Melanoma Classification Based on Ensemble Classification of Dermoscopy Image Features

Abstract

Malignant melanoma is the deadliest form of skin cancer and is one of the most rapidly increasing cancers in the world. If diagnosed early, it can be easily cured, and consequently early diagnosis of melanoma is of vital importance. In this paper, we present an effective approach to melanoma classification from dermoscopy images of skin lesions. First, we perform automatic border detection to delineate the lesion from the background skin. We then extract shape features from the border, while colour and texture features are obtained based on a division of the image into clinically significant regions using a Euclidean distance transform. The derived features are then used in a pattern classification stage for which we employ a dedicated ensemble learning approach to address the class imbalance in the training data. In particular, we employ a committee of one-class classifiers for that purpose. One-class classification uses samples from a single distribution to derive a decision boundary, and employing this method on the minority class can significantly boost its recognition rate and hence the sensitivity of our approach. We combine several one-class classifiers using a random subspace approach and a diversity measure to select members of the committee. Experimental results on a large dataset of dermoscopic skin lesion images show our approach to work well, the employed classifier selection stage to be crucial for achieving this performance, and the classifier ensembles to perform statistically better compared to several state-of-the-art ensembles.