Towards a Mobile and Fast Melanoma Detection System

Early detection of melanoma is crucial to avoid skin cancer deaths, but only with the recent advances of deep convolutional neural networks architectures, such as MobileNet, it is possible to create a reliable enough system to detect melanoma, that can be implemented on resource constrained environments such as mobile phones or embedded systems. With this aim, this work assesses the performance of the implementation of an early melanoma recognition system using MobileNet trained from the HAM10000 database. Besides, we explain in detail two strategies to improve melanoma classification task, i.e., data augmentation on an unbalanced dataset and a multiclass approach to address a binary classification problem. Numerical results in terms of AUC metric and ROC curves corroborate the validity of our model. The performance of the proposed model is also compared to the average dermatologist performance.