Skin cancer is one of the most prevalent cancers in the world, and the likelihood of patient survival is greatly enhanced by early and accurate diagnosis, especially in the case of melanoma. The typical first line diagnostic process is visual screening followed by dermoscopic/thermoscopic (also called dermatoscopic) imaging; manual interpretation is still time consuming and can vary from one observer to another, even among those who are trained in this technique. This paper reviews and synthesizes the use of deep convolutional neural networks (CNNs) applied to transfer learning on dermoscopic images from the International Skin Imaging Collaboration (ISIC) Archive for automated skin cancer detection and classification. The following widely used CNN backbones are presented: ResNet50, DenseNet121, Inception-v3, and EfficientNet, and are compared in terms of accuracy, sensitivity, specificity, F1-score and area-under-the-ROC-curve (AUC) for a standard preprocessing, augmentation and transfer-learning pipeline. The results presented in reported literature for HAM10000 in comparable transfer-learning settings range from 80.05% (Xception) to 94.29% (DenseNet121) in terms of classification accuracy; with attention and ensemble extensions of these backbones having first rank results in the ISIC 2019 Challenge (63.6% balanced test accuracy across eight diagnostic classes). The results suggest that transfer learning for deep, densely connected or residual CNN architectures with the judicious use of class-imbalance handling is a viable and increasingly promising approach to automated skin cancer screening at dermatologist level for diagnosis, while also underscoring the need for solutions to severe class imbalance, imaging-artifact variability and limited cross-population generalization before such systems can be extensively clinically utilized.

 

Keywords — skin cancer detection; deep learning; convolutional neural network; dermoscopic (thermoscopic) images; ISIC Archive; HAM10000; transfer learning; melanoma classification.