Generating Skin Lesion Hair Mask Pattern using Modified Pro-Gan

Abstract

A hair mask in dermoscopy refers to the appearance of hair follicles as dark lines surrounded by a lighter halo. The occlusion of skin lesions by hair affects the accuracy of disease detection algorithms. Hair masks can be used not only for hair removal but also for hair augmentation and as a realistic hair simulator. In previous work, hair masks are generated by pre-segmentation or manual drawing. Here a new approach to hair mask generation is used employing a modified version of the generative model Pro-GAN to facilitate the generation of hair mask patterns. The main contribution of this paper lies in generating large-sized hair mask patterns measuring 512x512 pixels using a generative model, coupled with modifications to the pro-GAN. These modifications include integrating minibatch standard deviation, pixel normalization, equalized learning rate, gradient penalty, and enhancements to the training loop. Initially, the generator produces small-sized images from noise. Each iteration progressively increases the image size until large images are attained. A customized dataset is used in this research and dermoscopic images of HAM10,000 have been used for hair mask extraction using U-Net. The performance of the proposed Pro-GAN in generating hair mask patterns is objectively evaluated using key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Perceptual Loss, yielding respective values of 31.473, 0.926, and 0.484. A pathway for hair mask generation has been established through this work.