Melanoma Screening Using Deep Neural Networks

Deep learning algorithms, in particular convolutional neural networks, have rapidly become a methodology of choice for analyzing images. This choice arises due to fact that Deep Learning reduces the feature engineering task which is very important for automatic analysis of various kinds of images including medical images. In this work we have implemented various Deep Convolutional Neural Networks architectures for the process of binary classification task efficient enough in predicting either the input RGB image of the skin lesion is melanoma or not. Diagnosing the skin lesion is the first step towards its treatment. In this work we have assimilated two approaches on the pretrained Convolutional Neural Networks on the ImageNet dataset. Firstly, we use the transfer learning approach without any fine tuning and obtained features are fed for linear classification task. Secondly, we use the transfer learning approach with fine tuning. The adopted method requires no lesion segmentation or any other higher level image processing tasks. Finally, we have also developed a real time android application to perform the binary classification task. The result obtained using our android application were enthusiastic.