Enhancing Skin Disease Classification Leveraging Transformer-based Deep Learning Architectures and Explainable AI

Significant accuracy improvement for skin disease diagnosis: The study achieved 96.48% test accuracy using DinoV2, representing nearly a 10% improvement over previous benchmark results on the same dataset, suggesting AI could meaningfully assist dermatologists in identifying skin conditions among 31 different disease classes.

Explainability is critical for clinical adoption: The researchers used explainable AI tools (GradCAM and SHAP) to show doctors exactly which parts of skin images the AI was analyzing, addressing a key barrier to trusting AI in medical settings where clinicians need to understand and verify the reasoning behind diagnostic recommendations.

Model robustness validated across multiple datasets: The best-performing model (DinoV2) was tested on three different skin disease datasets (with 31, 23, and 7 disease classes), consistently outperforming benchmarks, which suggests the technology could potentially work reliably across different clinical settings and patient populations rather than just one specific dataset.

Transfer learning enables practical deployment: By using pre-trained weights from ImageNet rather than training from scratch, the researchers demonstrated an efficient approach that could be more feasible for hospitals and clinics with limited computational resources to implement AI-assisted skin disease diagnosis.

Addresses a significant global health need: With skin diseases affecting over one-third of the global population, automating classification could help reduce diagnostic delays and improve access to dermatological care, particularly in regions with limited specialist availability.