Effect of image noise on the classification of skin lesions using deep convolutional neural networks

Skin lesions are in a category of disease that is both common in humans and a major cause of death. The classification accuracy of skin lesions is a crucial determinant of the success rate of curing lethal diseases. Deep Convolutional Neural Networks (CNNs) are now the most prevalent computer algorithms for the purpose of disease classification. As with all algorithms, CNNs are sensitive to noise from imaging devices, which often contaminates the quality of the images that are fed into them. In this paper, a deep CNN (Inception-v3) is used to study the effect of image noise on the classification of skin lesions. Gaussian noise, impulse noise, and noise made up of a compound of the two are added to an image dataset, namely the Dermofit Image Library from the University of Edinburgh. Evaluations, based on t-distributed Stochastic Neighbor Embedding (t-SNE) visualization, Receiver Operating Characteristic (ROC) analysis, and saliency maps, demonstrate the reliability of the Inception-v3 deep CNN in classifying noisy skin lesion images.