Skin Lesion Image Generation and Classification Based on SAE Framework
赵宇航 1闫天星 1伊力哈木·亚尔买买提1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 新疆大学电气工程学院,新疆 乌鲁木齐 830047
折叠
摘要
针对现有算法对于皮肤病变数据集形态复杂、各类样本不平衡导致分类精度低、特征提取能力不强等问题,提出了一种基于皮肤病变图像的风格对抗生成网络与设计的ECA-ConvNext分类网络结合的皮肤病变图像生成分类方法(SL-style-GAN2 and ECA-ConvNeXt Frame,SAE).首先,对风格对抗生成网络中对生成器重新设计,并且对判别器部分进行重构,使判别器可以同时为生成器提供局部和全局信息,从而生成更好的样本图片以供后续分类模型得到更好的效果.之后选用ConvNeXt-T为分类基础网络,设计了深层特征提取模块(Deep information extraction module,DIEM)使通道和权值之间直接联系,提高网络特征提取能力,从而提高模型精度.最后,在ISIC 2018 数据集上进行实验,实验结果表明,分类准确率达到94.0%,比原始ConvNeXt提高了 4.5%.
Abstract
A skin lesion image generation and classification method(SL-styleGAN2 and ECA ConvNeXt Frame,SAE)was proposed,which combines a style adversarial generation network based on skin lesion images with a de-signed ECA ConvNext classification network,to address the problems of low classification accuracy and weak feature extraction ability caused by the complex morphology and imbalanced samples of existing algorithms on skin lesion datasets.Firstly,the generator was redesigned in the style confrontation generation network,and the discriminator part was reconstructed,so that the discriminator could provide both local and global information to the generator,so as to generate better sample pictures for the subsequent classification model to get better results.Then,the ConvNeXt-T is selected as the classification basis network,and the Deep information extraction module is designed to make the chan-nel and weight directly related to improve the feature extraction ability of the network,so as to improve the accuracy of the model.Finally,the experimental results on ISIC 2018 dataset show that the classification accuracy reaches 94.0%,which is 4.5%higher than the original ConvNeXt.
关键词
图像处理/皮肤病变分类/生成对抗网络/高效通道注意力/计算机辅助识别
Key words
Image processing/Classification of skin lesions/Generative adversarial networks/Efficient channel at-tention/Machine-aided cognition
维普
万方数据