HI-MViT: A lightweight model for explainable skin disease classification based on modified MobileViT.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate and efficient skin disease classification is crucial for timely diagnosis and treatment.
• Existing models may lack the precision or deployability required for real-world clinical settings.
• Mobile deployment of AI diagnostic tools can significantly improve accessibility and speed of care.

Purpose of the Study:

• To develop a lightweight, explainable AI model for high-precision skin disease classification.
• To enable the deployment of this model on mobile terminals for practical use.
• To enhance the diagnostic capabilities of dermatologists through computer-assisted tools.

Main Methods:

• The HI-MViT model, based on Modified MobileViT, incorporates ordinary convolution, Improved-MV2, and MobileViT blocks.
• Improved-MV2 reduces computation while maintaining efficient information interaction and memory usage.
• Explainability is achieved through semantic feature dimensionality reduction and class activation mapping visualizations.

Main Results:

• HI-MViT achieved high performance on the ISIC-2018 dataset, with F1-Score, Accuracy, AP, and AUC scores of 0.931, 0.932, 0.961, and 0.977, respectively.
• The model outperformed top algorithms in ISIC-2018 and ISIC-2017 tasks, showing significant improvements in key metrics.
• Excellent performance on the PH² dataset demonstrated strong generalization capabilities.

Conclusions:

• The HI-MViT model offers excellent classification and generalization performance for skin diseases.
• Its explainable nature and lightweight design make it suitable for mobile deployment.
• This AI tool can assist dermatologists in classifying dermoscopic images more rapidly and reliably, supporting computer-assisted diagnostics.