Adaptive dermascopy application using machine learning

Abstract

Skin cancer is the most lethal because skin cells develop abnormally. Finding skin cancer early is very important and may help stop some kinds of skin cancer, like melanoma and focal cell carcinoma. Early detection and classification of skin cancer are difficult and costly. Recurrent networks and ConvNets can automatically extract complex data. This paper proposes to use a handmade features-based multi-layer perceptron and a cascaded ensembled network to upgrade ConvNet models. This convolutional neural network model detects non-handmade picture qualities and generates features like color moments and material properties. With ensembled DL, accuracy increased from 85.3% with convolutional neural networks to 98.3%.