Enhancing Monkeypox diagnosis and explanation through modified transfer learning, vision transformers, and federated learning

Abstract

The Monkeypox outbreak has emerged as a pressing global health challenge, evidenced by rising cases across nations. Individuals afflicted exhibit diverse dermatological symptoms that risk further transmission via contamination. Our study assessed the efficacy of three modified transfer learning models (M-VGG16, M-ResNet50, M-ResNet101) alongside vision transformers (ViT) across four investigations. We achieved high accuracy in discriminating Monkeypox cases, with M-VGG16 achieving 88%, 76%, and 77% accuracy in Studies One, Two and Four and M-ResNet50 achieving 89% in Study Three. To comprehend triggers for Monkeypox onset, we utilized Local Interpretable Model-Agnostic Explanations (LIME) to explain predictions visually. LIME alignments underscored our models' high accuracy, correlating with segmented identification of infected regions. Further, we implemented Federated Learning on decentralized data to evaluate generalization capabilities. Blending established deep learning with emerging decentralized learning and explanation techniques is vital in improving predictive accuracy and elucidating Monkeypox intricacies amid the persisting global outbreak. Our study emphasizes the continued relevance of pioneering techniques while introducing new approaches to address this major health challenge.

Keywords

• Transfer learning
• Vision transformer
• Disease diagnosis
• Image segmentation
• LIME