Intelligent skin disease identification system and screening method based on image features

Abstract

The invention discloses an intelligent skin disease recognition system and screening method based on image features, and relates to the technical field of artificial intelligence. Skin disease image data are collected and uploaded to the labeling system, and label labeling is completed through assistance of professional doctors. The constructed skin disease data set covers dermatoscope type data and clinical shot image data. Performing data enhancement, standardization, normalization and image size unified adjustment on the labeled data, and inputting the data into a deep learning model for training; a deep neural network DCSNeXt model suitable for skin disease image feature recognition is designed and constructed, a special classifier is designed according to data categories, and model training and performance testing are carried out on the two types of data; a plurality of existing common image recognition models are further used for carrying out training testing on the data set, and the performance of the data set is compared with the performance of a DCSNeXt model on a test set; in addition, two types of skin disease image data in related literatures in recent two years are collected, and a model with good performance is selected to carry out label-free prediction analysis on the skin disease image data. According to the invention, intelligent characterization analysis of dermatoscope and clinically shot dermatosis images can be realized, technical support is provided for feature extraction and classification of dermatosis images, the method can be widely applied to the fields of dermatology department clinical auxiliary analysis, telemedicine and medical teaching, and development of dermatosis intelligent analysis is promoted.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent skin disease recognition system and screening method based on image features. Background Art

[0002] As a widespread health problem, skin diseases are an important research subject in the field of dermatology because of their complex and diverse pathological characteristics and manifestations. In the course of the disease, skin diseases often show their characteristics through surface morphological changes and structural abnormalities, which provides a basis for diagnosis and treatment. However, the manifestations of different types of skin diseases may be highly similar, and some early symptoms are subtle and difficult to detect, which poses a huge challenge to the accurate identification of the disease. Especially in cases involving complex pathologies such as pigmented lesions, inflammatory skin lesions, and tumorous lesions, the traditional judgment method that relies on the doctor's experience is often affected by subjectivity and is difficult to deal with highly complex and diverse diseases. At present, the image data widely used in skin disease research mainly include two types of data: ordinary clinical shooting type data and dermoscopic type data. Ordinary clinical shooting type data records the macroscopic features of the skin surface and is suitable for preliminary screening of lesions, while dermoscopic type data uses special optical equipment to capture the microscopic structure of deep skin tissues, such as pigment distribution and vascular morphology. Although these data provide multidimensional information for diagnosis, the diversity of images and differences in details make manual interpretation vulnerable to visual fatigue and lack of experience. At the same time, the accumulation of large-scale image data has further increased the workload of analysis and put forward higher efficiency requirements for researchers and clinicians. With the development of artificial intelligence technology, the application of algorithms such as deep learning in medical image analysis has made significant progress, providing a new direction for the intelligent identification of skin diseases. Through learning and modeling of massive image data, AI models can efficiently extract key features and achieve accurate detection and classification of lesion areas. However, most existing studies focus on the analysis of a single image type, and the comprehensive utilization of different types of image data remains to be explored. How to integrate the characteristics of clinical shooting type data and dermoscopic type data, give full play to the complementary advantages of the two, and combine innovative deep learning model design is the key to achieving comprehensive and accurate judgment.

[0003] In view of this, the present invention proposes an intelligent skin disease recognition system and screening method based on image features. Summary of the invention

[0004] The purpose of the present invention is to propose an intelligent skin disease recognition system and screening method based on image features to solve the problems raised in the background technology. The present invention solves the problems of slow speed, low judgment accuracy and poor adaptability to multiple types of images in the existing skin disease image analysis process, further improves diagnostic efficiency and accuracy, and at the same time reduces the workload of doctors in large-scale image analysis.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] An intelligent skin disease recognition system based on image features, comprising:

[0007] Dermatology image dataset module, including dermoscopy type data and clinical photography type data;

[0008] The data augmentation module is used to expand the categories with small data volume in the skin disease image dataset by using random rotation, random flipping, brightness contrast adjustment, cutout and mixup data augmentation methods;

[0009] The DCSNeXt recognition network model is composed of 39 vertically stacked CSMRNext Block feature extraction modules combined with a classifier; the CSMRNext Block feature extraction module includes a 7×7 deep convolution module, a LayerNorm normalization module, a multi-scale fusion spatial attention (MAS) module, and a multi-layer perceptron (MLP) with GlobalResponseNorm.

[0010] Preferably, the dermoscopy type data is derived from the public ISIC database; the clinical photography type data is obtained through DermNet, Dermatology ATLAS database and related literature.

[0011] Preferably, the dermoscopic type data is high-resolution image data collected by professional dermoscopic equipment, which is used to observe the microscopic features of the skin surface and its underlying layer; the clinical photography type data is an image of skin lesions visible to the naked eye acquired by conventional photography methods, which is used to record the macroscopic features of the skin surface, including the size, shape, distribution, and color changes of the lesions.

[0012] Preferably, the multi-scale fusion spatial attention (MAS) module fuses global and local features through residual connections to improve the effectiveness of image feature extraction.

[0013] A skin disease screening method based on an image feature-based skin disease intelligent recognition system includes the following contents:

[0014] S1. Collect a data set of skin disease images, including dermoscopic type data and clinical photography type data; upload the dermoscopic type data and clinical photography type data in the data set to the annotation system developed by the laboratory based on PHP technology, and perform image annotation and category labeling with the assistance of professional doctors;

[0015] S2. For the categories with less data in the skin disease image dataset in S1, random rotation, random flipping, brightness contrast adjustment, Cutout and Mixup data enhancement methods are used to expand them; the images in the expanded dataset are standardized and normalized, and the processed dataset is divided into a training set and a test set;

[0016] S3. Construct a DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism, and use the image data in the training set to train the constructed model; introduce the AutoAugment automatic enhancement strategy in the training process to improve the robustness of the model, and propose the BHFLoss function to optimize the learning ability of the model;

[0017] S4. Use the test set data to test the constructed DCSNeXt recognition network model and the publicly available image recognition models of the same type, compare the performance of each model, and verify the effectiveness of the DCSNeXt recognition network model;

[0018] S5. Collect additional test set image data, perform unlabeled prediction on the constructed DCSNeXt recognition network model and publicly available image recognition models of the same type, and analyze the actual prediction performance and generalization ability of each model.

[0019] Preferably, the DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism is composed of 39 vertically stacked CSMRNeXt Block feature extraction modules combined with a classifier; the execution principle of the CSMRNeXt Block feature extraction module includes the following contents:

[0020] Use a 7×7 deep convolution module to perform deep convolution operations on the input feature map, acting only on the channel dimension to preserve spatial features and improve computing performance;

[0021] Use the LayerNorm normalization module to perform LayerNorm normalization on the feature map to stabilize the training process and improve model performance;

[0022] Combined with the Multi-Scale Fusion Spatial Attention (MAS) module to enhance the feature extraction of the spatial dimension, 3×3, 5×5 and 7×7 convolutions are used to extract multi-scale features, and the output is limited to [0,1] as the spatial attention weight. Combined with the residual connection, it promotes the transmission of feature information flow and enhances the model's attention to the target area.

[0023] The global dependencies between channels are constructed based on a multi-layer perceptron (MLP) consisting of two fully connected layers, activation function GELU, GlobalResponseNorm and Dropout.

[0024] Preferably, the DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism includes four parts: the first part is composed of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 128; the second part is composed of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 256; the third part is composed of 27 CSMRNeXt Block feature extraction modules with a channel dimension of 512; the fourth part is composed of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 1024;

[0025] When the model was trained using the dermoscopy type data training set, the output of the classifier was set to 10;

[0026] When training the model using the clinical imaging type data training section, the output of the classifier is set to 22.

[0027] Preferably, it is characterized in that the formula of the BHFLoss function in S3 is expressed as:

[0028]

[0029] L FL =α c ·ω c (1-p) γ ·L CE

[0030]

[0031] Among them, y i represents the true label, C represents the number of categories, ε represents the smoothing parameter, represents the smoothed label; L CE represents the cross entropy loss, p i represents the predicted probability; L FL represents the focal loss, α c represents the sample size scaling factor, ω c represents the category weight, γ represents the focus adjustment parameter; L harmonic represents the harmonic loss, represents the cosine similarity of smoothed labels; L total Indicates the final total loss; It represents the expected operation on the loss function value, that is, weighted average.

[0032] A computer device, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the above-mentioned skin disease screening method based on image features.

[0033] Compared with the prior art, the present invention provides an intelligent skin disease recognition system and screening method based on image features, which has the following beneficial effects:

[0034] The present invention proposes an intelligent skin disease recognition system and screening method based on image features, which can realize the characteristic intelligent analysis of skin disease images taken by dermatoscope and clinical examination, and provide technical support for the feature extraction and classification of skin disease images. In addition, the present invention proves through specific comparative experiments in the embodiments that the skin disease intelligent recognition system based on image features and the DCSNeXt recognition network model proposed in this application have a substantial actual prediction effect and generalization ability that are significantly superior to the existing similar image recognition models on the market, and can be widely used in the fields of dermatology clinical auxiliary analysis, telemedicine and medical teaching, promoting the development of intelligent analysis of skin diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of a skin disease screening method of a skin disease intelligent recognition system based on image features mentioned in an embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of dermoscopy type data and clinical photography type data mentioned in an embodiment of the present invention;

[0037] Figure 3 It is a statistical diagram of dermoscopy type data and clinical photography type data mentioned in the embodiment of the present invention;

[0038] Figure 4 Schematic diagram of a data augmented image mentioned in an embodiment of the present invention;

[0039] Figure 5 It is a structural diagram of the DCSNeXt recognition network model mentioned in an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the confusion matrix of the DCSNeXt recognition network model mentioned in an embodiment of the present invention in a test set of dermatoscope type data and clinical photography type data. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0042] Example:

[0043] The present invention proposes an intelligent skin disease recognition system based on image features, comprising:

[0044] Dermatology image dataset module, including dermoscopy type data and clinical photography type data;

[0045] The data augmentation module is used to expand the categories with small data volume in the skin disease image dataset by using random rotation, random flipping, brightness contrast adjustment, cutout and mixup data augmentation methods;

[0046] The DCSNeXt recognition network model is composed of 39 vertically stacked CSMRNext Block feature extraction modules combined with a classifier; the CSMRNext Block feature extraction module includes a 7×7 deep convolution module, a LayerNorm normalization module, a multi-scale fusion spatial attention (MAS) module, and a multi-layer perceptron (MLP) with GlobalResponseNorm.

[0047] Based on the above system, the present invention further proposes a skin disease screening method based on an image feature-based skin disease intelligent recognition system. Figure 1 , including the following steps:

[0048] S100, collect skin disease image dataset, see Figure 2-3 , which shows examples and quantity statistics of specific skin disease categories. Common dermoscopic type data includes lesion categories such as actinic keratosis (AK), basal cell carcinoma (BCC), dermatofibroma (DF), lichen planus keratosis (LK), melanoma (MEL), melanocytic nevus (NV), squamous cell carcinoma (SCC), seborrheic keratosis (SK), solar lentigo (SL) and vascular lesions (VASC), with a total of 37,334 images. The types of lesions included in the clinical photography type data are actinic keratosis, basal cell carcinoma, dermatofibroma, melanoma, melanocytic nevus, squamous cell carcinoma, acne, atopic dermatitis (AD), varicella (Cpox), lichen, lupus, psoriasis (PSO), pyoderma, ringworm, rosacea, seborrheic keratosis, vascular lesions, vitiligo, molluscum, scar, monkeypox and measles, with a total of 13,340 images.

[0049] The dermatology image dataset includes dermatoscope type data and clinical photography type data. Dermatoscope type data is high-resolution image data collected by professional dermatoscope equipment, which is mainly used to observe the microscopic features of skin lesions. Dermatoscope equipment can magnify the details of the skin, provide clearer skin lesion features, and help identify subtle skin structure changes. Clinical photography type data is skin lesion images visible to the naked eye obtained by conventional photography methods, which mainly record the macroscopic features of the skin surface. Clinical photography type data usually shows the size, shape, distribution, color changes, etc. of skin lesions, which is an intuitive record of the health status of the skin surface and is suitable for monitoring and research of various skin problems.

[0050] The dermoscopic type data mainly comes from the public ISIC database, and the clinical photography type data is obtained through the DermNet, Dermatology ATLAS database and related literature; the dermoscopic type data and clinical photography type data in the dataset are uploaded to the annotation system developed by the laboratory based on PHP technology, and image annotation and category labeling are performed with the assistance of professional doctors;

[0051] S102, for categories with less data in the skin disease image dataset in S100, please refer to Figure 4 , random rotation, random flipping, brightness and contrast adjustment, Cutout and Mixup data enhancement methods are used for expansion. Random rotation and flipping are operations to rotate the image at random angles or flip it horizontally and vertically. Brightness and contrast adjustment is to enhance the diversity of the image by changing the brightness and contrast of the image. Cutout is to enhance the robustness of the model by randomly blocking a certain area in the image. The blocked area is usually a rectangle of random size. The model needs to ignore the blocked area during training to learn the features of other parts. Mixup generates a new image by linearly combining two images in a certain proportion.

[0052] Standardize and normalize the images in the expanded dataset, and divide the processed dataset into training set, validation set and test set;

[0053] S104. Construct a DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism, and use the image data in the training set to train the constructed model; introduce the AutoAugment automatic enhancement strategy in the training process to improve the robustness of the model. The AutoAugmentt automatic enhancement strategy aims to automatically select the most suitable enhancement strategy for the current data set through the reinforcement learning algorithm. By defining a search space, it can automatically select and combine multiple data enhancement strategies to obtain the best training effect;

[0054] See also Figure 5 The DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism is composed of 39 vertically stacked CSMRNeXt Block feature extraction modules combined with a classifier; the execution principle of the CSMRNeXt Block feature extraction module includes the following contents:

[0055] Use a 7×7 deep convolution module to perform deep convolution operations on the input feature map, acting only on the channel dimension to preserve spatial features and improve computing performance;

[0056] Use the LayerNorm normalization module to perform LayerNorm normalization on the feature map to stabilize the training process and improve model performance;

[0057] Combined with the multi-scale fusion spatial attention module to enhance the feature extraction of the spatial dimension, 3×3, 5×5 and 7×7 convolutions are used to extract multi-scale features, and the output is limited to [0,1] as the spatial attention weight. Combined with the residual connection, it promotes the transmission of feature information flow and enhances the model's attention to the target area;

[0058] Based on a multi-layer perceptron consisting of two fully connected layers, activation function GELU, GlobalResponseNorm and Dropout, the global dependency between channels is constructed;

[0059] The DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism includes four parts: the first part consists of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 128; the second part consists of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 256; the third part consists of 27 CSMRNeXt Block feature extraction modules with a channel dimension of 512; the fourth part consists of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 1024;

[0060] When the model was trained using the dermoscopy type data training set, the output of the classifier was set to 10;

[0061] When training the model using the clinical shooting type data training section, the output of the classifier is set to 22;

[0062] The BHFLoss function is further proposed to optimize the learning ability of the model. The BHFLoss function combines multiple mechanisms such as label smoothing, focal loss, category balance, and harmonic loss, aiming to effectively solve the problems of category imbalance and model overfitting. Through label smoothing, the model's excessive dependence on a single category is reduced, while avoiding overconfident predictions; through the adjustment of focal loss, the weight of easy-to-classify samples is reduced, and the focus is on difficult-to-classify samples; the category weight and sample number scaling factor are combined to balance the loss contribution of different categories and alleviate the problem of uneven category distribution; the cosine similarity is introduced to calculate the harmonic loss, which further enhances the model's sensitivity to feature discrimination from a directional perspective. The function also supports adaptive parameter adjustment, so that the loss function can dynamically optimize its parameters to better adapt to different data distributions and task requirements. The BHFLoss function is expressed as follows:

[0063]

[0064] L FL =α c ·ω c (1-p) γ ·L CE

[0065]

[0066] Among them, y i represents the true label, C represents the number of categories, ε represents the smoothing parameter, represents the smoothed label; L CE represents the cross entropy loss, p i represents the predicted probability; L FL represents the focal loss, α c represents the sample size scaling factor, ω c represents the category weight, γ represents the focus adjustment parameter; L harmonic represents the harmonic loss, represents the cosine similarity of smoothed labels; L total Indicates the final total loss; It represents the expected operation on the loss function value, that is, weighted average.

[0067] S106. Using the test set data, the constructed DCSNeXt recognition network model and the publicly available image recognition models of the same type are tested, and the performance of each model is compared to verify the effectiveness of the DCSNeXt recognition network model.

[0068] The image recognition models used for comparison include VGG-19, AlexNet, DenseNet-201, MobileNet-V3, EfficientNet-V2, Resnet-101, RegNet, Xception, NASNet, EfficientViT-M5, ConViT-Base, CoatNet, CrossViT, FastViT, MobileViT-S, ViT-Base16, RepViT-M2 and BEiTV2-Base. These models are used to train the dermatoscope and clinical shooting type data datasets, and the test set performance is compared with the proposed DCSNeXt recognition network model. The comparison indicators include test set accuracy, F1 score (F1-score) and mean area under the curve (mAUC). The result values ​​are used to represent the accuracy values. The dermatoscope and clinical shooting type data datasets are divided into training set, validation set and test set in a ratio of 7:1.5:1.5. The parameters initialized for model training are as follows: the image size of the input model is unified as 224×224; model inference is performed through GPU, and the model weights are updated through back propagation; the initial learning rate is 0.0001, and the learning rate adjustment strategy uses ReduceLROnPlateau; the loss function used is the designed BHFLoss; the optimizer used is Adam; the number of model training rounds is 100; the specific result values ​​are shown in Table 1.

[0069] Table 1

[0070]

[0071] According to Table 1, the DCSNeXt recognition network model has the best accuracy of 0.934, F1-score of 0.886, and mAUC of 0.992 in the dermatoscope dataset, and the best accuracy of 0.903, F1-score of 0.901, and mAUC of 0.990 in the clinical dataset. Its overall performance is better than other models in the table. Figure 6 The confusion matrix of the target model DCSNeXt in the dermatoscope and clinical data test sets was visualized, and the model's recognition and judgment of each category can be clearly seen.

[0072] S108. An additional 1951 pieces of dermoscopic type data and 1391 pieces of clinical photography type data were collected from the literature in the past two years. Among them, there are 7 types of dermoscopic type data, including actinic keratosis, basal cell carcinoma, dermatofibroma, melanoma, melanocytic nevus, squamous cell carcinoma and vascular lesions, with the numbers being 110, 169, 50, 399, 1090, 81 and 52 respectively. There are 12 types of clinical photography type data, including actinic keratosis, basal cell carcinoma, dermatofibroma, melanoma, melanocytic nevus, squamous cell carcinoma, acne, atopic dermatitis, lichen rash, seborrheic keratosis, vascular lesions and vitiligo, with the numbers being 155, 170, 47, 292, 300, 110, 54, 60, 66, 62, 35 and 40 respectively. The dermoscopic and clinical weights of the previously trained DCSNeXt recognition network model were used to perform unlabeled predictions on the two additional types of data collected above, and the category with the highest model prediction probability was selected as the model's prediction result. The prediction accuracy of each category was used as the only evaluation and comparison indicator. Specifically, the results of dermoscopic prediction are shown in Table 2.

[0073] Table 2

[0074]

[0075] According to Table 2, the prediction accuracy of the DCSNeXt recognition network model for each of the above dermatoscopes is 0.627, 0.740, 0.940, 0.805, 0.964, 0.704 and 0.827 respectively, and its average accuracy is 0.878, which is better than other models with better performance in the test set.

[0076] Then, we selected the model with better performance in the test set among the previous comparison models, and performed unlabeled prediction on these data in the same way. We calculated specific indicators based on the prediction results of these models and then compared the models to further prove the predictive generalization ability of the model. The results of clinical prediction are shown in Table 3.

[0077] Table 3

[0078]

[0079]

[0080] According to Table 3, the prediction accuracy of the DCSNeXt recognition network model for each category of the above clinical data is 0.750, 0.710, 0.944, 0.871, 0.894, 0.742, 0.911, 0.860, 0.782, 0.839, 0.829 and 1.000, respectively, and its average accuracy is 0.845, which is better than other models with better performance in the test set in most categories, and the average accuracy achieves the optimal level.

[0081] In summary, the present invention proposes an intelligent skin disease recognition system and screening method based on image features, which can realize the characteristic intelligent analysis of skin disease images taken by dermatoscope and clinical photography, and provide technical support for the feature extraction and classification of skin disease images. In addition, the present invention proves through specific comparative experiments in the embodiments that the skin disease intelligent recognition system based on image features and the DCSNeXt recognition network model proposed in this application have greatly improved the actual prediction effect and generalization ability compared with the existing similar image recognition models on the market, and can be widely used in the fields of clinical auxiliary analysis of dermatology, telemedicine and medical teaching, which promotes the development of intelligent analysis of skin diseases.

[0082] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent skin disease recognition system based on image features, characterized in that: include: Dermatology image dataset module, including dermoscopy type data and clinical photography type data; The data augmentation module is used to expand the categories with small data volume in the skin disease image dataset by using random rotation, random flipping, brightness contrast adjustment, cutout and mixup data augmentation methods; The DCSNeXt recognition network model is composed of 39 vertically stacked CSMRNext Block feature extraction modules combined with a classifier; the CSMRNext Block feature extraction module includes a 7×7 deep convolution module, a LayerNorm normalization module, a multi-scale fusion spatial attention module and a multi-layer perceptron with GlobalResponseNorm.

2. The system according to claim 1, characterized in that The dermoscopy type data is derived from the public ISIC database; the clinical photography type data is obtained through DermNet, Dermatology ATLAS database and related literature.

3. The system according to claim 2, characterized in that The dermoscopic type data is high-resolution image data collected by professional dermoscopic equipment, which is used to observe the microscopic features of the skin surface and its underlying layer; the clinical photography type data is images of skin lesions visible to the naked eye acquired by conventional photography methods, which is used to record the macroscopic features of the skin surface, including the size, shape, distribution, and color changes of the lesions.

4. The system according to claim 1, characterized in that The multi-scale fusion spatial attention module fuses global and local features through residual connections to improve the effectiveness of image feature extraction.

5. The skin disease screening method implemented by the system according to any one of claims 1 to 4, characterized in that: Includes the following: S1. Collect a data set of skin disease images, including dermoscopic type data and clinical photography type data; upload the dermoscopic type data and clinical photography type data in the data set to the annotation system developed by the laboratory based on PHP technology, and perform image annotation and category labeling with the assistance of professional doctors; S2. For the categories with less data in the skin disease image dataset in S1, random rotation, random flipping, brightness contrast adjustment, Cutout and Mixup data enhancement methods are used to expand them; the images in the expanded dataset are standardized and normalized, and the processed dataset is divided into a training set and a test set; S3. Construct a DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism, and use the image data in the training set to train the constructed model; introduce the AutoAugment automatic enhancement strategy in the training process to improve the robustness of the model, and propose the BHFLoss function to optimize the learning ability of the model; S4. Use the test set data to test the constructed DCSNeXt recognition network model and the publicly available image recognition models of the same type, compare the performance of each model, and verify the effectiveness of the DCSNeXt recognition network model; S5. Collect additional test set image data, perform unlabeled prediction on the constructed DCSNeXt recognition network model and publicly available image recognition models of the same type, and analyze the actual prediction performance and generalization ability of each model.

6. The method according to claim 5, characterized in that The DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism is composed of 39 vertically stacked CSMRNeXt Block feature extraction modules combined with a classifier; the execution principle of the CSMRNeXt Block feature extraction module includes the following contents: Use a 7×7 deep convolution module to perform deep convolution operations on the input feature map, acting only on the channel dimension to preserve spatial features and improve computing performance; Use the LayerNorm normalization module to perform LayerNorm normalization on the feature map to stabilize the training process and improve model performance; Combined with the multi-scale fusion spatial attention module to enhance the feature extraction of the spatial dimension, 3×3, 5×5 and 7×7 convolutions are used to extract multi-scale features, and the output is limited to [0,1] as the spatial attention weight. Combined with the residual connection, it promotes the transmission of feature information flow and enhances the model's attention to the target area; The global dependency between channels is constructed based on a multi-layer perceptron consisting of two fully connected layers, activation function GELU, GlobalResponseNorm and Dropout.

7. The method according to claim 6, characterized in that The DCSNeXt recognition network model based on deep convolution and multi-scale attention mechanism includes four parts: the first part consists of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 128; the second part consists of 4 CSMRNeXt Block feature extraction modules with a channel dimension of 256; the third part consists of 27 CSMRNeXt Block feature extraction modules with a channel dimension of 512; the fourth part consists of 4 CSMRNeXtBlock feature extraction modules with a channel dimension of 1024; When the model was trained using the dermoscopy type data training set, the output of the classifier was set to 10; When training the model using the clinical imaging type data training section, the output of the classifier is set to 22.

8. The method according to claim 5, characterized in that The formula of the BHFLoss function described in S3 is expressed as: L FL =a c ·oh c ·(1-p) γ ·L CE Among them, y i represents the true label, C represents the number of categories, ε represents the smoothing parameter, represents the smoothed label; L CE represents the cross entropy loss, p i represents the predicted probability; L FL represents the focal loss, α c represents the sample size scaling factor, ω c represents the category weight, γ represents the focus adjustment parameter; L harmonic represents the harmonic loss, represents the cosine similarity of smoothed labels; L total Indicates the final total loss; It represents the expected operation on the loss function value, that is, weighted average.

9. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the skin disease screening method of the image feature-based skin disease intelligent recognition system as described in any one of claims 5-8.

Images