Melanoma classification using ensemble deep transfer learning

Abstract

Melanoma, a type of skin cancer, poses significant challenges in early detection and diagnosis. Several methods for early melanoma detection, including visual inspection and several machine learning models, face challenges with accuracy. To overcome these issues, deep learning has been widely adopted in various biomedical applications. In this work, we employ deep transfer learning methods to classify melanoma. Firstly, we collect publicly available datasets containing melanoma images, their corresponding ground truth for segmentation, and class labels. Subsequently, we perform data preprocessing, normalization, and label encoding to address issues of varied illumination, image noise, and data imbalance. Next, we conduct feature extraction utilizing the previously trained deep learning models, VGG, ResNet, InceptionResNet, and MobileNet. The characteristic vectors obtained from each model are fused to produce a comprehensive depiction among the provided pictures. In the classification stage, we employ ensemble learning using transfer learning models, including EfficientNet, Xception, and DenseNet. These models are trained on the final feature vector to classify melanoma images effectively. The effectiveness of the suggested method is verified using publicly available ISIC 2017–2020 datasets, these model reports average accuracy scores of 96.10%, 97.23%, 97.50%, 98.33%, and 98.60%, in that order.