Efficacy of Deep Learning Approach for Automated Melanoma Detection

Melanoma is considered the most perilous type of skin cancer. However, differentiating lesions of melanoma from non-melanoma lesions has proven difficult. For this purpose, many Computer-Assisted Automatic Diagnosis Technologies are proposed in the past. Due to the complicated visual properties of images of skin lesions, which include inhomogeneous fuzzy boundaries and features, they have been limited in their performance. The use of computer-aided systems to diagnose skin lesions is becoming more common. Researchers have recently expressed a growing interest in developing computer-assisted diagnosis methods. In this research, a deep learning system for automated melanoma segmentation and lesion classification addresses these constraints. For effective learning and feature extraction, an upgraded encoder-decoder with sub-networks is interconnected, bringing the encoder feature maps relatively closer to the feature maps of the decoder. For pixel-wise categorization of melanoma lesions, the system uses a multi-scale and multi-stage technique, as well as a soft-max classifier. A novel approach called Lesion-classifier divides lesions of skin into non- melanoma and melanoma on the basis of the findings of pixel-wise classification. These studies are on the ISIC 2017 dataset, with dice coefficient and accuracy of 92 and 95 percent, respectively, while on the PH2 datasets, the dice coefficient and accuracies are as 93 and 95 percent.