Breaking Diagnostic Barriers: Vision Transformers Redefine Monkeypox Detection.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Virology

Background:

• The global Monkeypox (Mpox) outbreak necessitates rapid and accurate diagnostic tools.
• Traditional methods like PCR are resource-intensive, driving the need for alternative approaches.
• Skin lesion image analysis presents a promising avenue for automated Mpox detection.

Purpose of the Study:

• To evaluate the effectiveness of vision transformers (ViTs) for automated Monkeypox detection using lesion images.
• To compare the performance of ViT-based transfer learning (TL) models against CNN-based TL and ViT models trained from scratch.

Main Methods:

• Fine-tuning a pre-trained ViT model on an Mpox lesion image dataset to create a TL model.
• Assessing model performance using accuracy, precision, recall, F1-score, and AUC.
• Utilizing a transferability measure and Grad-CAM for interpretability and feature analysis.

Main Results:

• ViT models significantly outperformed CNN models, achieving an AUC of 0.948 and accuracy of 0.942 (p<0.05).
• ViT models demonstrated superior transferability, effectively transferring learning weights to Mpox images.
• Grad-CAM visualizations confirmed ViT's focus on clinically relevant features, supporting diagnostic reliability.

Conclusions:

• ViT-based automated detection offers superior accuracy for Mpox diagnosis.
• This AI approach is a valuable tool for early Mpox detection in field settings with limited resources.
• The findings support faster outbreak response and improved public health resource allocation.