Multi-Dataset Training for Skin Lesion Classification on Multimodal and Multitask Deep Learning

Abstract

According to the World Health Organization, skin cancer represents approximately one third of every diagnosed cancer, reaching over 3 million cases over the world, annually. Similar to other types of cancer, though, early diagnosis is key for a good outcome, and computer-aided diagnosis has shown great promise in such task. In this paper we improve the results of previous work on skin lesion diagnosis by using a deep convolutional neural network trained on multimodal data, namely macroscopic and dermoscopic image and metadata. For a deep learning approach is important to have a large number of samples, which EDRA dataset does not present. We have improved the results of previous work in the field of multimodal and multitasking for skin lesion classification by performing transfer learning using similar datasets, which are predicting different skin conditions. By pre-training on datasets which belong to a similar domain, the network learns useful features which enhances the performances of the network.