Deep Neural Networks for Melanoma Detection from Optical Standard Images using Transfer Learning

摘要

Melanoma is the most serious type of skin cancer. Early detection of melanoma can increase survival rates. Recently, the emergence of deep learning approaches for medical image analysis has improved the development of computer-aided diagnosis systems that can help the expert to make a better decision about patient health. While dermoscopic image requires a specialized device for image acquisition, optical skin images acquired with a standard camera is an attractive modality for melanoma early diagnosis, yielding acceptable lower detection rates, reported in the state of the art, at a reasonable cost. In this paper, we consider this modality for improving the detection accuracy of melanoma. However, there are no large available datasets of optical lesion images as for dermatoscopic modality; for training deep networks. Thus, accordingly, we propose here a transfer learning paradigm that helps to overcome such limitations. A new CNN architecture and a set of deep learning networks have been trained. The best detection rates are obtained by the convolutional neural network, yielding a detection rate of 97%.