A method for initial screening of onychomycosis based on improved residual network

Abstract

The present invention discloses a method for primary screening of onychomycosis based on an improved residual network, comprising the following steps: preparing an onychomycosis image dataset; inputting the image dataset into a feature extractor to extract features; integrating the features by a classifier; and calculating the difference between the current value and the true value. The present invention uses a feature extractor based on a lightweight attention mechanism to extract relevant features, inputs them into an improved classifier for classification, and uses a regularization loss function to normalize the output prediction of the network. The effective features in the onychomycosis image are learned and identified through forward and back propagation mechanisms to achieve automatic discrimination of the primary screening classification of onychomycosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and medical image analysis, and in particular relates to a method for initial screening of onychomycosis based on an improved residual network. Background Art

[0002] Onychomycosis is a lesion caused by dermatophytes, yeasts and other molds invading the nail plate and nail bed. It is the most common disease in dermatology and is highly contagious and has a high incidence rate. Onychomycosis accounts for 50% of all nail diseases and 10% of all skin infections, and the number of patients accounts for about 2% - 18% of the world's population. In the early stage of onychomycosis, there is no obvious inflammatory reaction and pain, which leads to a certain cognitive misunderstanding of the disease in many patients, thinking that onychomycosis only affects the appearance of the nails. However, if the patient does not receive timely treatment, it will lead to a series of serious consequences. As a highly contagious disease that frequently occurs among the elderly, onychomycosis is spreading with the trend of aging of the world's population, which will bring great pressure to the medical system. As the number of patients increases year by year and onychomycosis is highly contagious, if it is not treated in time and allowed to cause a large-scale infection, it will bring great pressure to the medical system. If onychomycosis can be detected and treated early, it will greatly reduce the harm caused by this infectious disease and reduce medical costs.

[0003] At present, dermatologists mainly diagnose onychomycosis through laboratory fungal microscopy, fungal culture or histopathology, which usually requires patients to go to the hospital for testing. Due to various reasons such as insufficient knowledge of onychomycosis, high medical expenses, and long waiting time, patients are less willing to go to the hospital for early screening of onychomycosis. Therefore, it is very important to find a simpler and faster method for the initial screening of onychomycosis. However, the initial screening of onychomycosis often faces many challenges, such as the lack of public data sets for onychomycosis, scarcity of labeled samples, easy confusion with other nail diseases, the existing classification network is too heavyweight, and prior knowledge is insufficient. How to make a suitable data set and explore more effective and lightweight onychomycosis classification methods, and apply them to the initial screening of onychomycosis, is one of the important issues currently faced.

[0004] In recent years, with the development of high-tech and the advent of the digital age, a wave of technological innovation has been set off again in various fields such as science, industry and medicine. Deep learning has revolutionized the medical image computing method due to its ability to automatically learn features and superior performance. It has achieved many results in medical image classification, segmentation and registration, and has performed well in medical imaging fields such as pathology, ophthalmology, radiology and skin. Therefore, for the initial screening of onychomycosis, a residual network based on lightweight attention and regularization improvement is proposed, and the onychomycosis classification algorithm framework is built through this network. Summary of the invention

[0005] In order to solve the problem in the initial screening detection of onychomycosis, the present invention provides a method for initial screening detection of onychomycosis based on an improved residual network to achieve the binary classification problem of onychomycosis.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for initial screening of onychomycosis based on an improved residual network comprises the following steps:

[0008] S1) Create a dataset of onychomycosis images;

[0009] S2) inputting the image data set obtained in step S1) into a feature extractor to perform feature extraction;

[0010] S3) The classifier integrates the features obtained in step S2): The features after dimension reduction in step S2) are input into the improved classifier module. The dual fully connected layer in the module can integrate multiple local features, find the correlation between channels, and map the feature vector to the sample category space for effective classification;

[0011] S4) Calculate the difference between the current value and the true value: Based on the input image of the training sample and the corresponding category label, the model learns to identify the effective features in the onychomycosis image through forward and back propagation mechanisms, uses the regularization loss function to standardize the output prediction of the network, optimizes the weight parameters of the model, and improves the accuracy of onychomycosis classification.

[0012] Furthermore, in step S3), the Dropout technique can be used to randomly discard network nodes to avoid overfitting due to too many parameters.

[0013] Furthermore, the specific steps in step S1) are as follows:

[0014] S11) First, obtain the public onychomycosis samples as required, remove dirty data and perform batch processing and standardization operations.

[0015] S12) Use the target detection model to automatically identify the nail position and crop it, completing the expansion and unification of the data set samples.

[0016] S13) Finally, doctors re-label the data to obtain a dataset of onychomycosis images.

[0017] Furthermore, the specific steps in step S2) are as follows:

[0018] S21) The data in the training set flows through each convolution module in the network in turn, and the input image first enters the first convolution module Block1;

[0019] S22) then sends the features to four convolution modules Block2, Block3, Block4 and Block5. The implementation process can be seen Figure 1 .

[0020] Furthermore, the convolution module Block1 is composed of 64 large convolution kernels with a step size of 2 and a size of 7x7.

[0021] Furthermore, the convolution modules Block2, Block3, Block4 and Block5 are composed of 3, 4, 6 and 3 lightweight attention residual blocks respectively, wherein the lightweight attention residual block is realized by adding the ECA module.

[0022] Furthermore, the specific steps of processing data in the classifier are:

[0023] S31) First, the input feature map is flattened, and the one-dimensional array is input into the first fully connected layer.

[0024] S32) is then sent to another fully connected layer after passing through the ReLU activation function and the Dropout layer.

[0025] S33) Map the feature vector to the sample category space for binary classification of onychomycosis.

[0026] Beneficial effects: The present invention uses a feature extractor based on a lightweight attention mechanism to extract relevant features, inputs them into an improved classifier for classification, and uses a regularized loss function to normalize the output prediction of the network. Through forward and back propagation mechanisms, learning and identifying effective features in onychomycosis images can achieve automatic discrimination of onychomycosis initial screening classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A network structure diagram of the onychomycosis initial screening method based on the improved residual network of the present invention;

[0028] Figure 2 This is a detailed information diagram of the convolution layer in the feature extraction module of the present invention. DETAILED DESCRIPTION

[0029] The present invention is described below with reference to specific examples. It will be appreciated by those skilled in the art that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention in any way.

[0030] like Figure 1 As shown, a method for initial screening of onychomycosis based on an improved residual network comprises the following steps:

[0031] S1) Create a dataset of onychomycosis images. The specific steps are as follows:

[0032] S11) First, obtain the public onychomycosis samples as required, remove dirty data and perform batch processing and standardization operations.

[0033] S12) Use the target detection model to automatically identify the nail position and crop it, completing the expansion and unification of the data set samples.

[0034] S13) Finally, doctors relabeled the data to obtain a preliminary screening dataset for onychomycosis.

[0035] In this embodiment, the data set needs to include a large number of photos taken in real life, and try to cover various forms of nail images of onychomycosis and non-onychomycosis, including diseased nails infected by onychomycosis, dystrophic nails of non-onychomycosis, and normal nails.

[0036] S2) Input the image data set into the feature extractor for feature extraction. The specific steps are as follows:

[0037] S21) The data in the training set flows through each convolution module in the network in turn. The detailed information of the convolution module is as follows: Figure 1 As shown in the figure, the input image first enters the first convolution module Block1, which consists of 64 large convolution kernels with a step size of 2 and a size of 7x7;

[0038] S22) then feeds the features into four convolutional modules Block2, Block3, Block4, and Block5, which are composed of 3, 4, 6, and 3 lightweight attention residual blocks, respectively. The lightweight attention residual block is implemented by adding the ECA module. The implementation process can be seen in Figure 1 The convolution module can gradually extract high-level semantic features of the image, and the attention mechanism can prevent the loss of texture detail information in the lesion area.

[0039] S3) The classifier integrates the features obtained in step S2): the features after dimensionality reduction in step S2) are input into the improved classifier module. The dual fully connected layer in the module can integrate multiple local features, find the correlation between channels, and map the feature vector to the sample category space for effective classification; and use the Dropout technology to randomly discard network nodes to avoid overfitting caused by too many parameters. The specific steps of processing data in the classifier are:

[0040] S31) First, the input feature map is flattened, and the one-dimensional array is input into the first fully connected layer.

[0041] S32) is then sent to another fully connected layer after passing through the ReLU activation function and the Dropout layer.

[0042] S33) Map the feature vector to the sample category space for binary classification of onychomycosis.

[0043] S4) Calculate the difference between the current value and the true value: Based on the input image of the training sample and the corresponding category label, the model learns to identify the effective features in the onychomycosis image through forward and back propagation mechanisms, uses the regularization loss function to standardize the output prediction of the network, optimizes the weight parameters of the model, and improves the accuracy of onychomycosis classification.

[0044] In this embodiment, the R-Drop loss function is used to standardize the output prediction of the sub-network to overcome the problem that the structural randomness brought by the Dropout technology in the classifier makes multiple sub-networks differ greatly without any constraints. Specifically, the model minimizes the negative log-likelihood loss function so that the predicted value of the model is closer to the true value. The image samples in the data set will obtain different probability distributions after adding the model with the Dropout structure, and the Kullback-Leibler, KL divergence is used to measure the difference between the two probability distributions. By taking the weighted sum of the log-likelihood function and the KL divergence, R-Drop further standardizes the structure of the network and constrains the sub-network.

[0045] In this embodiment, the experimental results are mainly evaluated from two aspects: classification performance and network complexity. The classification performance includes accuracy, specificity, sensitivity, positive predictive value and F1 score, and the network complexity includes parameter quantity and floating point calculation quantity. The test results are shown in Table 1 and Table 2, where Table 1 shows the classification performance of the model in percentage, %, and Table 2 shows the network complexity of the model.

[0046] Table 1 shows the classification performance of the onychomycosis screening model

[0047] Model Accuracy Specificity Sensitivity Positive predictive value F1 score (VGG19+ANN) 90.37 90.13 90.65 88.94 89.79 <![CDATA[(VGG19+ResNet152) FF ]]> 89.59 88.34 91.09 86.59 88.78 <![CDATA[(VGG19+ResNet152) DF ]]> 91.94 89.29 95.37 87.29 91.15 proposed 91.12 90.68 91.50 90.23 90.45

[0048] Table 2 shows the network complexity of the onychomycosis screening model

[0049] Model Parameter quantity (M) Floating point calculation amount (B) (VGG19+ANN) 48.24 8.97 <![CDATA[(VGG19+ResNet152) FF ]]> 203.86 31.38 <![CDATA[(VGG19+ResNet152) DF ]]> 203.86 31.38 proposed 29.75 4.37

Claims

1. A method for initial screening of onychomycosis based on an improved residual network, characterized in that: The following steps are involved: S1) Create a dataset of onychomycosis images; S2) inputting the image data set obtained in step S1) into a feature extractor to perform feature extraction; S3) The classifier integrates the features obtained in step S2): The features after dimension reduction in step S2) are input into the improved classifier module. The dual fully connected layer in the module can integrate multiple local features, find the correlation between channels, and map the feature vector to the sample category space for effective classification; S4) Calculate the difference between the current value and the true value: Based on the input image of the training sample and the corresponding category label, the model learns to identify the effective features in the onychomycosis image through the forward and back propagation mechanism, uses the regularization loss function to standardize the output prediction of the network, optimizes the weight parameters of the model, and improves the accuracy of onychomycosis classification; The specific steps in step S2) are as follows: S21) The data in the training set flows through each convolution module in the network in turn, and the input image first enters the first convolution module Block1; S22) then sends the features to four convolutional modules Block2, Block3, Block4 and Block5; The convolution modules Block2, Block3, Block4 and Block5 are composed of 3, 4, 6 and 3 lightweight attention residual blocks respectively, among which the lightweight attention residual block is realized by adding the ECA module.

2. The method for primary screening of onychomycosis based on improved residual network according to claim 1, characterized in that: In step S3), the Dropout technique can also be used to randomly discard network nodes to avoid overfitting due to too many parameters.

3. The method for primary screening of onychomycosis based on improved residual network according to claim 1, characterized in that: The specific steps in step S1) are as follows: S11) First, obtain the public onychomycosis samples as required, remove dirty data and perform batch processing and standardization operations. S12) Use the target detection model to automatically identify the nail position and crop it, completing the expansion and unification of the data set samples. S13) Finally, doctors re-label the data to obtain a dataset of onychomycosis images.

4. The method for primary screening of onychomycosis based on improved residual network according to claim 1, characterized in that: The convolution module Block1 consists of 64 large convolution kernels with a step size of 2 and a size of 7x7.

5. The method for primary screening of onychomycosis based on improved residual network according to claim 1, characterized in that: The specific steps of processing data in the classifier are: S31) First, the input feature map is flattened, and the one-dimensional array is input into the first fully connected layer. S32) is then sent to another fully connected layer after passing through the ReLU activation function and the Dropout layer. S33) Map the feature vector to the sample category space for binary classification of onychomycosis.

Images