Enhancing skin disease classification leveraging transformer-based deep learning architectures and explainable AI

Area of Science:

• Dermatology
• Artificial Intelligence
• Computer Vision

Background:

• Skin diseases impact a significant portion of the global population.
• Accurate and timely diagnosis is crucial but challenging.
• Automating skin disease classification is essential for clinical support.

Purpose of the Study:

• To evaluate the efficacy of transformer models, specifically DinoV2, for automated skin disease classification.
• To compare DinoV2 performance against existing benchmarks and CNN architectures.
• To enhance model interpretability using explainable AI techniques.

Main Methods:

• Utilized Vision Transformers, Swin Transformers, and introduced DinoV2 on a 31-class skin disease dataset.
• Validated model robustness on HAM10000 and Dermnet datasets.
• Employed GradCAM and SHAP for explainable AI analysis.

Main Results:

• DinoV2 achieved state-of-the-art performance with 96.48% accuracy and 97.27% F1-Score on the 31-class dataset.
• Demonstrated significant improvement (nearly 10%) over existing benchmarks.
• Consistently outperformed prior models on HAM10000 and Dermnet datasets.

Conclusions:

• Transformer models, particularly DinoV2, show exceptional potential for dermatological applications.
• Explainable AI techniques provide valuable clinical insights for diagnosis and treatment planning.
• This approach enhances transparency and supports clinical decision-making in dermatology.