An app to detect melanoma using deep learning: An approach to handle imbalanced data based on evolutionary algorithms
Pedro B. C. Castro, Breno Krohling, André G. C. Pacheco, Renato A. Krohling
Abstract
Skin cancer is a quite common type of cancer. Its incidence is more often in caucasian people and melanoma is the most lethal one. In order to increase patient prognosis, developing tools to assist early diagnosis is quite important. In the last few years, several methods have been proposed to deal with automated melanoma detection. Nonetheless, most of them are based only in dermoscopy images and/or do not take into account lesion clinical information. In this paper, we developed an ease and accessible mobile tool to assist melanoma detection. The app is linked to a convolutional neural network (CNN) trained on images collected from smartphones and lesion clinical information. Since the occurrence of melanoma is much smaller than other skin lesions, most of the datasets for this problem are imbalanced. To deal with this issue, we present an approach based on evolutionary algorithm to balance datasets. The proposed approach obtained promising results in comparison with related works with a balanced accuracy of 92% and a recall of 94%. However, it is a preliminary study, therefore these metrics may vary when applied to other datasets.