Transformer-aided skin cancer classification using VGG19-based feature encoding.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Imaging and Diagnostics

Background:

• Skin cancer is a major global health concern, necessitating accurate and early diagnosis for improved patient outcomes.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), show promise for automated skin lesion classification but face limitations in dataset dependency, orientation sensitivity, and global context modeling.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model, VGG19-RSPDA-ViT, for enhanced skin lesion classification.
• To address limitations of existing CNNs by integrating fine-grained local feature extraction with global context modeling and rotation invariance.

Main Methods:

• A hybrid model combining VGG19 for local features and Vision Transformers (ViT) for global context was proposed.
• The Rotation and Shift-based Data Augmentation (RSPDA) technique was introduced to enforce rotation invariance and enrich feature representation, improving generalization on small datasets.
• The model's performance was validated on three benchmark datasets: MSK10000 (binary classification), HAM10000, and PH2 (multi-class classification).

Main Results:

• The VGG19-RSPDA-ViT model achieved high accuracies: 97.9% on MSK10000, 97.1% on HAM10000, and 98.67% on PH2.
• The model demonstrated consistently high macro-averaged precision, recall, specificity, and F1 scores across all tested datasets.
• Superior generalization capabilities were observed compared to existing state-of-the-art methods.

Conclusions:

• The proposed VGG19-RSPDA-ViT model is effective for classifying skin lesions, outperforming current methods.
• The integration of CNNs and Transformers with novel augmentation strategies shows significant potential for clinical application as an automated diagnostic tool in dermatology.