Intelligent auxiliary diagnosis system for skin diseases

Abstract

The invention relates to the technical field of disease diagnosis, in particular to a skin disease intelligent auxiliary diagnosis system which comprises a data input module, an image processing module, a feature extraction module, a diagnosis and classification module and a treatment recommendation module. Compared with the prior art which only depends on single-point lesion area image input and has the defect that the false positive rate is increased due to the fact that individual differences and environmental factor interference cannot be eliminated, the scheme adopts synchronous input of lesion area images and symmetrical skin images of the lesion area images on a human body symmetrical structure; by comparing and analyzing characteristic differences such as colors, textures and forms through two channels, the method has the advantages that physiological changes and pathological abnormalities are accurately distinguished (for example, abnormal pigmentation is judged if the symmetric part delta E chromatic aberration is larger than 3), and the environment illumination interference misjudgment rate is reduced by 38%.

Description

Technical Field

[0001] The present invention relates to the technical field of disease diagnosis, and in particular to an intelligent auxiliary diagnosis system for skin diseases. Background Art

[0002] Existing dermatological diagnostic technology mainly relies on doctors' naked eye observation and dermatoscope examination. Its diagnostic accuracy is limited by the doctors' experience level and equipment resolution. Especially in primary medical institutions, the lack of professional dermatologists leads to a high misdiagnosis rate.

[0003] Although traditional digital image processing technology attempts to perform AI-assisted diagnosis using ordinary mobile phone photos, it is difficult to restore the micron-level structure of the dermatoscope level (such as hair follicle morphology and epidermal texture) due to problems such as blur, color difference and light interference in ordinary photos. In addition, existing super-resolution algorithms generally use a single interpolation or simple convolutional network, which cannot effectively distinguish the characteristic differences between the diseased area and normal skin.

[0004] Furthermore, existing AI models are mostly trained on public datasets, which are skewed towards Caucasian skin tones, resulting in a misdiagnosis rate of up to 22% for dark-skinned patients. They also lack the ability to dynamically adapt to rare diseases (such as hereditary epidermolysis bullosa) and emerging skin conditions (such as the COVID-19 rash). When recommending treatment options, traditional rule-based engines rely solely on matching disease names to generic drug lists, ignoring the personalized impact of a patient's genetic profile (such as skin barrier defects caused by FLG mutations) and environmental exposures (such as oxidative stress caused by PM2.5), making it difficult to achieve precision medicine.

[0005] The above limitations have severely restricted the standardization and universalization of skin disease diagnosis and treatment, and urgently need to achieve technological breakthroughs through multimodal data fusion and adaptive learning mechanisms. Summary of the Invention

[0006] In order to overcome the problems raised in the above background technology, the present invention proposes an intelligent auxiliary diagnosis system for skin diseases.

[0007] The technical solution of the present invention is: an intelligent auxiliary diagnosis system for skin diseases, comprising: The data input module is used to input the user's normal skin close-up photos. The data input module supports users to input normal skin close-up photos. Compared with professional dermatoscope equipment, ordinary close-up photos are more convenient and less costly to obtain. Users do not need to use professional medical equipment to complete data collection, which greatly reduces the threshold for patients to seek medical treatment and helps with the early detection and diagnosis of diseases, especially for patients in remote areas or areas with scarce medical resources; The image processing module is used to pre-process the recent normal skin photos input by the data input module and convert them into dermatoscope-grade super-resolution images. The image processing module can pre-process the recent normal skin photos and convert them into dermatoscope-grade super-resolution images. The pre-processing process can remove noise from the image, adjust brightness and contrast, and improve image quality. The conversion to dermatoscope-grade super-resolution images can provide clearer and richer skin details, simulating the observation effect of a professional dermatoscope, providing a solid foundation for subsequent accurate diagnosis, helping to detect early and subtle skin lesions, and improving the sensitivity and specificity of diagnosis. The feature extraction module is used to extract features from the dermoscopic super-resolution image obtained after processing by the image processing module. The feature extraction module can accurately extract key features from the processed dermoscopic super-resolution image. These features may include skin texture, color distribution, lesion morphology and other multi-dimensional information. Through the automated feature extraction process, it can quickly and comprehensively capture information closely related to skin disease diagnosis, avoiding feature omissions that may be caused by subjective factors or lack of experience during manual diagnosis, and improving the accuracy and consistency of diagnosis. The Diagnosis and Classification Module diagnoses abnormalities based on the features extracted by the Feature Extraction Module and uses a classifier to classify abnormalities. This module utilizes advanced algorithms and models to quickly determine if skin lesions are abnormal and accurately classify the type of abnormality. This significantly shortens diagnosis time and improves diagnostic efficiency, helping patients receive timely treatment recommendations and avoid delays in treatment. The treatment recommendation module uses knowledge graph matching to generate and recommend personalized treatment plans based on the diagnostic results of the diagnosis and classification modules. The knowledge graph integrates a wealth of medical knowledge related to skin diseases, including treatment methods, medication information, and treatment case studies. By matching the patient's specific condition characteristics, the most appropriate treatment plan can be tailored for each patient, fully considering individual differences such as age, gender, physical condition, and allergy history, thereby improving the targeted and effective treatment.

[0008] Preferably, when the data input module inputs a user's recent photo of normal skin, the data input module specifically includes: S11: Multimodal data collection, recording the user's basic information, medical history, and symptom data. In addition to a recent skin image, the user's basic information (such as age, gender, occupation, etc.), medical history (previous medical history, family genetic history, etc.), and symptom data (onset time, symptoms, course of disease, etc.) are recorded, providing the diagnostic system with multi-dimensional and comprehensive information. This information complements and verifies each other, helping to more accurately determine the cause, type, and stage of skin diseases, avoiding misdiagnosis or missed diagnosis that may result from relying solely on image information, and improving the comprehensiveness and accuracy of diagnosis. S12: Image acquisition: Users use an image acquisition device to take a close-up photo of their skin and enter it into the data input module. Users use an image acquisition device (such as a common device like a mobile phone camera) to take a close-up photo of their skin and enter it into the system. The operation is simple and convenient, without the need for specialized medical equipment or complex operation procedures. This lowers the user threshold, allowing more patients to easily participate in the initial diagnosis of diseases. It is particularly suitable for telemedicine scenarios, allowing patients to conduct preliminary data collection at home, reducing medical costs and time. S13: Image quality detection: This system performs quality checks on close-up skin images taken by the user and reminds the user to retake the image if the quality test fails. It performs quality checks on close-up skin images taken by the user to ensure that the images input into the system are clear, complete, and meet diagnostic requirements. Low-quality images may contain noise, blur, uneven lighting, and other issues, which may affect the accuracy of subsequent image processing and analysis. By performing quality checks and reminding the user to retake the image, data quality can be guaranteed from the source, providing reliable image data for subsequent diagnostic modules and improving the performance of the entire diagnostic system. S14: Data transmission: The input data and recent skin images are encrypted using an encryption algorithm and then transmitted. This effectively prevents data theft, tampering, or leakage during transmission. Skin images and user-related information are considered personal privacy data. Encrypted transmission ensures data security and complies with relevant laws and regulations on data privacy protection. This enhances user trust in the system and increases patients' willingness to use it for disease diagnosis.

[0009] Preferably, when the user takes a close-up photo of the skin using an image acquisition device and inputs it into the data input module, the user inputs two sets of close-up photos of the skin: an image of the lesion area and an image of the skin symmetrically located in the lesion area on the human body. Simultaneously inputting the image of the lesion area and its symmetrical counterpart provides a natural comparison sample for the diagnostic system. Under normal circumstances, the skin in symmetrical areas of the human body has similar physiological structures and characteristics. By comparing the differences in color, texture, and morphology between the lesion area and the symmetrical skin, doctors or the system can more intuitively detect abnormalities in the lesion, helping to eliminate interference from normal skin changes caused by individual differences, environmental factors, and the like, thereby improving diagnostic sensitivity and specificity. For example, when determining whether skin pigmentation is a lesion, comparing the skin color of symmetrical areas can more accurately determine whether there is an abnormal increase or decrease in pigmentation in the lesion area.

[0010] Preferably, when the image processing module pre-processes the ordinary skin close-up photo input by the data input module and converts it into a dermatoscope-level super-resolution image, it specifically includes: S21: Image preprocessing, using a high-order degradation model to simulate the real dermoscopic imaging process, and using bicubic interpolation denoising and histogram equalization to enhance contrast and solve the blurring problem of ordinary photos; using a high-order degradation model to simulate the real dermoscopic imaging process, the preprocessed image is closer to the actual dermoscopic observation effect, providing a more reliable basis for subsequent analysis. Bicubic interpolation denoising can effectively remove the noise introduced by shooting equipment, environment and other factors in ordinary photos, reduce the interference of noise on image information, and make the lesion features clearer. Histogram equalization enhances contrast, which can highlight the detailed information in the image, solve the blurring problem that may exist in ordinary photos, and make features such as skin texture and lesion boundaries more obvious, which is convenient for subsequent feature extraction and diagnostic analysis. For example, for some early and subtle skin lesions, the difference between the lesion area and normal skin can be more clearly identified after enhancing the contrast; S22: Super-resolution reconstruction, based on conditional generative adversarial networks, uses residual dense blocks to generate dermoscopic-level images; Based on conditional generative adversarial networks, and using residual dense blocks to generate dermoscopic-level images, it can significantly improve the resolution of images. Dermoscopic-level images can provide clearer and more detailed information about the microstructure of the skin, capturing subtle lesion features that are difficult to detect in ordinary photos, such as subtle changes in skin texture and tiny blood vessel dilation. These subtle features are crucial for the accurate diagnosis of certain skin diseases. For example, when diagnosing early skin cancer, high-resolution images can help detect tiny signs of cancer cell infiltration or abnormal blood vessel patterns, thereby achieving early diagnosis and timely treatment; S23: Color correction, color gamut mapping for different skin tones, eliminating the impact of illumination bias on pigmentation assessment; people with different skin tones have significantly different skin colors, and changes in lighting conditions will also affect the color of skin images. Color gamut mapping for different skin tones can eliminate the impact of illumination bias on pigmentation assessment, making the system more accurate in analyzing skin pigmentation changes. In the diagnosis of skin diseases, abnormal pigmentation is one of the important characteristics of many diseases. For example, melanoma often manifests as darker pigmentation and uneven color. Through color correction, regardless of the patient's skin color, the system can accurately assess the pigmentation of the lesion area, avoiding misdiagnosis or missed diagnosis due to skin color and lighting factors.

[0011] Preferably, when performing super-resolution reconstruction, the specific principle steps include: S31: Input preprocessing: De-noising, contrast enhancement, and normalization of a close-up photo of normal skin to a standard size. De-noising, contrast enhancement, and normalization of a close-up photo of normal skin to a standard size effectively improve the quality of the input image. De-noising eliminates noise introduced by the camera and the environment, resulting in a clearer image. Contrast enhancement highlights the difference between the diseased area and normal skin, making key information easier to identify. Normalization to a standard size facilitates subsequent unified processing and analysis, ensuring the system operates stably with input images of varying sizes. This lays a solid foundation for subsequent feature extraction and diagnostic analysis, improving diagnostic reliability. S32: Feature Extraction and Mapping. The RRDB network extracts hierarchical features from low-resolution images through multi-level residual connections and uses the feature reuse mechanism within dense blocks to enhance the representation of high-frequency details. The RRDB network extracts hierarchical features from low-resolution images through multi-level residual connections and uses the feature reuse mechanism within dense blocks to enhance the representation of high-frequency details. Multi-level residual connections can solve the gradient vanishing problem during deep network training, allowing the network to learn deeper feature information. The feature reuse mechanism within dense blocks makes full use of features at different levels, avoiding information loss and better capturing subtle features in skin images, such as subtle changes in skin texture and tiny blood vessel dilation. This provides rich feature support for generating high-quality high-resolution images and helps improve the detection of skin lesions. S33: High-resolution image generation, the generator upsamples the intermediate feature map to the target resolution through the deconvolution layer, and combines the attention mechanism to dynamically allocate the generation weights of different regions, giving priority to restoring key structures; the generator upsamples the intermediate feature map to the target resolution through the deconvolution layer, and combines the attention mechanism to dynamically allocate the generation weights of different regions, giving priority to restoring key structures. The deconvolution layer upsampling process can gradually enlarge the low-resolution image to the target resolution, and the introduction of the attention mechanism enables the system to allocate computing resources according to the importance of different regions, giving priority to key areas such as lesion areas, and ensuring that the image details of these areas are better restored and presented. This helps doctors observe the structure and morphology of lesions more clearly and improve the accuracy of diagnosis, especially in the diagnosis of early skin lesions, and can discover some subtle features that are easy to overlook; S34: Post-processing and optimization. This technique uses wavelet-domain fidelity loss to suppress artifacts, optimize detail clarity in high-frequency subbands, and ensure natural transitions of microstructures. During the image generation process, artifacts or unnatural details may appear, affecting the image's visual quality and diagnostic value. The wavelet-domain fidelity loss function specifically suppresses these artifacts while optimizing high-frequency subbands. This allows for the natural and clear presentation of microstructures, such as skin cell texture and lesion edge details, further improving image quality and diagnostic reliability.

[0012] Preferably, when the feature extraction module extracts features from the dermatoscope-level super-resolution image obtained after processing by the image processing module, it specifically includes: S41: Identification of lesion areas uses an image segmentation algorithm to separate normal skin areas from lesion areas. Using an image segmentation algorithm to separate normal skin areas from lesion areas can accurately locate the lesion area, providing a clear target for subsequent feature extraction and analysis. This helps avoid interference with the normal skin area, allowing the system to focus on analyzing the characteristics of the lesion area, improving the efficiency and accuracy of diagnosis. At the same time, accurate lesion area positioning also provides an important basis for subsequent treatment plan formulation. Doctors can develop more personalized treatment plans based on information such as the size and location of the lesion area. S42: Color feature extraction: convert the image from RGB to LAB color space and calculate the color difference between the lesion area and the surrounding normal skin. The calculation principle formula is: ; in, is color difference, is brightness, is the red-green axis chromaticity, is the yellow-blue axis chromaticity, is the brightness difference between the lesion area and normal skin, and The red-green axis chromaticity and yellow-blue axis chromaticity differences between the lesion area and normal skin respectively; convert the image from RGB to LAB color space, and calculate the color difference between the lesion area and the surrounding normal skin. LAB color space is more consistent with the visual perception of the human eye and can more accurately describe the color difference. By calculating the color difference, the degree of difference in color between the lesion area and normal skin can be quantified. This difference information is of great significance for judging the type and severity of skin diseases. For example, certain skin diseases can cause specific changes in skin color, such as pigmentation, redness, yellowing, etc. By calculating the color difference, these changes can be evaluated more objectively, providing strong support for diagnosis; S43: Texture feature extraction, statistically analyzing the grayscale distribution of pixel pairs in the lesion area; statistically analyzing the grayscale distribution of pixel pairs in the lesion area can capture the texture characteristics of the lesion area. Changes in skin texture are one of the important manifestations of many skin diseases, such as rough, smooth, and scaly skin. By analyzing the grayscale distribution of pixel pairs, the complexity and directionality of the texture in the lesion area can be revealed. These features help distinguish different types of skin diseases, especially for some skin diseases with similar symptoms, where texture features can serve as an important basis for identification. S44: Morphological feature extraction, automatically draw an initial boundary based on the color difference of the lesion area, and make the outline fit the edge of the real lesion through iterative adjustment, and then calculate the aspect ratio difference and edge irregularity of the lesion area; automatically draw an initial boundary based on the color difference of the lesion area, and make the outline fit the edge of the real lesion through iterative adjustment, and then calculate the aspect ratio difference and edge irregularity of the lesion area. Automatically drawing and adjusting the lesion boundary can more accurately describe the morphology of the lesion area. Morphological features such as aspect ratio difference and edge irregularity can reflect the growth pattern and characteristics of the lesion. For example, the edges of benign lesions are usually more regular, while the edges of malignant lesions may be irregular, jagged, etc. Quantifying these morphological features can help determine whether the lesion is benign or malignant, and evaluate the development trend of the lesion; S45: Eliminate irrelevant differences. The generator in the adversarial network outputs lesion features, and the discriminator forces the feature distribution to be independent of skin color and age. The gradient reversal layer is used to reverse the gradient direction during backpropagation, reducing racial correlation in the feature space. The generator in the adversarial network outputs lesion features, and the discriminator forces the feature distribution to be independent of skin color and age. The gradient reversal layer is used to reverse the gradient direction during backpropagation, reducing racial correlation in the feature space. In the diagnosis of skin diseases, factors such as skin color, age, and race may have a certain impact on the manifestation of the lesion, resulting in differences in lesion characteristics among different populations. By eliminating the interference of these irrelevant factors, the system can focus more on the characteristics of the lesion itself, improving the universality and accuracy of diagnosis. This makes the system applicable to people of different skin colors, ages, and races, and provides accurate diagnostic services to a wider range of patients.

[0013] Preferably, when using an image segmentation algorithm to segment the normal skin area and the lesion area, the method specifically includes: S51: Normal skin feature extraction: Feature extraction is performed on the user-input image of the non-lesioned area, i.e., the skin image of the lesioned area symmetrically located on the human body. By extracting features from the skin image of the lesioned area symmetrically located on the human body, typical features of normal skin are obtained. This method of obtaining normal skin features based on symmetrical parts of the human body can minimize interference caused by individual differences, environmental factors, and other variations in normal skin features, providing a reliable benchmark for subsequent comparison with lesioned skin features, ensuring accurate positioning of the lesioned area. S52: Lesion skin feature extraction: This function extracts features from the lesion area image input by the user. This feature extraction can fully capture the unique characteristics of the lesion skin, including abnormalities in color, texture, morphology, etc. These features are the key basis for distinguishing lesions from normal skin, providing rich information for subsequent feature comparison and lesion area identification. S53: Feature comparison: Compare the features of normal skin with those of lesioned skin to identify abnormal features of the lesioned skin. This comparison and analysis method can highlight the specificity of the lesioned area, eliminate the interference of normal skin features on lesion identification, greatly improve the accuracy and reliability of lesion area identification, and provide a precise target area for subsequent diagnosis and treatment. S54: Image segmentation: Based on the identified abnormal features, the lesion skin contour is identified and image segmentation is performed to determine the lesion area. This lesion skin contour identification and image segmentation based on the identified abnormal features can accurately separate the lesion area from normal skin. Accurate image segmentation results provide clear lesion area boundaries for subsequent feature extraction, diagnosis, and classification, preventing interference from normal skin areas in lesion feature analysis, and helping to improve the accuracy and efficiency of the entire diagnostic system.

[0014] Preferably, the diagnosis and classification module, when performing abnormality diagnosis based on the features extracted by the feature extraction module and performing abnormality classification using a classifier, specifically includes: S61: Primary classification, using a support vector machine to quantitatively analyze the seven-point melanoma checklist features, including diameter >6mm, irregular margins, uneven color, and dynamic change indicators. This support vector machine-based quantitative analysis of the seven-point melanoma checklist features can quickly and accurately identify patients with typical melanoma features. The seven-point checklist features (diameter >6mm, irregular margins, uneven color, and dynamic change indicators) are important for melanoma diagnosis. By quantitatively analyzing these features, the support vector machine can perform preliminary screening for possible melanoma lesions at an early stage, improving diagnostic sensitivity and specificity and buying time for further diagnosis and treatment. S62: Secondary classification, using the EfficientNetV2 model, trained based on the HAM10000 dataset, and optimizing computational efficiency through compound scaling, wherein the confidence threshold is set to 0.85, and if the model output is lower than the threshold, a manual review process is triggered; The EfficientNetV2 model is trained based on the HAM10000 dataset, and compound scaling is used to optimize computational efficiency, which can improve the operating speed of the system while ensuring diagnostic accuracy. The EfficientNetV2 model has powerful feature extraction and classification capabilities, and through training on large-scale datasets, it can learn richer characteristic information about skin diseases. The confidence threshold is set to 0.85, and when the model output is lower than the threshold, a manual review process is triggered, which not only ensures accurate automatic diagnosis of most common skin diseases, but also performs manual intervention on some complex or uncertain cases, thereby improving the reliability and safety of diagnosis; S63: Three-level classification integrates the user's basic information, medical history, and symptom data to achieve precise classification. The diagnosis of skin diseases relies not only on skin image features but also closely correlates with basic information such as the patient's age, gender, family history, site of onset, and symptom progression. By integrating this multi-dimensional data, a more comprehensive understanding of the patient's condition can be achieved, leading to more accurate classification of skin diseases and providing strong support for developing personalized treatment plans.

[0015] Preferably, the treatment recommendation module generates and recommends personalized treatment plans based on the diagnosis results of the diagnosis and classification modules using a knowledge graph matching method, specifically including: S71: Data integration, integrating disease-drug-gene relationships, clinical guidelines, drug side effect databases, and environmental medicine research; integrating multi-source data such as disease-drug-gene relationships, clinical guidelines, drug side effect databases, and environmental medicine research to build a comprehensive and systematic knowledge system for skin disease treatment. Disease-drug-gene relationship data can help doctors understand the relationship between different skin diseases and specific drugs and genes, providing a basis for precision treatment; clinical guidelines provide a standardized reference for the formulation of treatment plans; the drug side effect database can remind doctors to pay attention to the potential risks of drugs during treatment; and environmental medicine research considers the impact of environmental factors on skin disease treatment. By integrating these data, more comprehensive and accurate information support can be provided for treatment recommendations; S72: Graph database implementation, using Neo4j to construct nodes and edges, where nodes represent diseases, drugs, genes, and environmental factors, and edges represent treatment associations, contraindications, and dosage rules. Using Neo4j to construct nodes and edges, where nodes represent diseases, drugs, genes, and environmental factors, and edges represent treatment associations, contraindications, and dosage rules, the complex relationships between various factors are presented in an intuitive graphical manner. This visual representation allows doctors to more clearly understand the connection between diseases and treatments, as well as the mutual influence of different factors, helping them to formulate treatment plans more quickly and accurately. S73: Query and matching logic: After entering the patient's diagnosis, the atlas automatically matches targeted drugs and adjuvant therapies, filtering them based on the patient's contraindications. After entering the patient's diagnosis, the atlas automatically matches targeted drugs and adjuvant therapies, filtering them based on the patient's contraindications, and providing personalized treatment recommendations. Each patient's condition and physical condition are unique. This personalized query and matching logic fully accounts for individual differences, avoiding the use of drugs with contraindications for the patient and improving the safety and effectiveness of treatment. Furthermore, the precise matching of targeted drugs and adjuvant therapies can improve the targeted nature of treatment, enhance treatment effectiveness, and reduce unnecessary drug use and side effects.

[0016] Preferably, the diagnosis and classification module includes a balanced self-paced learning algorithm, which is used to assign higher weights to rare disease samples and combine with a generator to generate synthetic data to expand the training set. The adversarial network uses a conditional generator to generate dermoscopic images with rare disease labels as conditions. Assigning higher weights to rare disease samples can enable the model to pay more attention to the feature learning of rare diseases. In traditional machine learning model training, due to the relatively small number of rare disease samples, it is often difficult for the model to fully learn the characteristics of rare diseases, resulting in low diagnostic accuracy for rare diseases. Assigning higher weights to rare disease samples through a balanced self-paced learning algorithm can guide the model to more deeply explore the characteristic information of rare diseases and improve the ability to identify rare diseases.

[0017] Beneficial effects of the present invention: 1. Compared to existing technologies that rely solely on single-point lesion area image input, which suffers from the inability to eliminate individual differences and environmental factors that lead to increased false positive rates, this solution uses a simultaneous input of the lesion area image and a symmetrical skin image on the symmetrical structure of the human body. Through dual-channel comparative analysis of differences in color, texture, morphology, and other characteristics, it can accurately distinguish physiological changes from pathological abnormalities (for example, a ΔE color difference of more than 3 in symmetrical areas is judged as abnormal pigmentation) and reduces the false positive rate due to environmental lighting interference by 38%; 2. Compared with existing technologies that use a single image input and simple contrast enhancement scheme, which suffer from insufficient image resolution and inability to restore skin microstructure, this scheme uses a high-order degradation model to simulate the dermatoscope imaging process. It combines a conditional generative adversarial network with residual dense blocks to achieve the conversion of ordinary photos to dermatoscope-level super-resolution, with the advantages of eliminating blurring artifacts and restoring micron-level details such as hair follicles and skin textures. 3. Compared to existing technologies that rely on single-modal image feature analysis and suffer from skin color bias, which leads to high misdiagnosis rates, this solution uses an adversarial learning framework and gradient reversal layer technology to force the feature space to be decoupled from irrelevant factors such as skin color and age. This has the advantages of reducing the misdiagnosis rate for dark-skinned patients to 7% and improving cross-racial diagnostic generalization capabilities. 4. Compared with the existing technology that uses static classification models and single threshold judgments, it has the disadvantages of weak rare disease recognition ability and inability to adapt to emerging diseases. This solution introduces a balanced self-paced learning algorithm to give high weights to rare disease samples, and combines it with a federated learning framework to achieve dynamic updates across institutions. It has the advantages of increasing the recognition rate of rare diseases such as hereditary epidermolysis bullosa by 18%, and adapting to emerging diseases such as COVID-19-related rashes through monthly iterations. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Shown is a schematic diagram of the structure of the intelligent auxiliary diagnosis system for skin diseases of the present invention; Figure 2Shown is a schematic diagram of the workflow of the feature extraction module in the intelligent auxiliary diagnosis system for skin diseases of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and examples.

[0020] See also Figure 1-2 The present invention provides an embodiment: an intelligent auxiliary diagnosis system for skin diseases, comprising: Data input module Used to input a user-entered normal skin photo. The specific steps are: Multimodal data acquisition: inputting the user's basic information, medical history, and symptom data; image acquisition: the user uses the image acquisition device to take a close-up photo of the skin and inputs it into the data input module; image quality detection: performing quality detection on the close-up skin photo taken by the user and reminding the user to retake it if the quality test fails; data transmission: encrypting the input data and the close-up skin photo using an encryption algorithm and transmitting the data; When a user takes a close-up photo of the skin through an image acquisition device and inputs the photo into the data input module, the input close-up photo of the skin is two sets, namely, an image of the lesion area and an image of the skin symmetrical to the lesion area on the symmetrical structure of the human body; Image processing module It is used to pre-process the normal skin close-up photo input by the data input module and convert it into a dermatoscope-level super-resolution image. The specific steps are as follows: Image preprocessing uses a high-order degradation model to simulate the real dermatoscope imaging process, and uses bicubic interpolation to denoise and histogram equalization to enhance contrast and resolve the blurring problem of ordinary photos. Super-resolution reconstruction uses residual dense blocks based on conditional generative adversarial networks to generate dermatoscope-level images. Color correction performs color gamut mapping for different skin tones to eliminate the impact of lighting deviation on pigment assessment. Among them, the principle of super-resolution reconstruction is: Input preprocessing: denoises and enhances contrast of close-up photos of normal skin, and normalizes them to a standard size. Feature extraction and mapping: The RRDB network extracts hierarchical features from low-resolution images through multi-level residual connections, and utilizes the feature reuse mechanism within dense blocks to enhance the representation of high-frequency details. For high-resolution image generation, the generator upsamples the intermediate feature maps to the target resolution through a deconvolution layer, and dynamically assigns generation weights to different regions using an attention mechanism to prioritize the recovery of key structures. Post-processing and optimization: Wavelet domain fidelity loss is used to suppress artifacts, optimize detail clarity in high-frequency subbands, and ensure a natural transition of microstructures. Feature extraction module It is used to extract features from the dermatoscope-level super-resolution image obtained after processing by the image processing module. The specific steps are as follows: Lesion area identification uses an image segmentation algorithm to segment the normal skin area and the lesion area; S42: color feature extraction converts the image from RGB to LAB color space and calculates the color difference between the lesion area and the surrounding normal skin. The calculation principle formula is: ;in, is color difference, is brightness, is the red-green axis chromaticity, is the yellow-blue axis chromaticity, The brightness difference between the lesion area and the normal skin. and The red-green axis chromaticity and yellow-blue axis chromaticity differences between the lesion area and the skin of the lesion are respectively calculated; texture feature extraction is carried out by statistically analyzing the grayscale distribution of pixel pairs in the lesion area; morphological feature extraction is carried out by automatically drawing an initial boundary based on the color difference of the lesion area, and through iterative adjustment, the outline is aligned with the actual lesion edge, and then the aspect ratio difference and edge irregularity of the lesion area are calculated; irrelevant differences are eliminated by outputting lesion features through the generator in the adversarial network, and the discriminator forces the feature distribution to be independent of skin color and age, and uses the gradient reversal layer to reverse the gradient direction during back propagation to reduce racial correlation in the feature space; The specific process of lesion area identification is as follows: Normal skin feature extraction: extract features from the user-input image of the non-lesion area, i.e., the skin image of the lesion area that is symmetrical to the human body; lesion skin feature extraction: extract features from the user-input image of the lesion area; feature comparison: compare the normal skin features and the lesion skin features to identify abnormal features of the lesion skin; image segmentation: identify the lesion skin contour based on the identified abnormal features, and perform image segmentation to obtain the lesion area; The diagnosis and classification module uses a balanced self-paced learning algorithm to assign higher weights to rare disease samples and combines it with a generator to generate synthetic data to expand the training set. The adversarial network uses a conditional generator to generate dermatoscope images based on rare disease labels. Diagnosis and classification module It is used to diagnose anomalies based on the features extracted by the feature extraction module and classify anomalies using a classifier. The specific steps include: The first-level classification uses a support vector machine to quantitatively analyze the seven-point melanoma checklist features, including diameter greater than 6mm, irregular margins, uneven color, and dynamic change indicators. The second-level classification uses the EfficientNetV2 model, trained on the HAM10000 dataset, and optimizes computational efficiency through compound scaling. A confidence threshold of 0.85 is set, and if the model output falls below this threshold, a manual review process is triggered. The third-level classification integrates the user's basic information, medical history, and symptom data to achieve accurate classification. Treatment recommendation module It is used to generate and recommend personalized treatment plans based on the diagnosis results of the diagnosis and classification modules using the knowledge graph matching method. The specific steps include: Data integration: integrating disease-drug-gene relationships, clinical guidelines, drug side effect databases, and environmental medicine research; graph database implementation: building nodes and edges through Neo4j, where nodes represent diseases, drugs, genes, and environmental factors, and edges represent treatment associations, contraindications, and dosage rules; query and matching logic: input patient diagnosis results, the graph automatically matches targeted drugs and adjuvant therapies, and filters based on the patient's contraindications.

[0021] Example 1: Early screening for melanoma and personalized treatment recommendations Application scenarios: Melanoma screening and precision treatment for high-risk groups (such as those with a history of ultraviolet exposure or family genetic history).

[0022] Technical implementation: Step 1: Data input and preprocessing The user uploads a close-up photo of normal skin (including a symmetrical comparison image) showing a pigmented patch on the back. The system simulates dermoscopic imaging using a high-order degradation model, generates a super-resolution image (up to 2000×2000 pixels) using residual dense blocks, and eliminates skin color deviations through LAB color space conversion. Color difference calculation: A ΔE value greater than 8 triggers a warning of malignant lesions (for example, a significant increase in the Δa value indicates a red shift).

[0023] Step 2: Feature extraction and diagnosis: The Swin Transformer segmentation model was used to identify lesion boundaries, and the asymmetry index was quantified using the ABCD method (an asymmetry score > 0.75 indicates high risk). A support vector machine (SVM) primary classification system was used: lesions with a combined score of 4 or higher, including diameter > 6 mm, irregular margins (curvature standard deviation > 0.15), and color heterogeneity (LAB color gamut dispersion > 20%), were considered malignant.

[0024] Step 3: Treatment recommendations: Knowledge Graph Matching: If a BRAF gene mutation is detected, targeted drugs (such as the dabrafenib + trametinib combination) are recommended first, and drugs with hepatotoxicity risks are excluded through Neo4j-linked contraindication databases. Dynamic Optimization: Based on patient treatment feedback (e.g., lesion reduction rate <30% / month), the system automatically adjusts the phototherapy dose (20 → 35 mJ / cm²).

[0025] Effect verification: In tests at Xiangya Second Hospital, the accuracy rate for identifying benign and malignant tumors reached 95.8%, and the misdiagnosis rate was reduced by 18% compared with traditional methods.

[0026] Example 2: Intelligent Identification and Classification of Psoriasis and Eczema Application scenario: Rapid identification of erythematous and scaly skin diseases by general practitioners in primary hospitals.

[0027] Technical implementation: Step 1: Image enhancement and feature extraction For the erythema area, the system generates dermatoscope-level images through conditional generative adversarial networks (cGAN) to enhance the details of the scale texture (GLCM contrast > 0.3 indicates the typical "candle wax phenomenon" of psoriasis).

[0028] Step 2: Morphological feature analysis The standard deviation of the curvature of the border of psoriasis plaques is <0.1 (smooth), and the standard deviation of the curvature of the edge of eczema is >0.25 (jagged).

[0029] Step 3: Stratified Diagnosis Secondary classification (EfficientNetV2 model): Trained on the HAM10000 dataset, it achieved a confidence level of 0.92 for psoriasis identification (above the threshold of 0.85) and an accuracy rate of 88.6% for eczema. Tertiary classification: Integrating IL-36RN gene test results, it distinguishes between plaque-type (mutation-positive) and arthritic-type psoriasis (HLA-B27-positive).

[0030] Step 4: Treatment Recommendations Multi-dimensional Matching: Ceramide moisturizers are recommended for patients with FLG gene mutations, and vitamin C antioxidants are added to the regimen when PM2.5 levels are >75. Federated Learning Update: Misdiagnosed cases trigger model iterations, with monthly optimizations (e.g., a 12% increase in the accuracy of identifying COVID-19-related rashes). Grassroots Applications: Tests at the Hangzhou Community Health Service Center showed that the diagnostic accuracy of general practitioners increased from 68% to 87%, and patient satisfaction increased by 23%.

[0031] Example 3: Auxiliary diagnosis and treatment of rare diseases such as hereditary epidermolysis bullosa (EB) Application scenario: Precise management of EB patients by rare disease diagnosis and treatment centers.

[0032] Technical implementation: Step 1: Data enhancement and weight optimization Balanced Self-paced Learning (BSPL): EB samples were weighted 10x, synthetic data (FID < 30) was generated using a conditional GAN, and the training set was expanded to 3,000 cases. Adversarial Network Debiasing: The Gradient Reversal Layer (GRL) reduced the misdiagnosis rate for dark-skinned patients from 22% to 7%.

[0033] Step 2: Feature extraction and classification Morphological analysis: High-frequency ultrasound examination of epidermal separation depth greater than 1.5 mm indicates recessive dystrophic EB (associated with COL7A1 mutations). Multimodal fusion: Combining gene sequencing data (such as KRT5 / KRT14 mutations) with skin CT images can distinguish between simple and junctional EB.

[0034] Step 3: Personalized treatment Knowledge graph-driven: Matching gene-editing therapies (such as CRISPR-Cas9 repair of COL7A1) and linking to a contraindication database to exclude sulfonamides (which may cause blistering). Disease management: The Ruifu AI platform monitors blister reduction and dynamically adjusts the dosage of biologics (such as etanercept). Clinical value: Applications at Peking Union Medical College Hospital demonstrate an EB typing accuracy rate of 98.3%, shortening treatment cycles by 30%.

[0035] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. An intelligent auxiliary diagnosis system for skin diseases; characterized by: include: A data input module, used to input a user-entered normal skin close-up photo; An image processing module, configured to pre-process the normal skin close-up photo input by the data input module and convert it into a dermatoscope-level super-resolution image; A feature extraction module is used to extract features from the dermatoscope-level super-resolution image obtained after processing by the image processing module; The diagnosis and classification module is used to diagnose anomalies based on the features extracted by the feature extraction module and classify anomalies using a classifier; The treatment recommendation module is used to generate and recommend personalized treatment plans based on the diagnosis results of the diagnosis and classification modules using the knowledge graph matching method.

2. The intelligent auxiliary diagnosis system for skin diseases according to claim 1, characterized in that: When the data input module inputs a user's normal skin close-up photo, it specifically includes: S11: Multimodal data collection, entering the user's basic information, medical history and symptom data; S12: Image acquisition, the user takes a close-up photo of the skin through an image acquisition device and inputs it into the data input module; S13: Image quality detection, performing quality detection on the close-up skin image taken by the user, and reminding the user to retake the image if the quality detection fails; S14: Data transmission: encrypt the input data and the skin close-up image through an encryption algorithm and transmit the data.

3. The intelligent auxiliary diagnosis system for skin diseases according to claim 2, characterized in that: When the user takes close-up skin photos through an image acquisition device and inputs them into the data input module, the input close-up skin photos are two groups, namely, images of the lesion area and images of the skin symmetrical to the lesion area on the symmetrical structure of the human body.

4. The intelligent auxiliary diagnosis system for skin diseases according to claim 3, characterized in that: The image processing module pre-processes the normal skin close-up photo input by the data input module and converts it into a dermatoscope-level super-resolution image, specifically including: S21: Image preprocessing, using a high-order degradation model to simulate the real dermatoscope imaging process, and using bicubic interpolation to denoise and histogram equalization to enhance contrast and solve the blur problem of ordinary photos; S22: Super-resolution reconstruction, based on conditional generative adversarial networks, using residual dense blocks to generate dermatoscope-level images; S23: Color correction, performing color gamut mapping for different skin tones to eliminate the impact of lighting deviation on pigment evaluation.

5. The intelligent auxiliary diagnosis system for skin diseases according to claim 4, characterized in that: When performing super-resolution reconstruction, the specific principles and steps include: S31: Input preprocessing: denoising, contrast enhancement, and normalization of normal skin close-up photos to a standard size; S32: Feature Extraction and Mapping, the RRDB network extracts hierarchical features of low-resolution images through multi-level residual connections and uses the feature reuse mechanism within dense blocks to enhance the representation of high-frequency details; S33: High-resolution image generation: The generator upsamples the intermediate feature map to the target resolution through a deconvolution layer, and dynamically allocates generation weights to different regions using an attention mechanism to prioritize the restoration of key structures. S34: Post-processing and optimization, suppressing artifacts through wavelet domain fidelity loss, optimizing detail clarity in high-frequency sub-bands, and ensuring natural transitions of microstructures.

6. The intelligent auxiliary diagnosis system for skin diseases according to claim 5, characterized in that: The feature extraction module performs feature extraction on the dermatoscope-level super-resolution image obtained after processing by the image processing module, specifically including: S41: Identify the lesion area, using image segmentation algorithm to segment the normal skin area and the lesion area; S42: Color feature extraction, converting the image from RGB to LAB color space and calculating the color difference between the lesion area and the surrounding normal skin; S43: Texture feature extraction, statistical grayscale distribution of pixel pairs in the lesion area; S44: Morphological feature extraction: automatically draw an initial boundary based on the color difference of the lesion area, and iteratively adjust the outline to fit the actual lesion edge. Then, calculate the aspect ratio difference and edge irregularity of the lesion area. S45: Eliminate irrelevant differences by outputting lesion features through the generator in the adversarial network, and the discriminator forces the feature distribution to be independent of skin color and age. The gradient reversal layer is used to reverse the gradient direction during back propagation, thereby reducing racial correlation in the feature space.

7. The intelligent auxiliary diagnosis system for skin diseases according to claim 6, characterized in that: When using image segmentation algorithms to segment normal skin areas and lesion areas, the following steps are specifically included: S51: extracting normal skin features, performing feature extraction on the image of the non-lesion area input by the user, i.e., the skin image of the lesion area that is symmetrical to the human body structure; S52: Extracting features of lesion skin, extracting features from the image of the lesion area input by the user; S53: Feature comparison: compare the features of normal skin and diseased skin to identify abnormal features of diseased skin; S54: Image segmentation: Based on the identified abnormal features, the contour of the diseased skin is identified and the image is segmented to obtain the diseased area.

8. The intelligent auxiliary diagnosis system for skin diseases according to claim 7, characterized in that: The diagnosis and classification module performs abnormal diagnosis based on the features extracted by the feature extraction module and classifies abnormalities using a classifier, specifically including: S61: First-level classification, quantitative analysis of the 7-point checklist features of melanoma based on support vector machine, including diameter >6 mm, irregular margins, uneven color, and dynamic change indicators; S62: Secondary classification, using the EfficientNetV2 model, trained on the HAM10000 dataset, with compound scaling to optimize computational efficiency. A confidence threshold of 0.85 is set, and manual review is triggered if the model output falls below this threshold. S63: Three-level classification, integrating the user's basic information, medical history and symptom data to achieve accurate classification.

9. The intelligent auxiliary diagnosis system for skin diseases according to claim 8, characterized in that: The treatment recommendation module generates and recommends personalized treatment plans based on the diagnosis results of the diagnosis and classification modules using the knowledge graph matching method. Specifically, it includes: S71: Data integration, integrating disease-drug-gene relationships, clinical guidelines, drug side effect databases, and environmental medicine research; S72: Graph database implementation, using Neo4j to build nodes and edges, where nodes represent diseases, drugs, genes, and environmental factors, and edges represent therapeutic associations, contraindications, and dosage rules; S73: Query and matching logic: input the patient’s diagnosis results, the map automatically matches targeted drugs and adjuvant treatments, and filters according to the patient’s contraindications.

10. The intelligent auxiliary diagnosis system for skin diseases according to claim 9, characterized in that: The diagnosis and classification module includes a balanced self-paced learning algorithm, which is used to give higher weights to rare disease samples, and is combined with a generator to generate synthetic data to expand the training set. Among them, the adversarial network uses a conditional generator to generate dermatoscope images based on the rare disease label.

Images