Melanocytic Pigmented Skin Lesion Detection and Classification using Hybrid Deep Features based on Fully Convolutional Network

One of the life-threatening skin cancers around the world is "Malignant Melanoma", which is difficult to treat and can be fatal. But it can be curable when detected and treated early. Detecting skin lesions whether it is Malignant melanoma or not in the early stage is a crucial task. So, in this paper we proposed a method which involves two stages of melanoma detection, Feature Classification and Extraction. Here for feature extraction two Fully convolutional networks (FCNs) are incorporated for improving the accuracy of feature extraction. And then classification is done by using Random Forest/SVM. To find the best Fully Convolutional Network s for feature extraction process we select five best architectures which can alone perform well, and experimented on all possible combinations of two of those Fully Convolutional Networks for feature extraction. The architectures we used are VGG16, VGG19, ResNet50, AlexNet, and GoogleNet. And the best result was given by VGG16 and ResNet50 with a promising accuracy of 92.3% when used in the International Society for Biomedical Imaging 2016 dataset.