基于神经架构搜索的色素性皮损图像分类方法

Pigmented Skin Lesions Image Classification Based on Neural Architecture Search

针对卷积神经网络设计高度依赖专家经验、需要大量参数调优和效率低的问题,提出了一种基于单路径激活搜索策略的神经架构搜索方法(SPA-NAS),并应用于色素性皮损图像分类。该方法将搜索空间构建为一个过参数化神经网络架构,该架构包含了所有的路径,并且每条路径都被分配一个架构参数以表示路径的占比强度。为了避免搜索所有路径,提出了一种单路径激活策略对构建的过参数化神经网络架构进行路径剪枝,以得到一个更加精简的子架构。搜索时,采用梯度下降法学习和优化架构参数,得到最佳子架构。最后,采用子架构堆叠方式构建色素性皮损图像分类卷积神经网络。实验表明,该方法自动构建的卷积神经网络取得了比Dilated-VGG19和ARL-CNN等SOTA方法更高的分类准确性,在ISIC2017和HAM10000数据集上的平均敏感度分别为62.4%和69.8%。

Because design of convolutional neural network(CNN)requires abundant expert experience,considerable parameter tuning and low efficiency,we propose a neural architecture search approach based on single path activation search strategy(SPA-NAS)applied to the image classification of pigmented skin lesions.This method constructs the search space as an hyperparameter neural network architecture,which contains all paths,and each path is assigned an architecture parameter to represent the proportion of the path.In order to avoid searching all paths,a single-path activation strategy is proposed to prune the constructed hyperparameter network architecture to obtain a more streamlined child architecture.When searching,the gradient descent method is used to learn and optimize the architecture parameters to search the best child architecture.Finally,a CNN for pigmented skin lesion image classification is constructed by using child architecture stacking.The test results showed that the CNN built with the proposed approach was more accurate than state-of the-art(SOTA)approaches such as Dilated-VGG19 and ARL-CNN in terms of image classification.The average sensitivity of the proposed method on the ISIC2017 and HAM10000 datasets was 62.4%and 69.8%,respectively.