DermViT: Diagnosis-Guided Vision Transformer for Robust and Efficient Skin Lesion Classification

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Early skin cancer diagnosis is crucial for patient survival.
• Current classification models struggle with lesion variability, artifacts, and lack physician-like reasoning.

Purpose of the Study:

• To develop a medically-driven deep learning architecture, DermViT, for improved skin lesion classification.
• To address challenges like semantic entanglement, intra-class variability, and artifactual interference in dermoscopic images.

Main Methods:

• DermViT utilizes a modular design inspired by physician diagnostic processes.
• Key modules include Dermoscopic Context Pyramid (DCP) for multi-scale analysis, Dermoscopic Hierarchical Attention (DHA) for focused lesion identification, and Dermoscopic Feature Gate (DFG) for artifact suppression.

Main Results:

• DermViT achieved 86.12% classification accuracy on ISIC2018 and ISIC2019 datasets, a 7.8% improvement over ViT-Base.
• The model demonstrated a 40% reduction in parameters compared to ViT-Base.
• Visualization confirmed DermViT's ability to accurately locate lesions even with interference.

Conclusions:

• DermViT offers a more logical feature extraction and decision-making process for medical diagnosis.
• The architecture provides an efficient and reliable solution for dermoscopic image analysis, enhancing early skin cancer detection.