Melanoma Detection from Dermatoscopic Images using Deep Convolutional Neural Networks

A transfer learning-based architecture for melanoma detection from dermatoscopic images is presented. Ten pre-existing well-known deep convolutional neural networks were re-trained to detect melanoma using automatically preprocessed and segmented skin images. Preprocessing consisted of automatic hair removal from the dermatoscopic images and segmentation consisted of detection of the skin lesion area. The evaluation dataset used consisted of lesions of several skin pathologies including melanoma and the experimental results showed that the best performing re-trained deep convolutional neural network model was ResNet-101 with melanoma detection accuracy equal to 97.72% and sensitivity equal to 85.18%.