Screening Feasibility and Comparison of Deep Artificial Neural Networks Algorithms for Classification of Skin Lesions

Abstract

Deep convolutional neural networks (CNNs) have proven its potential for many tasks related to object identification and classification. This study aims to show the performance of several convolutional neural networks architectures applied to the diagnosis and screening of skin lesions in patients using different training techniques: Random weights initialization, feature extraction and extending model. A dataset of 1000 clinical images proven by biopsy or consensus among specialists were the examples applied at the various architectures which were end-to-end trained from images directly, using only pixels and disease labels as inputs. The predictions provided from the models intended to claim whether the lesion could be treated by doctors with images only on a teledermatology approach or if it is necessary to prescribe a biopsy or referral to a face-to-face consultation. The model can also tell the urgency of the case and the group of diseases which that lesion belongs to. Performances of deep neural networks in all proposed tasks demonstrated that artificial intelligence has the potential to perform the screening of skin lesions with a level of competence comparable to dermatologists. It is projected 6.3 billion signatures of smartphone by the year 2021 [38]. Therefore, deep neural networks incorporated in mobile devices can amplify the reach of dermatologists outside their offices providing universal low-cost access to dermatological diagnostics.