Sampling Methods to Balance Classes in Dermoscopic Skin Lesion Images

Abstract

Convolutional neural networks are used to classify dermoscopic skin lesion images. The high accuracy of deep learning models is well documented; however, those models do not perform very well on testing (unseen data) sets due to imbalanced classes of images. To tackle this problem, over-sampling and under-sampling methods are explored in this study. Part 1 of the study focuses on the details of these sampling techniques, while Part 2 highlights the architecture of the deep learning model and its performance when using both sampling approaches. The results of Part 1 show that through the use of unsupervised learning techniques, namely, Hierarchical Clustering, Self-Organizing Maps, and K-Means, similar images are clustered, based on the skin lesions’ shape and color. Using augmentation for oversampling, 32,731 images are included for the training task in total. For undersampling, unsupervised learning techniques suggested 3 or 4 sub-groups of melanocytic nevi. Going through those clusters, the image background color also affects the way unsupervised learning techniques group similar images together.