Skin_Hair Dataset: Setting the Benchmark for Effective Hair Inpainting Methods for Improving the Image Quality of Dermoscopic Images

Abstract

Dermoscopic images are often contaminated by artifacts including clinical pen markings, immersion fluid air bubbles, dark corners, and most importantly hair, which makes interpreting them more challenging for clinicians and computer-aided diagnostic algorithms. Hence, automated artifact recognition and inpainting systems have the potential to aid the clinical workflow as well as serve as an preprocessing step in the automated classification of dermoscopic images. In this paper, we share the first release of a public dermoscopic image dataset with hair artifacts which can be accessed here https://skin-hairdataset.github.io/SHD/ . The Skin_Hair dataset contains over 252 dermoscopic images including artificial hair and will be expanded over time. Furthermore, we present the primary results of applying machine learning algorithms and GAN based architectures to the hair inpainting problem in dermoscopic images. We envision that these results will serve as a benchmark for researchers who might work on the hair detection and reconstruction tasks with this dataset in the future. In this work, we present a skin lesion image dataset based on the ISIC dataset containing dermoscopic images, images containing artificial hairs and the corresponding ground-truth masks. Furthermore, we use four hair inpainting methods including Navier-Stokes, Telea, Hair_SinGAN and R-MNet architectures which we evaluate using image quality assessment metrics MSE, PSNR, UQI and SSIM. The R-MNet architecture achieved the highest SSIM score of 0.960.