Improved performance on melanoma skin cancer classification using deep learning based ensemble technique

Abstract

Skin cancer, particularly melanoma, arises from DNA damage that leads to abnormal cell growth in the epidermis. Early detection is crucial as melanoma can spread rapidly, but it is highly curable if identified promptly. Detecting and diagnosing melanoma early are essential to reduce mortality rates associated with this type of cancer. In the literature, various ensemble techniques have been proposed to improve the performance. This paper introduces a deep learning based ensemble method aimed at enhancing the accuracy of melanoma skin cancer detection. Additionally, it presents a thorough performance evaluation of five ensemble techniques. Initially, the dataset underwent pre-processing, involving the removal of artifacts through hair removal, and achieving a balance in the distribution of images for each class through image augmentation techniques. Then, the architecture of 16 pre-trained models was modified by adding additional layers to improve their performance. The models that achieved the highest melanoma accuracy were selected for ensembling. Since VGG16, MobileNetV2, and DenseNet169 achieved the highest melanoma accuracy, they were chosen for ensembling. Five ensemble techniques, namely, weighted average, voting, bagging, boosting, and stacking, were applied to the modified architectures of fine-tuned pre-trained models such as VGG16, MobileNetV2, and DenseNet169 to classify skin cancer images. The experiments were performed on a combined dataset of HAM10000 and ISIC2019, which contains images of seven skin lesion classes. The results demonstrate that the weighted average ensemble model achieves highest overall accuracy of 81.99% and melanoma classification accuracy of 89.85%. The positive outcomes affirm that employing ensemble techniques with adjusted model architectures enhances performance, thereby demonstrating their potential utility in the classification of skin cancer images.