[0058] The graph enhancement unit performs subgraph sampling, edge weight perturbation, and node masking on the preliminary lesion graph structure to generate a multifractal enhanced graph while keeping the fractal structure of the skin lesion region graph unchanged.
[0059] The graph representation unit performs node-level message propagation on the multifractal enhanced graph to extract higher-order semantics; it converges to obtain the global graph representation vector and generates graph embedding pairs.
[0060] The training unit constructs a contrastive learning loss function, combines it with graph embedding pairs, trains a graph neural network, and obtains a trained graph neural network; through the trained graph neural network, a discriminative graph representation is generated.
[0061] Example 7, based on Example 5, specifically includes the following training unit: Introducing a fractal dimension weighting method and a contrastive learning loss function to construct a fractal-contrast learning loss function. This loss function dynamically adjusts the weights of positive and negative sample pairs based on the fractal dimension differences between the enhanced graphs, achieving accurate modeling and optimization of structural consistency. Combined with graph embedding pairs, a graph neural network is trained to obtain a trained graph neural network. Through the trained graph neural network, a discriminative graph representation is generated. The formula used is as follows:
[0062] Fractal dimension difference weighting formula:
[0063] ;
[0064] in, Indicates the graph sample index. Indicates the first A diagram of the skin lesion area. Indicates the first A multi-fractal enhancement map, that is, for The image obtained after enhancement; It is defined as, express fractal dimension, express fractal dimension; Indicates the scaling factor. Represents an exponential function; express and fractal dimension weighting factor;
[0065] ;
[0066] in, Indicates the index of candidate positive and negative samples. Indicates the first Fractal-contrast learning loss value for each image sample. Representation diagram The graph representation of vectors, Representation of augmented graph The graph representation of vectors, express and Similarity between them Indicates the temperature coefficient. Represents an exponential function. Indicates the number of samples; Indicates the first A diagram of the skin lesion area. Indicates the first A multi-fractal enhancement map, Representation of augmented graph The graph representation of vectors, express and The similarity between them; express and The fractal dimension weighting factor.
[0067] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.
Claims
1. An image recognition system for skin disease analysis, comprising a data acquisition module, wherein the data acquisition module acquires a high-dimensional skin image dataset; characterized in that: The system also includes an image recognition module, an image analysis module, and an image processing module; The image recognition module constructs an OFA three-objective extended distillation recognition model, processes high-dimensional skin image datasets through the OFA three-objective extended distillation recognition model, and generates image recognition results. The OFA three-objective extended distillation recognition model includes a benchmark training and evaluation unit, a structure extension control unit, and a distillation training unit. The benchmark training and evaluation unit is electrically connected to the structure extension control unit, and the structure extension control unit is electrically connected to the distillation training unit; The image analysis module processes high-dimensional skin image datasets to construct skin lesion area maps. A fractal-enhanced contrast image neural model is constructed. This model is used to process skin lesion area images and generate discriminative image representations. The fractal-enhanced contrast image neural model includes graph enhancement units, graph representation units, and training units. Graph enhancement units are electrically connected to graph representation units, and graph representation units are electrically connected to training units. The image processing module performs image annotation, image segmentation, and heatmap generation based on image recognition results and discriminative graph representation.
2. The image recognition system for skin disease analysis according to claim 1, characterized in that: The benchmark training and evaluation unit constructs a candidate network structure set and performs phased training on the candidate network structure set through an incremental training data subset sequence to serve as a benchmark model. The three-objective evaluation function is used to evaluate the recognition accuracy, structural robustness and cross-population fairness of each structure in the candidate network structure set, generating an initial performance profile.
3. The image recognition system for skin disease analysis according to claim 2, characterized in that: Based on the initial performance profile, the structure expansion control unit uses a three-bit binary control vector to perform structure expansion control on the three dimensions of the candidate network structure set, generating an expanded network structure.
4. The image recognition system for skin disease analysis according to claim 3, characterized in that: The distillation training unit uses the output of the baseline model as a soft target and employs the cross-distillation loss function to train the extended network structure, generating an adaptive distillation model. Based on the adaptive distillation model, image recognition results are generated.
5. The image recognition system for skin disease analysis according to claim 1, characterized in that: The OFA three-target extended distillation identification model also includes a structure generation unit and a data partitioning unit.
6. The image recognition system for skin disease analysis according to claim 1, characterized in that: The graph enhancement unit constructs a preliminary lesion graph structure. By performing subgraph sampling, edge weight perturbation, and node masking on the preliminary lesion graph structure, a multifractal enhancement graph is generated.
7. An image recognition system for skin disease analysis according to claim 6, characterized in that: The graph representation unit generates graph embedding pairs by performing node-level message propagation on the multifractal enhanced graph.
8. An image recognition system for skin disease analysis according to claim 7, characterized in that: The training unit constructs a fractal-contrast learning loss function by introducing a fractal dimension weighting method and a contrastive learning loss function. Combined with graph embedding pairs, it trains a graph neural network to obtain a trained graph neural network. Through the trained graph neural network, discriminative graph representations are generated.