Children skin disease multi-classification system based on improved ResNet

Abstract

The invention discloses a children skin disease multi-classification system based on improved ResNet. The children skin disease multi-classification system comprises a computer memory and a classification model stored in the computer memory. The classification model comprises a feature extraction unit and a prediction unit; the feature extraction unit comprises a residual module structure, a mesoscale expansion convolution structure, a wide scale expansion convolution structure and a multi-scale fusion dilated convolution structure; an image sequentially passes through a convolution layer, a maximum pooling layer, a convolution layer and a batch regularization layer and then enters the four structures which are used in parallel from feature extraction, and respective feature maps are output; the prediction unit fuses feature maps output by each structure and inputs the fused feature maps into a batch regularization layer, and the fused features are standardized; carrying out down-sampling on the standardized feature map through an average pooling layer; and finally, converting the features into classification probability distribution of skin diseases through a softmax layer. According to the invention, the recognition and classification accuracy of common skin diseases of children can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical artificial intelligence, and in particular relates to a pediatric skin disease multi-classification system based on improved ResNet. Background Art

[0002] Common childhood skin diseases (atopic dermatitis, urticaria, hemangioma, diaper dermatitis, etc.) seriously affect the quality of life of children, and can easily lead to limited daily activities, sleep disorders, poor school performance, slow growth and development, social disorders, anxiety, depression, etc. Their family members will suffer from lack of sleep, fatigue, anxiety, depression, and impact on work and life due to taking care of the children and constantly seeking medical treatment. The children and their families suffer serious distress, which endangers their physical and mental health and even their lives, causing heavy economic and social burdens.

[0003] Dermatology is a highly intuitive morphological science, and its diagnosis and treatment monitoring largely rely on the morphology of various skin diseases. Therefore, the traditional diagnosis of skin diseases is mainly based on the dermatologist's judgment of the disease based on the patient's medical history, clinical manifestations, and image information such as the lesion morphology, size, and color. However, this method has certain subjectivity and errors. Therefore, the research in the field of skin disease image classification aims to use computer vision and image processing technology to automatically classify and diagnose skin disease images, thereby improving the diagnostic efficiency and accuracy of skin diseases.

[0004] For example, Chinese patent document with publication number CN117333710A discloses a skin disease image classification method based on deep learning; Chinese patent document with publication number CN111444960A discloses a skin disease image classification system based on multimodal data input.

[0005] Although some progress has been made in the field of skin disease image classification, there are still many challenges. For example, the diversity and complexity of skin disease images, the similarities between different skin diseases, and the impact of image quality all make it difficult to classify and diagnose skin disease images. Summary of the invention

[0006] The present invention provides a pediatric skin disease multi-classification system based on improved ResNet, which can realize automatic classification of pediatric skin disease images and improve the recognition and classification accuracy of common pediatric skin diseases.

[0007] A pediatric skin disease multi-classification system based on an improved ResNet, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein a trained classification model is stored in the computer memory;

[0008] The classification model adopts an improved ResNet network model, which includes a feature extraction unit and a prediction unit;

[0009] The feature extraction unit comprises a residual module structure, a medium-scale dilated convolution structure, a wide-scale dilated convolution structure and a multi-scale fusion hole convolution structure; the image is sequentially passed through a convolution layer, a maximum pooling layer, a convolution layer and a batch normalization layer, and features are extracted from the image, and then enter the above four structures, which are used in parallel to output their own feature maps;

[0010] In the prediction unit, the feature maps output by each structure are fused and input into a batch normalization layer, and the fused features are standardized by normalization technology; then the standardized feature maps are downsampled by an average pooling layer, and the feature maps are compressed into a single vector, reducing the spatial dimension of the feature maps while retaining the most important feature information; finally, the features are converted into classification probability distributions of skin diseases by a softmax layer;

[0011] The computer processor implements the following steps when executing the computer program: inputting the common children's skin disease images to be classified into the trained classification model to obtain the skin disease classification results.

[0012] Furthermore, the training process of the classification model is as follows:

[0013] Collect images of affected areas of children with atopic dermatitis, urticaria, hemangioma, and diaper dermatitis and skin images of normal children, and preprocess the image data;

[0014] 70% of the data are randomly selected as the training set, 10% of the data are randomly selected as the validation set, and the remaining 20% ​​of the data are used as the test set; the training set is used for model training; the validation set is used for model selection and hyperparameter adjustment; and the test set is used to finally evaluate the model performance.

[0015] The image data is preprocessed, including the following steps:

[0016] Crop the skin image to reduce irrelevant areas; Image normalization: Normalize the image by mean and standard deviation to ensure that all images have the same brightness and contrast levels; Image enhancement: Use histogram equalization and color space conversion technology to enhance image details; Size normalization: Adjust all input images to a uniform size.

[0017] The residual module structure comprises three residual modules, two of which are connected in series and used in parallel with another residual module;

[0018] Each residual module contains two parallel branches, one branch contains a 3×3 convolutional layer, a batch normalization layer, a 3×3 convolutional layer and a batch normalization layer connected sequentially, and the other branch is a 1×1 convolutional layer; the features obtained by the two branches are added and then passed through a batch normalization layer to obtain the output feature map.

[0019] The mesoscale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution, and a batch normalization layer connected in sequence; wherein the dilation rate of the 3×3 dilated convolution is 2, and the mesoscale dilated convolution structure is used to simulate the receptive field of a 7×7 convolution kernel.

[0020] The wide-scale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, and a batch normalization layer connected in sequence; wherein the dilation rate of the 3×3 dilated convolution layer is 2, and the wide-scale dilated convolution structure is used to simulate the receptive field of a 15×15 convolution kernel.

[0021] The multi-scale fusion dilated convolution structure includes two parallel branches, one branch includes three dilated convolutions connected in sequence, each dilated convolution uses a different dilation rate to expand the receptive field, and after the input feature map is subjected to three dilated convolution operations in sequence, the intermediate feature maps corresponding to the first two convolutions and the feature map of the last convolution are concatenated; the other branch is a 1×1 convolution; the outputs of the two branches are added with features and then subjected to a 1×1 convolution to obtain the output feature map.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The classification model in the present invention adopts an improved Resnet network model, which uses three different methods composed of residual module, medium-scale dilated convolution, wide-scale dilated convolution and multi-scale fused dilated convolution to simulate various convolution kernel sizes for feature extraction. This means that the network can decide how to balance and compensate between different ways of simulating convolution kernels, so that the model can effectively learn feature information at different scales.

[0024] 2. The present invention improves the accuracy of skin disease recognition and classification by learning features from a large number of skin disease images, provides doctors with auxiliary diagnosis suggestions, shortens diagnosis time, helps primary medical institutions improve their diagnostic capabilities, and enables more children to receive timely and accurate medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is an implementation flow chart of a pediatric skin disease multi-classification system based on an improved ResNet according to an embodiment of the present invention;

[0026] Figure 2It is a schematic diagram of the overall structure of the classification model in an embodiment of the present invention;

[0027] Figure 3 Schematic diagram of the residual module structure in an embodiment of the present invention;

[0028] Figure 4 Schematic diagram of the mesoscale dilated convolution structure in an embodiment of the present invention;

[0029] Figure 5 Schematic diagram of a wide-scale dilated convolution structure in an embodiment of the present invention;

[0030] Figure 6 Schematic diagram of the multi-scale fusion dilated convolution structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0032] A classification system for common children's skin diseases comprises a computer memory, a computer processor and a computer program stored in the computer memory and executable on the computer processor. A trained classification model is stored in the computer memory.

[0033] like Figure 1 As shown, the implementation process of the entire system is as follows:

[0034] 1. Data Collection

[0035] Collect images of affected areas of children with atopic dermatitis, urticaria, hemangioma, diaper dermatitis and skin images of normal children, and pre-process the image data. All images should be annotated by professional dermatologists to ensure the accuracy and consistency of the data.

[0036] 2. Image preprocessing

[0037] The preprocessing of the image includes the following steps: cropping the skin image to reduce irrelevant areas, focusing the model on key areas, and improving the efficiency and performance of the model; image standardization: normalizing the image by mean and standard deviation to ensure that all images have the same brightness and contrast levels; image enhancement: using histogram equalization, color space conversion and other techniques to enhance image details; size standardization: adjusting all input images to a uniform size for easy model processing.

[0038] 3. Data Grouping

[0039] 70% of the data are randomly selected as the training set, 10% of the data are randomly selected as the validation set, and the remaining 20% ​​of the data are used as the test set. The training set is used for model training; the validation set is used for model selection and hyperparameter adjustment; and the test set is used to finally evaluate the model performance.

[0040] 4. Model construction

[0041] In the present invention, the classification model adopts an improved ResNet network model, which includes a feature extraction unit and a prediction unit.

[0042] The feature extraction unit includes a residual module structure, a medium-scale dilated convolution structure, a wide-scale dilated convolution structure, and a multi-scale fusion hole convolution structure. The image passes through a convolution layer, a maximum pooling layer, a convolution layer, and a batch normalization layer in sequence to extract features from the image and enter the above four structures. The four structures are used in parallel to output their own feature maps.

[0043] In the prediction unit, the feature maps output by each structure are fused and input into a batch normalization layer, and the fused features are standardized through normalization technology; then the standardized feature maps are downsampled through the average pooling layer, and the feature maps are compressed into a single vector, reducing the spatial dimension of the feature maps while retaining the most important feature information; finally, the features are converted into the classification probability distribution of skin diseases through the softmax layer.

[0044] like Figure 2 As shown, in this embodiment, the model inputs a skin disease picture of size [3,224,224], which means that the height and width of the image are both 224 pixels and the number of channels is 3; the feature map passes through the convolution layer, the maximum pooling layer, and the convolution layer respectively. The convolution layer extracts features from the image and generates a new feature map. The two convolution layers capture the patterns and features of the image at different scales. The maximum pooling layer is used to reduce the size of the feature map while retaining important features. Maximum pooling can reduce the spatial dimension of the feature map, thereby reducing the computational cost.

[0045] The residual module is one of the basic building blocks for building residual networks, which aims to improve the training and performance of the network through residual connections. The design of the residual module enables the network to learn and represent complex features more effectively.

[0046] like Figure 3 As shown in the figure, the residual module structure includes three residual modules, two of which are connected in series and used in parallel with another residual module. Each residual module contains two parallel branches, one of which contains a 3×3 convolution layer, a batch normalization layer, a 3×3 convolution layer, and a batch normalization layer connected in sequence, and the other branch is a 1×1 convolution layer; the features obtained by the two branches are added and then passed through a batch normalization layer to obtain the output feature map.

[0047] In this embodiment, 1 and 2 residual modules are connected. The specific structure of each residual module is as follows: the input scale size is [c, w, h] feature map, where c represents the number of channels, w and h represent the width and height of the feature map respectively; the 1×1 convolution layer is used to change the number of channels of the feature map to match the number of channels required by the subsequent convolution layer; the 3×3 convolution layer is responsible for extracting features; nonlinear activation functions are added to the two convolution layers to introduce linear features; the batch normalization layer is used to standardize the input features, which helps to speed up the training process and prevent the gradient from disappearing or exploding; the addition operation adds the newly learned features to the input features, so that deeper networks can be trained.

[0048] like Figure 4 As shown in , the medium-scale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution, and a batch normalization layer connected in sequence; the dilation rate of the 3×3 dilated convolution is 2. Figure 5 As shown in the figure, the wide-scale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, and a batch normalization layer connected in sequence; among which, the dilation rate of the 3×3 dilated convolution layer is 2.

[0049] In this embodiment, the medium-scale dilated convolution and the wide-scale dilated convolution use dilated convolution to reduce the parameters required by the larger convolution kernel, allowing the network to understand more advanced features. The specific structure of this module is as follows: the input scale size is [c,w,h] feature map, where c represents the number of channels, w and h represent the width and height of the feature map respectively; the 3×3 convolution layer is used to learn features, the dilated convolution layer can model the expansion of the receptive field, which helps to capture long-distance dependencies, the medium-scale dilated convolution simulates the receptive field of the 7×7 convolution kernel, and the wide-scale dilated convolution simulates the receptive field of the 15×15 convolution kernel; the batch normalization layer is used to standardize the input features, which helps to speed up the training process and prevent the gradient from disappearing or exploding.

[0050] like Figure 6 As shown in the figure, the multi-scale fusion dilated convolution structure contains two parallel branches. One branch contains three dilated convolutions connected in sequence. After the input feature map undergoes three dilated convolution operations in sequence, the intermediate feature maps corresponding to the first two convolutions and the feature map of the last convolution are concatenated; the other branch is a 1×1 convolution; the outputs of the two branches are added together and then subjected to a 1×1 convolution to obtain the output feature map.

[0051] In this embodiment, each atrous convolution layer uses a different atrous rate to expand the receptive field. The input feature map has a scale of [c, w, h], where c represents the number of channels, and w and h represent the width and height of the feature map, respectively; after three convolution operations, the intermediate feature maps are connected to obtain a feature map of scale [3c, w, h], and 1×1 convolution is used for channel compression. Finally, the compressed feature map is combined with the input feature map through a residual connection. The output feature map simultaneously aggregates the receptive fields of these three different scales, which enables the model to effectively learn feature information at different scales.

[0052] like Figure 2 As shown, in the network architecture of the present invention, all the module structures mentioned above are used in parallel, so that the network can use the behavior it considers to be optimal in each step, and simulate various convolution kernel sizes in three different ways. Different modules output feature maps of the same size [c, w, h]. The feature maps of different modules are spliced ​​in the channel c dimension by splicing, retaining the integrity of the feature map features output by different modules, and making full use of the feature information obtained by different modules. The network can decide how to balance and compensate between different ways of simulating convolution kernels. Having a variety of convolution kernel sizes means that it can find the general area of ​​the target while also correctly finding the edges.

[0053] After the feature maps output by each module are fused, they are input into the batch normalization layer, and the fused features are standardized through normalization technology to keep the data distribution stable; then the feature map is downsampled through the average pooling layer, and the feature map is compressed into a single vector, reducing the spatial dimension of the feature map while retaining the most important feature information; finally, the features are converted into outputs in the form of probability distribution through the softmax layer, and the scores of each category are converted into probability values. The output of the softmax layer can be regarded as the probability of each category.

[0054] 5. Model training and classification

[0055] The training set is sent to the classification model, and the model parameters are updated through model training and back propagation. At the same time, the model is evaluated on the validation set, and the model parameters are optimized according to the evaluation results. After several rounds of training, the trained model is tested on the test set to evaluate the final performance of the classification model.

[0056] 6. Model Evaluation

[0057] After the model training is completed, the classification effect is evaluated on an independent test set, and indicators such as accuracy, recall, F1 value and confusion matrix are used to evaluate the performance of the model on the test set.

[0058] The multi-classification system for common childhood skin diseases based on improved ResNet provides an efficient and accurate method for auxiliary diagnosis of common childhood skin diseases. Through carefully designed data collection, model training and optimization strategies, a powerful auxiliary diagnosis tool can be built to help doctors better identify and manage childhood skin diseases.

[0059] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-classification system for children's skin diseases based on improved ResNet, characterized in that: The invention comprises a computer memory, a computer processor and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory stores a trained classification model; The classification model adopts an improved ResNet network model, which includes a feature extraction unit and a prediction unit; The feature extraction unit comprises a residual module structure, a medium-scale dilated convolution structure, a wide-scale dilated convolution structure and a multi-scale fusion hole convolution structure; the image is sequentially passed through a convolution layer, a maximum pooling layer, a convolution layer and a batch normalization layer, and features are extracted from the image, and then enter the above four structures, which are used in parallel to output their own feature maps; In the prediction unit, the feature maps output by each structure are fused and input into a batch normalization layer, and the fused features are standardized by normalization technology; then the standardized feature maps are downsampled by an average pooling layer, and the feature maps are compressed into a single vector, reducing the spatial dimension of the feature maps while retaining the most important feature information; finally, the features are converted into classification probability distributions of skin diseases by a softmax layer; The computer processor implements the following steps when executing the computer program: inputting the common children's skin disease images to be classified into the trained classification model to obtain the skin disease classification results.

2. The pediatric skin disease multi-classification system based on improved ResNet according to claim 1, characterized in that: The training process of the classification model is as follows: Collect images of affected areas of children with atopic dermatitis, urticaria, hemangioma, and diaper dermatitis and skin images of normal children, and preprocess the image data; 70% of the data are randomly selected as the training set, 10% of the data are randomly selected as the validation set, and the remaining 20% ​​of the data are used as the test set; the training set is used for model training; the validation set is used for model selection and hyperparameter adjustment; and the test set is used to finally evaluate the model performance.

3. The pediatric skin disease multi-classification system based on improved ResNet according to claim 2, characterized in that: The image data is preprocessed, including the following steps: Crop the skin image to reduce irrelevant areas; Image normalization: Normalize the image by mean and standard deviation to ensure that all images have the same brightness and contrast levels; Image enhancement: Use histogram equalization and color space conversion technology to enhance image details; Size normalization: Adjust all input images to a uniform size.

4. The pediatric skin disease multi-classification system based on improved ResNet according to claim 1, characterized in that: The residual module structure comprises three residual modules, two of which are connected in series and used in parallel with another residual module; Each residual module contains two parallel branches, one branch contains a 3×3 convolutional layer, a batch normalization layer, a 3×3 convolutional layer and a batch normalization layer connected sequentially, and the other branch is a 1×1 convolutional layer; the features obtained by the two branches are added and then passed through a batch normalization layer to obtain the output feature map.

5. The pediatric skin disease multi-classification system based on improved ResNet according to claim 1, characterized in that: The medium-scale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution, and a batch normalization layer connected sequentially; among them, the dilation rate of the 3×3 dilated convolution is 2, and the medium-scale dilated convolution structure is used to simulate the receptive field of the 7×7 convolution kernel.

6. The pediatric skin disease multi-classification system based on improved ResNet according to claim 1, characterized in that: The wide-scale dilated convolution structure includes a 3×3 convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, a batch normalization layer, a 3×3 dilated convolution layer, and a batch normalization layer connected in sequence; among them, the dilation rate of the 3×3 dilated convolution layer is 2, and the wide-scale dilated convolution structure is used to simulate the receptive field of a 15×15 convolution kernel.

7. The pediatric skin disease multi-classification system based on improved ResNet according to claim 1, characterized in that: The multi-scale fusion dilated convolution structure contains two parallel branches. One branch contains three dilated convolutions connected in sequence. After the input feature map undergoes three dilated convolution operations in sequence, the intermediate feature maps corresponding to the first two convolutions and the feature map of the last convolution are concatenated; the other branch is a 1×1 convolution; the outputs of the two branches are added together and then subjected to a 1×1 convolution to obtain the output feature map.

8. The pediatric skin disease multi-classification system based on improved ResNet according to claim 7, characterized in that: Each dilated convolution uses a different dilation rate to expand the receptive field.

Images