Robust Feature Spaces from Pre-Trained Deep Network Layers for Skin Lesion Classification

The incidence of skin cancer in the world population is a public health concern, and the first diagnosis takes into account the appearance of lesions on skin. In this context, automated methods to aid the screening for malign lesions can be an important tool. However, the efficiency of developed methods depends directly on the quality of the generated feature space which may vary when considering different image datasets and sources. We present a detailed study of feature spaces obtained from deep convolutional networks (CNNs), using the benchmark PH2 dataset, considering three CNN architectures, as well as investigating different layers, impact of dimensionality reduction, use of colour quantisation and noise addition. Our results show that, features have discriminative capability comparable to competing methods with balanced accuracy 94%, and 95% with noise injection. Additionally, we present a study of fine-tuning and generalisation across image quantisation and noise levels, contributing to the discussion of learning features from deep networks and offering a guideline for future works.