Deep Ensemble Architectures for Skin Lesion Detection

Abstract

Skin cancer is the most serious form of cancer. The five percent survival rate of Melanoma demands the requirement of an automated Artificial Intelligence (AI) based skin diagnosis system to assist physicians with early diagnosis. The study focuses on discussing how the conventional therapeutic approaches have influenced the automation of diagnosis using digital images. Recent trends and advancements in AI-based skin lesion detection models are analyzed. Machine learning systems, transfer learned models, and ensembles of networks are comparatively studied. We conducted experiments on selecting the best combination of pre-trained ensemble models in the most challenging ISIC2018 dataset. Over 4% gain in model performance is achieved in terms of sensitivity, specificity, and balanced accuracy. Furthermore, a few challenges and potential prospects to improve future research and enhance the performance of diagnosis are discussed.