Transfer Learning Approach and Analysis for Skin Cancer Detection

Abstract

Climate and other environmental humans are being more exposed to the Sun’s ultraviolet(UV) rays. This exposure has increased the chances of humans developing skin cancer. The early detection of skin cancer is vital for curing this disease. The recent development in Neural Networks and state of the art deep learning techniques have an essential role in detecting skin cancer. In this paper, we extensively go through the deep neural network architectures and compare their results. We have tried to propose an analysis report using the state-of-the-art transfer and comparing their prediction capabilities to detect Skin Cancer by testing it on the HAM1000 dataset. We propose a solution by using transfer learning which have very heavily trained dense trainable layers with weights from imagenet. Finally, we show that among all the models, Resnet50, U-net, Densenet121, MOBILENETV1, InceptionV3 and U Net achieved the highest accuracy with 91% test accuracy becoming the top approach. To draw comparison, we have also used CNN architecture models which was previously structured using Keras Sequential API. We see the newer dense transfer learning models provide better testing accuracy. Hence, this paper aims to detect skin in its early stage to start treatment and hopefully save valuable lives in the process.