Classification of Melanoma Based on Clustering Channels RGB of Skin Scan Using Adaptive Light Weight Deep Learning System

Abstract

Melanoma is considered to be a type of skin cancer that is characterised by symptoms of
poor prognostic responses. In this paper, we propose a deep learning method using DCNNs, modifying the
output layer and enhancing the features of skin scan images collected from Kaggle to be distinguished into
two groups: melanoma and non-melanoma cells. The proposed modified three types of DCNN (MobilNetv2, ResNet-18 and Squeeze Net) models have been tested in two experiments. In the first, the obtained values
of training accuracy are (93, 95 and 91) % and the testing accuracy is (90.09, 90.54 and 90.4) %, using
original datasets only. In the second experiment, the obtained values of training accuracy are (99.7, 96.3
and 92) % and the testing accuracy is (94.41, 94.14 and 91.43) %, The experimental findings show that the
model utilized produces enhanced photos with more accuracy than original images.
Keywords: Deep Learning (DL), Light Weight Convolutional Neural Networks (LWCNN),
Melanoma Classification