Style Transfer Technique for Dermatological Imaging of Skin Lesions in Various Phototypes

Abstract

In dermatological practice, accurate identification and classification of skin lesions in various skin phototypes are essential for early and effective diagnosis. In this context, advances in artificial intelligence have highlighted the potential of Convolutional Neural Networks (CNNs) as powerful tools for generating realistic medical images. This study focuses specifically on dermatological image generation by applying Style Transfer (ST) of seven skin lesions in six skin phototypes (Fitzpatrick scale) using the CNN model (VGG19). Once generated, the images are evaluated using two metrics: the structural similarity index (SSIM) and the Kullback-Leibler (KL) divergence. This work aims to improve visual data availability to support computer-aided diagnosis. Beyond extending the existing dataset with traditional data augmentation methods, we seek to enrich the quality and diversity of the images generated for each skin phototype. Seven hundred clinical images of seven common skin lesions were collected from the HAM10000 dataset, and 420 images were generated with the VGG19 model. Malignant and benign skin lesions were classified using the EfficienNet B0 and B1 models. The results suggest that the ST technique can be applied as an alternative method for diversifying skin phototypes in the HAM10000 data set.