A Convolutional Neural Network for skin lesion classification in Dermoscopic Images using discrete wavelet transform

The objective of this paper is to investigate a new technique for skin lesion classification based on discrete wavelet transforms and convolutional networks. More precisely, for preprocessing the images, two 2D family wavelets and Low-Low (LL), High-High (HH) bands were used to generate subimages for each skin lesion image as melanoma and normal nevi. On each subimage datasets of nevi and melanoma, five convolutional neural networks (CNNs) were tested. The experiments were conducted on the public Mednode dataset which contains 170 images (70 melanoma and 100 nevi cases). The accuracy, precision, recall, and F1-score metrics were extracted from confusion matrix for each test performed on the proposed CNN architecture. All metrics show that LL subimages are recommended to be used in classifying melanoma vs. nevi.