Dermoscopic Image Analysis for Skin Lesion Classification Using Dense Convolutional Network

Abstract

Skin cancer has become a more common disease in most parts of the world and accounts for high-mortality rates. Accurate diagnosing of skin cancer or lesion by observing or with hand-crafted methods is a more challenging task for expert dermatologists. Therefore, an automated and intelligent system is essential for the accurate detection of lesions. In this work, we have used a dense convolutional network (DenseNet) for automated skin lesions classification of seven types of skin lesions comprising of basal cell carcinoma, melanoma, melanocytic nevi, actinic keratosis/bowens disease, dermatofibroma, benign keratosis, and vascular lesion. The dense convolutional network used comprised of 5 dense blocks and each dense block comprised of batch normalization layer, ReLU, and two convolution layers which empower the maximum information flow between layers. The Softmax loss combined with image feature similarity restraint is further used to minimize the misclassification loss and to preserve the intra-class feature similarity scatter. We evaluated our model on the HAM10000 dataset, which consists of 10,015 images of seven skin lesion types. Our method achieved AUC of 0.96 and a precision of 0.92 for melanocytic nevi. Classification accuracy of the model is 88.86% in validation set which is approximately equivalent to trending performance of traditional models in skin lesion classification.