Comparison of Convolutional Neural Network Architectures on Dermastopic Imagery

In this paper, six convolutional neural networks of varying architecture and/or depth are compared along with methods of data pre-processing to explore their effects on classification performance of skin lesions. Skin cancer is one of the most common forms of cancer in the world. An early diagnosis along with rapid treatment is essential to a healthy recovery. However, the diagnosis of skin lesions can be a difficult task due to high inter-class similarities as well as large variances between cases. In recent years, the usage of Convolutional Neural Networks (CNN's), a type of image classification network, to diagnosis & predict diseases has become popular in the medical community. Their effectiveness, however, is dependent upon the quality and quantity of data available. Numerous networks with varying depths and designs exist; however, their performance on small datasets needs to explored further. The dataset used is the Human Against Machine with 10000 training images (HAM10000).