Analyzing Digital Image by Deep Learning for Melanoma Diagnosis

Abstract

Image classi cation is an important task in many medical applications, in order to achieve an adequate diagnostic of di erent le- sions. Melanoma is a frequent kind of skin cancer, which most of them can be detected by visual exploration. Heterogeneity and database size are the most important di culties to overcome in order to obtain a good classi cation performance. In this work, a deep learning based method for accurate classi cation of wound regions is proposed. Raw images are fed into a Convolutional Neural Network (CNN) producing a probability of being a melanoma or a non-melanoma. Alexnet and GoogLeNet were used due to their well-known e ectiveness. Moreover, data augmentation was used to increase the number of input images. Experiments show that the compared models can achieve high performance in terms of mean ac- curacy with very few data and without any preprocessing.