Enhancing Monkeypox Diagnosis with Transformers: Bridging Explainability and Performance with Quantitative...

Area of Science:

• Artificial Intelligence in Dermatology
• Medical Image Analysis
• Machine Learning for Infectious Diseases

Background:

• Monkeypox diagnosis is challenging due to visual similarity with other skin conditions.
• Existing AI models for monkeypox classification primarily use Convolutional Neural Networks (CNNs), with limited use of Transformer architectures.
• Explainability in AI for dermatology often relies on unverified heatmap techniques.

Purpose of the Study:

• To apply Transformer-based models for monkeypox classification.
• To introduce and evaluate a novel hybrid explainability approach for AI models.
• To enhance the transparency and clinical relevance of AI in diagnosing dermatological conditions.

Main Methods:

• Fine-tuning Vision Transformer (ViT) and Data-Efficient Image Transformer (DeiT) for binary and multi-class monkeypox classification.
• Integrating Gradient-weighted Class Activation Mapping (Grad-CAM), Layer-wise Relevance Propagation (LRP), and Attention Rollout (AR) for model interpretability.
• Developing a hybrid explanation method combining heatmaps via Principal Component Analysis (PCA) and assessing reliability with deletion/insertion metrics.

Main Results:

• ViT models demonstrated superior performance, achieving an AUC of 0.9192 for binary and 0.9784 for multi-class classification.
• The hybrid explainability approach (Grad-CAM + LRP) provided more informative explanations than individual methods.
• Quantitative assessment confirmed enhanced clinical reliability of the hybrid explanations.

Conclusions:

• This research pioneers the use of Transformer models with systematically evaluated hybrid explainability for monkeypox classification.
• The study improves both predictive accuracy and interpretability of AI in dermatology.
• Future research should focus on expanding datasets and incorporating clinical metadata for broader applicability.