Epidermis Segmentation in Melanoma Whole Slide Images: A Comparative Analysis of Deep Learning Architectures

Skin cancer is one of the most common types of cancer worldwide, and its incidence rate is increasing day by day. Melanoma, a type of skin cancer, represents a small proportion of all skin cancers, although it is responsible for half of all skin cancer-related deaths. Early staging is a factor that significantly affects survival in cancer treatment. However, traditional methods of examining tissue under a microscope and measuring the depth of invasion, the deepest point of the tumor, require substantial time and specialization. This can lead to delayed early diagnosis and some variability between specialists that affects the accuracy of the diagnosis. Automating this process could shorten the diagnostic time and improve diagnostic accuracy. Therefore, the first step to determine tumor stage in melanoma is to automate epidermis segmentation. This study conducted an analysis of model performance on 69 melanoma whole slide image (WSI) samples employing various architectures, including base UNet, UNet++, UNet3+, Efficient-UNet, Swin-UNet, and TransUNet. According to our findings, the Trans-UNet architecture achieved the highest success ratio among the 33 test WSI samples, with a Dice Coefficient Score of 92%.