Research On Pigmented Skin Lesions Image Classification Based On Neural Architecture Search

Skin cancer is one of the main types of cancer caused by various skin diseases.Early accurate diagnosis and treatment can prevent the deterioration of benign skin cancer and significantly improve the survival rate of patients.However,the accurate identification of lesions is highly challenging due to factors such as visual similarities between different types of lesions(e.g.,melanoma and non-melanoma lesions),low contrast between lesions and skin,background noise,and artifacts.Compared with traditional machine learning methods,the deep learning model based on convolutional neural network(CNN)have been widely used in the task of automatically recognizing lesion diseases with high accuracy.However,to a large extent,designing an excellent CNN architecture require researchers have rich experience in neural network design and spend a lot of working time on parameter tuning.To solve these problems,this paper studies a neural architecture search(NAS)approach to automatically build a CNN network for image classification of pigmented skin lesions.Around this method,the main contributions of this thesis are as follows:(1)A neural network architecture search method based on macro operation mutation(MOM-NAS)is designed.In the process of search,the evolutionary algorithm based on macro operation mutation is used to iteratively change the operation type and connection mode of the parent cell,so that the macro operation is inserted into the child cell by the way of virus injection into host DNA,so as to continuously update the parent cell.Finally,the best cells searched are stacked to build a CNN for image classification of pigmented skin lesions.In addition,an improved cell-oriented search space containing micro operations and macro operations is designed,which makes the search space includes both basic micro operations such as convolution layers and these manually designed and sophisticated macro operations,expanding the scope of the searched neural network architecture.The experimental results show that the CNN searched by MOM-NAS method achieves the classification performance close to or better than that of state-of-the-art(SOTA)methods,and has a good generalization ability.(2)Further,considering that MOM-NAS approach regards the NAS as a discrete search space optimization problem,a large number of candidate neural network architectures need to be evaluated for performance.Therefore,a neural network architecture search method based on single path activation(SPA-NAS)is studied.It constructs the search space as an over parameterized neural network architecture that includes all paths,and each path is assigned an architecture parameter to represent the proportion intensity of the path.In order to avoid searching all paths,a single path activation strategy is proposed to prune the paths of the constructed over parameterized neural network architecture to obtain a more streamlined subarchitecture.When searching,gradient descent method is used to learn and optimize architecture parameters to obtain the best sub-architecture.Finally,the sub-architecture is stacked to construct the classification convolution neural network of pigmented skin lesions.Experimental results and analysis verify the effectiveness of the proposed method,which can obtain better classification results.