Dermoscopy Image Classification Based on StyleGAN and DenseNet201

Melanoma is considered one of the most lethal skin cancers. However, skin lesion classification based on deep learning diagnostic techniques is a challenging task owing to the insufficiency of labeled skin lesion images and intraclass-imbalanced datasets. It is thus necessary to utilize data augmentation methods based on generative adversarial networks (GANs) to assist skin lesion classification and help dermatologists reach more accurate diagnostic decisions. Moreover, insufficient samples can cause a low classification accuracy in a model by using deep learning in medical diagnosis and reduce the accuracy of skin lesion classification. To solve the above problems, this paper proposes a new skin lesion image classification framework based on a skin lesion augmentation style-based GAN (SLA-StyleGAN) according to the basic architecture of style-based GANs and DenseNet201. The proposed framework redesigns the structure of style control and noise input in the original generator and reconstructs the discriminator to adjust the generator to efficiently synthesize high-quality skin lesion images. We introduce a new loss function that reduces the intraclass sample distance and expands the sample distance between different classes, which can improve the balanced multiclass accuracy (BMA). The experimental results show that our classification framework performs well on the ISIC2019 dataset, and the BMA reaches 93.64%. The proposed method improves the accuracy of skin lesion image classification, assists dermatologists in determining and diagnosing different types of skin lesions, and analyzes skin lesions at different stages as well as those that are difficult to distinguish.

Automated summary

This study proposes a new skin lesion image classification framework based on SLA-StyleGAN, which improves the application of deep learning in skin lesion classification through data augmentation methods, increasing classification accuracy. Experimental results show that the framework performs well on the ISIC2019 dataset, with a BMA of 93.64%.