A GAN-based image synthesis method for skin lesion classification
A GAN-based image synthesis method for skin lesion classification
复制标题
DOI：
10.1016/j.cmpb.2020.105568
复制
发表时间：
2020-10-01
影响因子：
6.1
通讯作者：
Xue, Yongbo
中科院分区：
文献类型：
--
作者：
Qin, Zhiwei;Liu, Zhao;Xue, Yongbo
背景和目的：皮肤癌的类型很多，黑色素瘤是其中最致命的一种。皮肤镜是筛查黑色素瘤和其他皮肤病变的重要成像技术。然而，由于标签数据的稀缺性和类别不平衡的数据集，基于计算机辅助诊断技术的皮肤病变分类是一个具有挑战性的任务。为了帮助皮肤科医生做出更准确的诊断决策，有必要将基于产生式对抗网络(GANS)的数据增强技术应用于皮肤病变分类。方法：建立了基于GAN的数据增强技术提高皮肤病变分类性能的全过程。首先，根据基于风格的GANS的基本体系结构，提出了基于皮肤病变风格的GANS。该模型对原生成器中的风格控制和噪声输入结构进行了改进，同时对生成器和鉴别器进行了调整，从而有效地合成了高质量的皮肤病变图像。在图像分类方面，采用传递学习的方法，将分类器建立在预先训练好的深度神经网络上。结果：本文提出的基于皮肤病变风格的GAN模型经过了初始评分(IS)、Frechet初始距离(FID)、准确度和召回率的评估，在这些量化评价指标上均优于其他比较的GAN模型。通过将合成图像添加到训练集中，在国际皮肤成像协作(ISIC)2018挑战赛的数据集上，准确率、敏感度、特异度、平均精确度和均衡多类准确率分别为95.2%、83.2%、74.3%、96.6%和83.1%，与CNN模型相比分别提高了1.6%、24.4%、3.6%、23.2%和5.6%。结论：基于皮肤病变样式的GANS能够高效地生成高质量的皮肤病变图像，从而提高分类模型的性能。该工作为基于深度学习的医学图像分析提供了有价值的参考。(C)2020爱思唯尔B.V.保留所有权利。
Background and Objective: There are many types of skin cancer, and melanoma is the most lethal one. Dermoscopy is an important imaging technique to screen melanoma and other skin lesions. However, Skin lesion classification based on computer-aided diagnostic techniques is a challenging task owing to the scarcity of labeled data and class-imbalanced dataset. It is necessary to apply data augmentation technique based on generative adversarial networks (GANs) to skin lesion classification for helping dermatologists in more accurate diagnostic decisions.Methods: A whole process of using GAN-based data augmentation technology to improve the skin lesion classification performance has been established in this article. First of all, the skin lesion style-based GANs is proposed according to the basic architecture of style-based GANs. The proposed model modifies the structure of style control and noise input in the original generator, adjusts both the generator and discriminator to efficiently synthesize high-quality skin lesion images. As for image classification, the classifier is constructed on the pretrained deep neural network using transfer learning method. The synthetic images from the proposed skin lesion style-based GANs are finally added to the training set to help train the classifier for better classification performance.Results: The proposed skin lesion style-based GAN has been evaluated by Inception Score (IS), Frechet Inception Distance (FID), Precision and Recall, and is superior to other compared GAN models in these quantitative evaluation metrics. By adding the synthesized images to the training set, the main classification indicators like accuracy, sensitivity, specificity, average precision and balanced multiclass accuracy are 95.2%, 83.2%, 74.3%, 96.6% and 83.1% on the dataset of International Skin Imaging Collaboration (ISIC) 2018 Challenge, which have been improved by 1.6%, 24.4%, 3.6%, 23.2% and 5.6% respectively compared to the CNN model.Conclusions: The proposed skin lesion style-based GANs can generate high-quality skin lesion images efficiently, leading to the performance improvement of the classification model. This work provides a valuable reference for medical image analysis based on deep learning. (c) 2020 Elsevier B.V. All rights reserved.