Abstract
Skin cancer remains one of the deadliest kinds of cancer, with a survival rate of about 18-20%. Early diagnosis and segmentation of the most lethal kind of cancer, melanoma, is a challenging and critical task. To diagnose medicinal conditions of melanoma lesions, different researchers proposed automatic and traditional approaches to accurately segment the lesions. However, visual similarity among lesions and intraclass differences are very high, which leads to low-performance accuracy. Furthermore, traditional segmentation algorithms often require human inputs and cannot be utilized in automated systems. To address all of these issues, we provide an improved segmentation model based on depthwise separable convolutions that act on each spatial dimension of the image to segment the lesions. The fundamental idea behind these convolutions is to divide the feature learning steps into two simpler parts that are spatial learning of features and a step for channel combination. Besides this, we employ parallel multidilated filters to encode multiple parallel features and broaden the view of filters with dilations. Moreover, for performance evaluation, the proposed approach is evaluated on three different datasets including DermIS, DermQuest, and ISIC2016. The finding indicates that the suggested segmentation model has achieved the Dice score of 97% for DermIS and DermQuest and 94.7% for the ISBI2016 dataset, respectively.