An Ensemble Classifier Based on Diverse Convolutional Neural Networks for Skin Lesions Classification

Skin cancer can be known as a malignant skin lesion and it is one of the deadliest diseases in all of the world. By using state-of-the-art deep learning approaches, it is possible to early diagnose skin lesions and reduce the mortality rate. Deep convolutional neural networks have demonstrated superior performance compared to the classical machine learning methods for automated skin lesion classification on dermoscopic images. In the presented work, we have focused on developing an accurate ensemble classifier constructed on convolutional neural networks to classify skin lesions. To develop a generalized ensemble classifier, we improved the level of diversity to make the classifier more robust and accurate. For this purpose, we have created diversity in both data and classifier levels. To train, validate, and test the model, the ISIC 2018 challenge dataset containing over 13,000 dermoscopic images was used. Based on the experiments, the proposed ensemble classifier achieved an accuracy of 90.15% which outperforms the base convolutional neural networks and other competing classifiers in terms of accuracy and robustness.