Transfer Learning with Ensembles of Deep Neural Networks for Skin Cancer Detection in Imbalanced Data Sets

Abstract Early diagnosis plays a key role in prevention and treatment of skin cancer. Several machine learning techniques for accurate detection cancer from medical images have been reported. Many these are based on pre-trained convolutional neural networks (CNNs), which enable training the models limited amounts data. However, classification accuracy still tends to be severely by scarcity representative malignant tumours. We propose novel ensemble-based network (CNN) architecture where multiple CNN models, some trained only data at hand, along with auxiliary form metadata associated input images, combined using meta-learner. The proposed approach improves model’s ability handle imbalanced demonstrate benefits technique dataset 33,126 dermoscopic 2056 patients. evaluate performance terms F1-measure, area under ROC curve (AUC-ROC), PR-curve (AUC-PR), compare it that seven different benchmark methods, including two recent CNN-based techniques. compares favourably all evaluation metrics.