Training method, device, equipment and medium based on skin disease classification model

Abstract

The present invention relates to the field of image recognition detection technology, and discloses a training method, device, equipment and medium for a classification model based on skin diseases, including: preliminary training of an original convolutional neural network, and adding a graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network; inputting image set data containing skin diseases as parameters into the preliminary convolutional neural network to extract feature vector set data; processing the feature vector set data to obtain a network prediction value; obtaining a cross entropy loss set data based on the network prediction value and the original label of the skin disease; training the training parameters in the classification model based on the cross entropy loss set data to obtain a classification model based on skin diseases. The present invention can improve the classification accuracy of the trained classification model.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition and detection, and in particular to a training method, device, equipment and medium for a classification model based on skin diseases. Background Art

[0002] Most existing image classification models can only classify natural image data. For skin disease image data, different skin diseases have many similar visual features, and the same skin disease may have large differences in performance. At the same time, since the skin disease data set comes from the disease data collected by the partners during the online consultation process, the data distribution is uneven. As a result, when using skin disease image data to train the existing image classification model, it will affect the convergence process of the network in the image classification model, causing the image classification model to overfit, thereby affecting the training results. Summary of the invention

[0003] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a training method, device, equipment and medium for a classification model based on skin diseases, which can improve the classification accuracy of the trained classification model.

[0004] To achieve the above objectives and other related objectives, in a first aspect, the present invention provides a training method for a classification model based on skin diseases, comprising:

[0005] Performing preliminary training on the original convolutional neural network, and adding the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network;

[0006] Using the image set data containing skin diseases as parameters, inputting them into the preliminary convolutional neural network to extract feature vector set data;

[0007] Processing the feature vector set data to obtain a network prediction value;

[0008] According to the network prediction value and the original label of the skin disease, a cross entropy loss set data is obtained;

[0009] The training parameters in the classification model are trained according to the cross entropy loss set data to obtain a classification model based on skin diseases.

[0010] In a second aspect, a training device for a classification model based on skin diseases is provided, comprising:

[0011] The first training module is used to perform preliminary training on the original convolutional neural network and add the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network;

[0012] A feature vector extraction module, used for inputting the image set data containing skin diseases as parameters into a preliminary convolutional neural network to extract feature vector set data;

[0013] A processing module, used for processing the feature vector set data to obtain a network prediction value;

[0014] A loss module, used to obtain cross entropy loss set data according to the network prediction value and the original label of the skin disease; and

[0015] The second training module is used to train the training parameters in the classification model according to the cross entropy loss set data to obtain a classification model based on skin diseases.

[0016] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned training method of the skin disease-based classification model when executing the computer program.

[0017] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the training method of the above-mentioned skin disease-based classification model are implemented.

[0018] In the scheme implemented by the training method, device, equipment and medium based on the skin disease classification model, the learning process of the network can be made relatively smooth, and it always converges to an effective optimization direction without overfitting. The network is always optimized based on the mutual mixing status under its current parameters, so that the similarity distribution of the image set data can be adaptively learned with trainable edges, so that the classification accuracy of the trained classification model is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0020] Figure 1 The figure is a flow chart of a training method of a skin disease classification model according to the present invention.

[0021] Figure 2 for Figure 1 A schematic flow chart of a specific implementation of step S10 in FIG.

[0022] Figure 3 for Figure 1A schematic flow chart of a specific implementation of step S30 in FIG.

[0023] Figure 4 The figure is a schematic structural diagram of a training device based on a classification model of skin diseases according to the present invention.

[0024] Figure 5 for Figure 4 A structural schematic diagram of the first training module in .

[0025] Figure 6 for Figure 4 A structural schematic diagram of a processing module in.

[0026] Figure 7 It is a structural diagram of a computer device in one embodiment of the present invention.

[0027] Component number description:

[0028] 10. First training module; 11. Data acquisition module; 12. Input module; 13. Replacement module; 20. Feature vector extraction module; 30. Processing module; 31. First processing module; 32. Second processing module; 33. Third processing module; 40. Loss module; 50. Second training module. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] See also Figure 1-7 It should be noted that the illustrations provided in this embodiment are only used to illustrate the basic concept of the present invention in a schematic manner, and the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0031] The training method of the classification model based on skin diseases provided in the embodiment of the present invention can be applied in the neural network model. Among them, the traditional convolutional neural network in the classification model can be trained by the image set data containing skin diseases, so as to obtain a preliminary converged classification model, and the convolutional neural network can extract relatively effective high-level semantic information of the image. However, since different skin diseases share many similar visual features, there may also be large performance differences between the same type of skin diseases. The image set data containing skin diseases has an uneven data distribution, and there is a high degree of visual similarity between some diseases. Therefore, these differences will affect the convergence process of the internal network of the classification model from different aspects, resulting in model overfitting. Although a very small loss can be converged during the training process, it performs poorly on the overall training set, the overall predicted distribution shifts in the direction of large-scale data, the recall rate of categories with a small proportion of quantity is low, and there are many mixed samples between similar categories, so the classification model after preliminary training cannot accurately classify skin diseases. In the present invention, for the classification in the skin disease scenario, the traditional convolutional neural network in the classification model can be trained first, and then the graph convolutional neural network replaces the classification head in the convolutional neural network, and the classification model is retrained, the obtained vector set data is dimensionally transformed, and finally the cross entropy loss set data is obtained. The neural network inside the classification model is trained by the cross entropy loss set data, so that the accuracy of the trained classification model can be improved. The present invention is described in detail below through specific embodiments.

[0032] See also Figure 1 As shown, Figure 1 A flowchart of a training method for a skin disease classification model provided by an embodiment of the present invention includes the following steps:

[0033] Step S10: perform preliminary training on the original convolutional neural network, and add the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network.

[0034] Step S20: input the image set data containing skin diseases as parameters into a preliminary convolutional neural network to extract feature vector set data.

[0035] Step S30: Process the feature vector set data to obtain a network prediction value.

[0036] Step S40: Obtain cross entropy loss set data according to the network prediction value and the original label of the skin disease.

[0037] Step S50: training the training parameters of the preliminary convolutional neural network, graph convolutional neural network and fully connected layer network in the classification model according to the cross entropy loss set data to obtain a classification model based on skin diseases.

[0038] See also Figure 2 As shown, in one embodiment of the present invention, when performing step S10, the original convolutional neural network is preliminarily trained, and the graph convolutional neural network is added to the original convolutional neural network to obtain a preliminary convolutional neural network. Specifically, the sub-steps of step S10 may include:

[0039] Step S11, acquiring image set data containing skin diseases.

[0040] Step S12: input the image set data as parameters into the original convolutional neural network in the classification model to perform preliminary training on the training parameters of the original convolutional neural network.

[0041] Step S13: adding the graph convolutional neural network to the trained original convolutional neural network and replacing the head classifier of the original convolutional neural network to obtain a preliminary convolutional neural network.

[0042] In one embodiment of the present invention, when step S11 is performed, image set data containing skin diseases is obtained. Specifically, the image set data containing skin diseases can come from a collection of images collected by a cooperative unit (such as a hospital, etc.) during a consultation process. The image set data can also come from images obtained by the patient's selfie, so as to obtain a collection after the patient uploads the image. The image set data can also actively search for various public images through the Internet to form a collection of images. The image set data may include multiple images, and the images may include different types of skin diseases, so that each image has corresponding skin disease data. The image set data has the characteristics of inter-class co-occurrence, intra-class diversity, one-hot label, and data imbalance.

[0043] In one embodiment of the present invention, specifically, inter-class co-occurrence refers to that different skin diseases share many similar visual features, such as papules, blisters, redness, plaques with clear boundaries, etc., and they are all stable location features that are only distributed on the face or hands. Intra-class diversity refers to that there may also be large differences in performance between the same type of skin diseases, such as the early, middle and late performances of dyshidrotic eczema are different, the middle stage is blisters similar to eczema, and the late stage is skin lesions similar to tinea manuum. One-hot labels refer to that in actual diagnosis scenarios, doctors often give multiple preliminary diagnosis conclusions first, and then get the final diagnosis conclusion based on some consultation information. That is to say, although the annotations we get are one-hot, some diseases are often not suitable for simple expression with exclusive labels such as 1 or 0 due to high visual similarity. Data imbalance refers to the fact that the image set data comes from disease data collected by partners during online consultations. The data distribution is determined by factors such as disease incidence and the feasibility of online consultations. Large chronic diseases such as dermatitis and eczema account for a large proportion and have various manifestations. Diseases with a more complicated diagnosis process, such as lupus erythematosus, appear less frequently in online databases.

[0044] In one embodiment of the present invention, when performing step S12, the image set data is used as a parameter and input into the original convolutional neural network in the classification model to perform preliminary training on the training parameters of the original convolutional neural network. Specifically, the classification model may be an efficientNet-v2s classification model. When the image set data is used as a parameter and input into the original convolutional neural network in the classification model, the original convolutional neural network will scan and read each image in the image set data in a certain order, so that the training parameters in the original convolutional neural network can be preliminarily trained. At this time, the original convolutional neural network in the classification model after preliminary training can extract relatively effective high-level semantic information of the image.

[0045] In one embodiment of the present invention, when performing step S13, the graph convolutional neural network is added to the original convolutional neural network after training, and the head classifier of the original convolutional neural network is replaced to obtain a preliminary convolutional neural network. Among them, the convolutional neural network (Convolutional Neural Networks, referred to as CNN) is a type of feedforward neural network that includes convolution calculations and has a deep structure, and is one of the representative algorithms of deep learning. The convolutional neural network has the ability to represent learning and can perform translation-invariant classification of input information according to its hierarchical structure. The convolutional neural network contains a head classifier, which can be removed, and the graph convolutional neural network is added to the convolutional neural network to replace the head classifier. The graph convolutional neural network (Graph Convolutional Network, referred to as GCN) actually has the same function as CNN, which is a feature extractor, and the object of GCN is graph data. GCN has ingeniously designed a method for extracting features from graph data, so that these features can be used to perform node classification, graph classification, and link prediction on graph data, and can also obtain graph embedding representation (graph embedding).

[0046] In one embodiment of the present invention, when performing step S20, the image set data is input as a parameter into a preliminary convolutional neural network to extract feature vector set data. Specifically, before performing step S20, the preliminary convolutional neural network has been preliminarily trained, and the image set data can be input as a parameter into the preliminary convolutional neural network in the classification model, so that the image set data is extracted through the preliminary convolutional neural network to obtain the corresponding feature vector set data.

[0047] See also Figure 3 As shown, in one embodiment of the present invention, when step S30 is performed, the feature vector set data is processed to obtain a network prediction value. Specifically, the sub-steps of step S30 may include:

[0048] Step S31, process the feature vector set data to obtain high-dimensional feature vector set data, expressed as X[b, D, c], where b is the number of parallel samples, c is the feature dimension, and D is the dimension of the high-dimensional feature vector set data.

[0049] Step S32: input the high-dimensional feature vector set data into the fully connected layer network to obtain the first feature vector set data for different skin disease categories, expressed as X[b, D, N], and perform dimensionality change processing on the first feature vector set data to obtain the second feature vector set data, expressed as X[b, N, D], wherein N is the number of skin disease categories in the graph structure, the graph structure is expressed as A[N, N], and A represents the relationship between different skin disease categories.

[0050] Step S33: input the second feature vector set data as a parameter into the graph convolutional neural network to obtain the output vector set data Y, expressed as Y=ReLU(AXW+B), wherein the dimension of the output vector set data Y is [b, N, D1], D1 is the dimension of the second feature vector set data, W is the weight of the graph convolutional layer, and B is the offset of the graph convolutional layer.

[0051] Step S34: perform dimension change processing on the output vector set data to obtain a network prediction value, wherein the dimension of the network prediction value is [b, N].

[0052] In one embodiment of the present invention, when step S31 is performed, the feature vector set data is processed to obtain high-dimensional feature vector set data, expressed as X[b, D, c], wherein b is the number of parallel samples, c is the feature dimension, and D is the dimension of the high-dimensional feature vector set data. Specifically, the output of the original convolutional neural network is a feature map of [b, h, w, c], where h and w represent the dimensions of the feature vector, b is the number of parallel samples, and c is the feature dimension. After the feature vector set data is obtained, the feature vector set data can be expanded in the spatial dimension to obtain high-dimensional feature vector set data, expressed as X[b, D, c].

[0053] In one embodiment of the present invention, when performing step S32, the high-dimensional feature vector set data is input into the fully connected layer network to obtain the first feature vector set data for different skin disease categories, expressed as X [b, D, N], and the first feature vector set data is processed by dimension change to obtain the second feature vector set data, expressed as X [b, N, D]. Specifically, in the CNN structure, after multiple convolutional layers and pooling layers, one or more fully connected layers are connected, and each neuron in the fully connected layer is fully connected with all neurons in the previous layer. The fully connected layer can integrate the local information with category distinction in the convolutional layer or the pooling layer. In order to improve the network performance of CNN, the excitation function of each neuron in the fully connected layer generally adopts the ReLU function. When the high-dimensional feature vector set data is input into the fully connected layer network as a parameter, the spatial dimension of the high-dimensional feature vector set data can be compressed, so that the output feature can be converted into a feature vector for each category, that is, the first feature vector set data for different skin disease categories is obtained, expressed as X [b, D, N]. The first feature vector set data can then be dimensionalized by swapping the last two dimensions to obtain the second feature vector set data, expressed as X[b, N, D], where each row vector is used as the feature vector of each node in the graph convolution input layer.

[0054] In one embodiment of the present invention, when performing step S33 and step S34, the second feature vector set data is used as a parameter and input into the graph convolutional neural network to obtain the output vector set data Y, which is expressed as Y=ReLU(AXW+B), wherein the dimension of the output vector set data Y is [b, N, D1], D1 is the dimension of the second feature vector set data, W is the weight of the graph convolution layer, and B is the offset of the graph convolution layer. Specifically, when each row vector of the second feature vector set data is used as the feature vector of each node of the graph convolution input layer and input into the graph convolutional neural network, the corresponding output vector set data Y can be obtained. The output vector set data can then be processed for dimensionality change to obtain the corresponding network prediction value.

[0055] In one embodiment of the present invention, when performing step S40 and step S50, the original label may be the corresponding label of the skin disease in the image set data, and each skin disease may correspond to a corresponding label. Afterwards, the training parameters in the classification model may be trained according to the cross entropy loss set data, so as to obtain a trained classification model. Among them, the training parameters may include the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network, and the training parameters of the fully connected layer network. Specifically, the cross entropy loss set data may be input into the classification model, and the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network, and the training parameters of the fully connected layer network in the classification model may be iteratively trained, and the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network, and the training parameters of the fully connected layer network may be optimized based on the results of the iterative training, so as to obtain a trained classification model.

[0056] It can be seen that in the above scheme, the skin disease category is represented by a graph node, and each node is represented by a high-dimensional feature vector. This feature vector can be a comprehensive description of the high-level semantic information of each sample by the neural network. The similarity relationship between different skin disease categories can be represented by the edges between nodes, and represented by an (N, N) undirected graph, that is, the adjacency matrix, which can be understood as the similarity relationship between classes. In the initial state, the relationship between classes in the image set data is unclear, so this undirected graph can be set to a full-one matrix, that is, the distance between classes is considered to be equal. During the training process, the values ​​in the adjacency matrix will also be gradually updated with the gradient backpropagation, and the intensity of information exchange between each node (category) and other nodes will also be adjusted with the change of the similarity relationship. Using the network structure to build inter-class similarity, for highly similar categories, the information interaction between graph nodes is stronger, which can promote the extraction of difference points between mixed categories and improve separability. For categories with low correlation, it is obvious that the extracted representation information is not needed for decision-making, so the information interaction is weak. This can make the network learning process always smoother and always converge towards an effective optimization direction without overfitting. The network is always optimized based on the mutual mixing status under its current parameters, so that the similarity distribution of image set data can be adaptively learned with trainable edges, thereby improving the classification accuracy of the trained classification model.

[0057] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0058] See also Figure 4As shown, the present invention also provides a training device based on a classification model for skin diseases, which corresponds one-to-one to the training method based on the classification model for skin diseases in the above embodiment. The training device based on the classification model for skin diseases may include a first training module 10, a feature vector extraction module 20, a processing module 30, a loss module 40, and a second training module 50. Each functional module is described in detail as follows:

[0059] The first training module 10 is used to perform preliminary training on the original convolutional neural network and add the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network.

[0060] The feature vector extraction module 20 is used to input the image set data containing skin diseases as parameters into the preliminary convolutional neural network to extract feature vector set data.

[0061] The processing module 30 is used to process the feature vector set data to obtain a network prediction value.

[0062] The loss module 40 is used to obtain cross entropy loss set data according to the network prediction value and the original label of the skin disease.

[0063] The second training module 50 is used to train the training parameters of the preliminary convolutional neural network, the graph convolutional neural network and the fully connected layer network in the classification model according to the cross entropy loss set data to obtain a classification model based on skin diseases.

[0064] See also Figure 5 As shown, in one embodiment of the present invention, the first training module 10 may include a data acquisition module 11, an input module 12 and a replacement module 13. Among them, the data acquisition module 11 can be used to obtain image set data containing skin diseases. Specifically, the image set data containing skin diseases can come from a collection of images collected by a cooperative unit (such as a hospital, etc.) during a consultation process. The image set data can also come from images obtained by the patient's self-portrait, so that the patient uploads the image and obtains the collection. The image set data can also actively search for various public images through the Internet to form a collection of images. The image set data may contain multiple images, and the images may contain different types of skin diseases, so that each image has corresponding skin disease data. The image set data has the characteristics of inter-class co-occurrence, intra-class diversity, one-hot label, and data imbalance.

[0065] In one embodiment of the present invention, the input module 12 can be used to input the image set data as a parameter into the original convolutional neural network in the classification model to perform preliminary training on the training parameters of the original convolutional neural network. Specifically, the classification model can be an efficientNet-v2s classification model. When the image set data is input as a parameter into the original convolutional neural network in the classification model, the original convolutional neural network will scan and read each image in the image set data in a certain order, so that the training parameters in the original convolutional neural network can be preliminarily trained. At this time, the original convolutional neural network in the classification model after preliminary training can extract relatively effective high-level semantic information of the image.

[0066] In one embodiment of the present invention, the replacement module 13 can be used to add a graph convolutional neural network to the original convolutional neural network after training, and replace the head classifier of the original convolutional neural network to obtain a preliminary convolutional neural network. Specifically, convolutional neural networks (CNN) are a type of feedforward neural network that includes convolution calculations and has a deep structure, and is one of the representative algorithms of deep learning. Convolutional neural networks have representation learning capabilities and can perform translation-invariant classification of input information according to their hierarchical structure. The convolutional neural network contains a head classifier, which can be removed, and the graph convolutional neural network is added to the convolutional neural network to replace the head classifier. The graph convolutional neural network (GCN) actually has the same function as CNN, and is a feature extractor. The object of GCN is graph data. GCN has ingeniously designed a method for extracting features from graph data, so that these features can be used to perform node classification, graph classification, and link prediction on graph data, and can also obtain graph embedding representation (graph embedding).

[0067] In one embodiment of the present invention, the feature vector extraction module 20 can be used to input the image set data as a parameter into a preliminary convolutional neural network to extract feature vector set data. Specifically, the preliminary convolutional neural network has been preliminarily trained, and the image set data can be used as a parameter to input into the preliminary convolutional neural network in the classification model, so that the image set data can be extracted through the preliminary convolutional neural network to obtain the corresponding feature vector set data.

[0068] See also Figure 6As shown, in one embodiment of the present invention, the processing module 30 may include a first processing module 31, a second processing module 32, a third processing module 33 and a fourth processing module 34. Among them, the first processing module 31 can be used to process the feature vector set data to obtain high-dimensional feature vector set data. Specifically, the output of the original convolutional neural network is a feature map of [b, h, w, c], where h and w represent the dimensions of the feature vector, b is the number of parallel samples, and c is the feature dimension. After obtaining the feature vector set data, the feature vector set data can be expanded in the spatial dimension to obtain high-dimensional feature vector set data, expressed as X[b, D, c].

[0069] In one embodiment of the present invention, the second processing module 32 can be used to input the high-dimensional feature vector set data into the fully connected layer network to obtain the first feature vector set data for different skin disease categories. Specifically, in the CNN structure, after multiple convolutional layers and pooling layers, one or more fully connected layers are connected, and each neuron in the fully connected layer is fully connected with all neurons in the previous layer. The fully connected layer can integrate the local information with category distinction in the convolutional layer or the pooling layer. In order to improve the network performance of CNN, the excitation function of each neuron in the fully connected layer generally adopts the ReLU function. When the high-dimensional feature vector set data is input as a parameter into the fully connected layer network, the spatial dimension of the high-dimensional feature vector set data can be compressed, so that the output feature can be converted into a feature vector for each category, that is, the first feature vector set data for different skin disease categories is obtained, expressed as X [b, D, N]. The first feature vector set data can then be dimensionalized by swapping the last two dimensions to obtain the second feature vector set data, expressed as X[b, N, D], where each row vector is used as the feature vector of each node in the graph convolution input layer.

[0070] In one embodiment of the present invention, the third processing module 33 can be used to input the second feature vector set data as a parameter into the graph convolutional neural network to obtain output vector set data. The fourth processing module 34 can be used to perform dimensionality change processing on the output vector set data to obtain a network prediction value. Specifically, when each row vector of the second feature vector set data is used as the feature vector of each node of the graph convolution input layer and input into the graph convolutional neural network, the corresponding output vector set data can be obtained. After that, the output vector set data can be processed for dimensionality change to obtain the corresponding network prediction value.

[0071] In one embodiment of the present invention, the loss module 40 can be used to obtain cross entropy loss set data according to the network prediction value and the original label. The second training module 50 can be used to train the training parameters of the preliminary convolutional neural network, the graph convolutional neural network and the fully connected layer network in the classification model according to the cross entropy loss set data to obtain a trained classification model. Specifically, the original label can be the corresponding label of the skin disease in the image set data, and each skin disease can correspond to a corresponding label. After that, the training parameters in the classification model can be trained according to the cross entropy loss set data, so as to obtain a trained classification model. The training parameters may include the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network and the training parameters of the fully connected layer network. Specifically, the cross entropy loss set data can be input into the classification model, and the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network and the training parameters of the fully connected layer network in the classification model are iteratively trained, and the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network and the training parameters of the fully connected layer network are optimized based on the results of the iterative training, so as to obtain a classification model based on skin diseases.

[0072] The present invention provides a training device for a classification model based on skin diseases, which can make the learning process of the network always relatively smooth, and always converge to an effective optimization direction without overfitting. The network is always optimized based on the mutual mixing status under its current parameters, so that the similarity distribution of image set data can be adaptively learned with trainable edges, so that the classification accuracy of the trained classification model is improved.

[0073] For the specific definition of the training device based on the classification model of skin diseases, please refer to the definition of the training method based on the classification model of skin diseases mentioned above, which will not be repeated here. Each module in the above-mentioned training device based on the classification model of skin diseases can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0074] See also Figure 7As shown, the present invention also provides a computer device, which includes a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client through a network connection. When the computer program is executed by the processor, the functions or steps of a training method based on a classification model of skin diseases are implemented.

[0075] In one embodiment of the present invention, a computer device may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are implemented:

[0076] Perform preliminary training on the original convolutional neural network, and add the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network;

[0077] The image set data is input as a parameter into a preliminary convolutional neural network to extract feature vector set data;

[0078] Processing the feature vector set data to obtain the network prediction value;

[0079] According to the network prediction value and the original label, the cross entropy loss set data is obtained;

[0080] The training parameters of the preliminary convolutional neural network, graph convolutional neural network and fully connected layer network in the classification model are trained according to the cross entropy loss set data to obtain a trained classification model.

[0081] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant descriptions on the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0082] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0083] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0084] In the description of this specification, the description with reference to the terms "this embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0085] The embodiments of the present invention disclosed above are only used to help illustrate the present invention. The embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A training method for a classification model based on skin diseases, characterized in that: Includes steps: Preliminarily training the original convolutional neural network, and adding the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network; Using the image set data containing skin diseases as parameters, inputting them into the preliminary convolutional neural network to extract feature vector set data; Processing the feature vector set data to obtain high-dimensional feature vector set data; the high-dimensional feature vector set data is represented by X [b, D, c], where b is the number of parallel samples, c is the feature dimension, and D is the dimension of the high-dimensional feature vector set data; Inputting the high-dimensional feature vector set data into a fully connected layer network to obtain first feature vector set data for different skin disease categories; the first feature vector set data is represented as X[b, D, N]; Performing dimension change processing on the first feature vector set data to obtain the second feature vector set data; the second feature vector set data is represented as X[b, N, D], wherein N is the number of skin disease categories in the graph structure, and the graph structure is represented as A[N, N], and A represents the relationship between different skin disease categories; The second feature vector set data is used as a parameter and input into the graph convolutional neural network to obtain output vector set data; the output vector set data Y is expressed as Y=ReLU(AXW+B), wherein the dimension of the output vector set data Y is [b, N, D1], D1 is the dimension of the second feature vector set data, W is the weight of the graph convolution layer, and B is the offset of the graph convolution layer; Performing dimension change processing on the output vector set data to obtain a network prediction value; the dimension of the network prediction value is [b, N], where b is the number of parallel samples; According to the network prediction value and the original label of the skin disease, a cross entropy loss set data is obtained; The training parameters in the classification model are trained according to the cross entropy loss set data to obtain a classification model based on skin diseases.

2. The training method of the skin disease classification model according to claim 1, characterized in that: The sub-steps of preliminarily training the original convolutional neural network and adding the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network include: Using the image set data as parameters, inputting them into the original convolutional neural network, and performing preliminary training on the training parameters of the original convolutional neural network; The graph convolutional neural network is added to the trained original convolutional neural network to replace the head classifier of the original convolutional neural network to obtain the preliminary convolutional neural network.

3. The training method of the skin disease classification model according to claim 2, characterized in that: Before the step of inputting the image set data as parameters into the original convolutional neural network to perform preliminary training on the training parameters of the original convolutional neural network, the method further includes the following steps: Get image set data containing skin diseases.

4. The training method of the skin disease classification model according to claim 1, characterized in that: The training parameters in the classification model include the training parameters of the preliminary convolutional neural network, the training parameters of the graph convolutional neural network and the training parameters of the fully connected layer network.

5. A training device based on a classification model for skin diseases, characterized in that: include: The first training module is used to perform preliminary training on the original convolutional neural network and add the graph convolutional neural network to the original convolutional neural network to obtain a preliminary convolutional neural network; A feature vector extraction module, used for inputting the image set data containing skin diseases as parameters into a preliminary convolutional neural network to extract feature vector set data; A processing module is used to process the feature vector set data to obtain high-dimensional feature vector set data; the high-dimensional feature vector set data is expressed as X[b, D, c], where b is the number of parallel samples, c is the feature dimension, and D is the dimension of the high-dimensional feature vector set data; the high-dimensional feature vector set data is input into a fully connected layer network to obtain first feature vector set data for different skin disease categories; the first feature vector set data is expressed as X[b, D, N]; the first feature vector set data is subjected to dimensionality change processing to obtain the second feature vector set data; the second feature vector set data is expressed as X[b, N, D], where N is the image The number of skin disease categories in the structure, the graph structure is represented as A[N, N], A represents the relationship between different skin disease categories; the second feature vector set data is used as a parameter and input into the graph convolutional neural network to obtain output vector set data; the output vector set data Y is represented as Y=ReLU(AXW+B), wherein the dimension of the output vector set data Y is [b, N, D1], D1 is the dimension of the second feature vector set data, W is the weight of the graph convolution layer, and B is the offset of the graph convolution layer; the output vector set data is processed by dimension change to obtain a network prediction value; the dimension of the network prediction value is [b, N], b is the number of parallel samples; A loss module, used to obtain cross entropy loss set data according to the network prediction value and the original label of the skin disease; and The second training module is used to train the training parameters in the classification model according to the cross entropy loss set data to obtain a classification model based on skin diseases.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the training method of the skin disease-based classification model as described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the training method of the skin disease-based classification model as claimed in any one of claims 1 to 4 are implemented.

Images