Automated Hair Segmentation in Dermoscopy Images with U-Net Based Approaches

Abstract

The presence of hair in dermoscopy images poses significant challenges for the diagnosis and analysis of skin lesions, particularly for melanoma, the most lethal type of skin cancer. Hair occlusion can significantly hinder the diagnostic process, making accurate segmentation critically important. This study evaluates the effectiveness of Convolutional Neural Network (CNN) based architectures U-Net, ResU-Net, and AttU-Net in the task of hair segmentation in dermoscopy images, including comparisons across various image resolutions. Simulation results indicate that the proposed system can detect and segment hair with mean scores of 0.7438 and 0.6211 for the Dice score and Jaccard index, respectively. The findings suggest that especially in high-resolution dermoscopy images, U-Net models significantly improve hair segmentation accuracy, potentially playing a critical role in early melanoma diagnosis.