Skin focus classification method based on improved Adaboost and lightweight convolutional neural network

Abstract

The invention discloses a skin lesion classification method based on an improved Adaboost and lightweight convolutional neural network, and aims to solve the problems of high calculation cost of an existing large model and insufficient precision of a lightweight model under the conditions that the number of skin lesion image samples is limited and the sample category proportion is unbalanced. According to the method, the lightweight convolutional neural network is used as a base learning device, and the weight parameters of the training samples are updated through the improved Adaboost algorithm, so that the lightweight model pays more attention to the samples with classification errors, and the classification capability of minority class samples is improved. According to the method, based on relatively low running resources and hardware cost, the classification accuracy is remarkably improved on the public data set of the skin lesions, intelligent and rapid screening and judgment are provided for clinical diagnosis and treatment of the skin lesions, and the method is suitable for portable and mobile terminal medical equipment and has a good application prospect.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a skin lesion classification method based on an improved Adaboost and lightweight convolutional neural network. Background Art

[0002] Skin diseases are a common health problem worldwide, and timely and accurate diagnosis is crucial for the treatment and recovery of patients. However, image samples of skin lesions are usually few, and it is very difficult to obtain a large number of high-quality image samples, especially in the case of some rare skin diseases. In addition, the uneven distribution of sample numbers between different categories of skin lesions poses a great challenge to the training and performance of classification methods, and easily leads to insufficient recognition ability of classification models for minority class samples.

[0003] At present, convolutional neural networks have achieved remarkable results in image classification tasks and are widely used in the classification of skin lesion images. However, many existing convolutional neural network models have complex structures, large parameters, and high computing resource requirements. Although such oversized models can improve classification accuracy to a certain extent, they also significantly increase computing costs and time overhead. In order to address these problems, lightweight convolutional neural networks have gradually become a research hotspot. Lightweight models can run on limited hardware resources by reducing the number of parameters and the amount of calculation, thus reducing the cost of medical examinations. However, while lightweight models reduce computing resource consumption, they are often accompanied by a decrease in classification accuracy. Especially when dealing with datasets with few skin lesion samples and unbalanced categories, the performance of lightweight models is difficult to meet the requirements of practical applications.

[0004] To address this problem, the present invention proposes a skin lesion classification method based on an improved Adaboost and lightweight convolutional neural network, which is used in the classification task of skin lesion images. This method can improve the accuracy and robustness of classification while maintaining low computational complexity, and has high practical value and promotion prospects. Summary of the invention

[0005] In view of the problems and shortcomings of the above-mentioned existing methods, the technical problem to be solved by the present invention is to provide a skin lesion classification method based on an improved Adaboost and a lightweight convolutional neural network, which improves the classification accuracy of the lightweight model while maintaining a low computational complexity, thereby ensuring the reliability and accuracy of the skin lesion image classification results.

[0006] To achieve the above-mentioned invention object, the present invention adopts the following technical scheme: a skin lesion classification method based on improved Adaboost and lightweight convolutional neural network, comprising the following steps:

[0007] Step 1: Given a skin lesion dataset, which includes a training set, a validation set, and a test set, set the number of integrated iterative training times T, and initialize the sample weight parameters of the training set;

[0008] Step 2, use The fold cross validation method is to perform fold, and select one sub-data set in turn as fold cross validation sub-validation set, and the rest The sub-datasets are Fold cross validation sub-training set, get A combination of sub-training set and sub-validation set;

[0009] Step 3, by fold cross validation method, training A lightweight convolutional neural network model;

[0010] Step 4: Select the model with the highest classification accuracy on the validation set and adjust its The sample weights of the fold cross validation sub-validation set;

[0011] Step 5: fold cross validation sub-validation set and the rest The sub-training sets are merged and shuffled to construct a new training set;

[0012] Step 6: Repeat steps 2 to 5 until the number of iterations is reached. , according to the accuracy of the base learner on the validation set, from In the iterative training, the base learner with the best classification accuracy on the validation set is selected, and the final strong learner is obtained through weighted sum.

[0013] Furthermore, in step 1, it also includes:

[0014] Given a training set:

[0015]

[0016] in Represents a sample, express Class labels, initially, each sample has equal weight:

[0017]

[0018] in .

[0019] Furthermore, in step 3, it also includes:

[0020] The lightweight convolutional neural network models are GhostNet, ShuffleNetV2, and MobileNetV3-Large;

[0021] In each training iteration During the cross-validation process, a network structure is selected from GhostNet, ShuffleNetV2 and MobileNetV3-Large for model training. After one iteration, the An independently trained model.

[0022] Furthermore, in step 4, it also includes:

[0023] The specific steps for adjusting the sample weights of the fold cross-validation sub-dataset are as follows:

[0024] S1: Round ensemble iterative training Among the models, the model with the highest classification accuracy on the validation set is selected as the Optimal Basis Learner for Iterative Learning ;

[0025] S2: Calculate the optimal base learner above in its Weighted classification error on the 1-fold cross-validation sub-validation set:

[0026]

[0027] in, is the indicator function, when When it is, it returns 1, otherwise it returns 0;

[0028] S3: Calculate the weight of the optimal base learner based on the weighted classification error :

[0029]

[0030] in represents the weight of the optimal base learner in the strong learner;

[0031] S4: Update sample weights , the update formula is:

[0032]

[0033] To keep the weights normalized, Perform normalization:

[0034]

[0035] Furthermore, in step 6, it also includes:

[0036] Strong Learner is the weighted sum of the best base learners in all iterative training:

[0037]

[0038] in It is The weight of the best base learner in the iterative training, It is The optimal basis learner in the training iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the improved Adaboost skin lesion classification method based on a lightweight convolutional neural network of the present invention.

[0040] Figure 2 Flowchart of iterative training for Adaboost ensemble learning.

[0041] Figure 3 This is a flowchart of the weight adjustment process of the sub-validation set of the improved Adaboost cross-validation based on the lightweight convolutional neural network.

[0042] Figure 4 This is a comparison chart of the relationship between the floating-point calculation amount, parameter amount and accuracy of the lightweight model.

[0043] Figure 5 Accuracy statistics and comparison chart of a single lightweight model and an improved Adaboost integrated lightweight model on the test set. DETAILED DESCRIPTION

[0044] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. 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 making creative work are within the scope of protection of the present invention.

[0045] The present invention provides a skin lesion classification method based on improved Adaboost and lightweight convolutional neural network. Figure 1 A flow chart showing an embodiment of the present invention includes:

[0046] Step 1: Given a skin lesion image dataset, take the ISIC2018_TASK3 public dataset as an example. The dataset includes a training set, a validation set, and a test set. Record the number of training set samples m and the number of training set samples ,in Represents a sample, express Skin lesion labels. Initially, each sample has an equal weight:

[0047]

[0048] in .

[0049] Set the number of ensemble learning iterations , ;

[0050] Step 2, such as Figure 2 The Adaboost ensemble learning iterative training flow chart shown in the figure uses the 5-fold cross-validation method to divide the training set into 5 folds, selects one sub-dataset as the 5-fold cross-validation sub-validation set, and the remaining four sub-datasets as the 5-fold cross-validation sub-training sets, to obtain five sub-training set and sub-validation set combinations;

[0051] Step 3: Select a lightweight convolutional neural network model structure from GhostNet, ShuffleNetV2, and MobileNetV3-Large, and train five lightweight convolutional neural network models using a 5-fold cross-validation method.

[0052] In one iteration, each model has the same structure but the models are trained independently;

[0053] Step 4: Calculate the classification accuracy of the five models on the validation set, select the model with the highest classification accuracy, and adjust the sample weight of its 5-fold cross-validation sub-validation set. The specific steps are as follows: Figure 3 The weight adjustment process of the sub-validation set of the improved Adaboost cross-validation based on the lightweight convolutional neural network is shown in the figure, including:

[0054] S1: Among the five models in the round of ensemble iterative training, the model with the highest classification accuracy on the validation set is selected as the first Optimal Basis Learner for Iterative Learning ;

[0055] S2: Calculate the optimal base learner above Weighted classification error on its 5-fold cross-validation sub-validation set:

[0056]

[0057] in, is the indicator function, when When it is, it returns 1, otherwise it returns 0;

[0058] S3: Calculate the weight of the optimal base learner based on the weighted classification error :

[0059]

[0060] in represents the weight of the optimal base learner in the strong learner;

[0061] S4: Update sample weights , the update formula is:

[0062]

[0063] To keep the weights normalized, Perform normalization:

[0064]

[0065] Step 6: Repeat steps 2 to 5 until the number of iterations is reached. According to the accuracy of the base learner on the validation set, the base learner with the best classification accuracy on the validation set is selected from the three iterative trainings, and the final strong learner is obtained by weighted sum :

[0066]

[0067] in It is The weight of the best base learner in the iterative training, It is The optimal basis learner in the training iteration.

[0068] The comparison of the floating-point calculation amount, parameter amount, and accuracy of the lightweight model on the skin lesion validation set and test set is shown in the figure below. Figure 4 As shown, a single trained lightweight model, a lightweight model trained by an improved Adaboost ensemble, a lightweight model using a Bagging ensemble method, an improved Adaboost 2 ensemble lightweight model under the same network structure, an improved Adaboost 3 ensemble lightweight model under the same network structure, and an improved Adaboost ensemble lightweight model under a hybrid network structure are compared.

[0069] Compared with the lightweight model with single training, the improved Adaboost ensemble learning method of the present invention is While maintaining a low number of parameters and computational complexity, the accuracy of the model is further improved, which can effectively improve the generalization ability of the model and the classification accuracy of skin lesion images.

[0070] The accuracy statistics and comparison of the single trained lightweight model and the improved Adaboost integrated lightweight model on the test set are as follows: Figure 5 As shown, it reflects that the present invention significantly improves the accuracy of skin lesion image classification.

Claims

1. A skin lesion classification method based on improved Adaboost and lightweight convolutional neural network, characterized by: Step 1: Given a skin lesion dataset, which includes a training set, a validation set, and a test set, set the number of integrated iterative training times T, and initialize the sample weight parameters of the training set; Step 2, use The fold cross validation method is to perform fold, and select one sub-data set in turn as fold cross validation sub-validation set, and the rest The sub-datasets are Fold cross validation sub-training set, get A combination of sub-training set and sub-validation set; Step 3, by fold cross validation method, training A lightweight convolutional neural network model; Step 4: Select the model with the highest classification accuracy on the validation set and adjust its The sample weights of the fold cross validation sub-validation set; Step 5: fold cross validation sub-validation set and the rest The sub-training sets are merged and shuffled to construct a new training set; Step 6: Repeat steps 2 to 5 until the number of iterations is reached. , according to the accuracy of the base learner on the validation set, from In the iterative training, the base learner with the best classification accuracy on the validation set is selected, and the final strong learner is obtained through weighted sum.

2. The skin lesion classification method according to claim 1, characterized in that: In step 1, given the training set ,in Represents a sample, express Class labels, initially, each sample has equal weight: in .

3. The skin lesion classification method according to claim 1, characterized in that: The lightweight convolutional neural network models in step 3 are GhostNet, ShuffleNetV2, and MobileNetV3-Large.

4. The skin lesion classification method according to claim 1, characterized in that: The steps for adjusting the sample weights of the cross-validation sub-dataset in step 4 are: S1: Round ensemble iterative training Among the models, the model with the highest classification accuracy on the validation set is selected as the Optimal Basis Learner for Iterative Learning ,in ; S2: Calculate the optimal base learner above In its Weighted classification error on the 1-fold cross-validation sub-validation set: in, is the indicator function, when When it is, it returns 1, otherwise it returns 0; S3: Calculate the weight of the optimal base learner based on the weighted classification error : in represents the weight of the optimal base learner in the strong learner; S4: Update sample weights , the update formula is: To keep the weights normalized, Perform normalization: 。 5. The skin lesion classification method according to claim 1, characterized in that: The strong learner in step 6 is the weighted sum of the best base learners in all iterative training: in It is The weight of the best base learner in the iterative training, It is The optimal basis learner in the training iteration.

Images