Enhancing Early Monkeypox Diagnosis with an Interpretable ResNet-50 Model

Abstract

Monkeypox is an emerging viral disease that is re-emerging and with increased areas of transmission represents a growing public health threat. Despite its critical importance, diagnosis of the early stage of the outbreak is often difficult, and traditional methods including clinical examination or PCR testing have objectives, high cost and length of time to yield results. Although deep learning has become a promising tool for automated medical image analysis, the majority of contemporary models lack interpretability, generalization, and the ability to ensure the privacy of the data. An interpretable ResNet-50 based deep learning framework is proposed for early Monkeypox detection from dermatological images in this study. In contrast to conventional CNN models, the proposed one combines Grad-CAM, LIMEs, and SHAPs to bring explanation ensuring transparency in decision making. In addition to that, in combination with pre trained ImageNet weights, transfer learning improves the feature extraction and data augmentation and dropout regularization increases the model robustness. With the aim to address privacy concerns, the framework with federated learning is incorporated to train the models collaboratively for multiple institutions preserving patient data confidentiality at the same time. In the experimental tests, high classification accuracy of 99.04 % is achieved, with the F1-score being enhanced to above 95 %, and the AUC-ROC above 0.95 on various test sets. This study shows, that such a Monkeypox diagnosis AI model is both clinically viable, interpretable and privacy respectful.