An efficient multi-class classification of skin cancer using optimized vision transformer

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Skin cancer is a significant global health concern.
• Accurate and early detection of skin malignancies is crucial.
• Computer-based analysis of skin lesion images offers a promising approach for diagnosis.

Purpose of the Study:

• To develop and evaluate an optimized vision transformer model for accurate skin cancer classification.
• To improve the identification of skin tumors using advanced image analysis techniques.

Main Methods:

• Image pre-processing including color constancy, hair artifact removal, and noise reduction using filters (piecewise linear bottom hat, adaptive median, Gaussian) and gradient intensity.
• Image segmentation using the self-sparse watershed algorithm.
• Feature extraction via hybrid Walsh-Hadamard Karhunen-Loeve expansion.
• Skin cancer classification using an improved vision transformer.

Main Results:

• The proposed methodology achieved high performance metrics: 99.81% accuracy, 96.65% precision, 98.21% sensitivity, 97.42% F-measure, 99.88% specificity, 98.21% recall, 98.54% Jaccard coefficient, and 98.89% MCC.
• The system demonstrated superior performance compared to existing methodologies.
• Experiments were conducted on the International Skin Imaging Collaboration (ISIC) 2019 database.

Conclusions:

• The optimized vision transformer approach provides a highly effective and accurate method for skin cancer classification.
• The integrated methodology, combining advanced pre-processing, segmentation, and feature extraction, significantly enhances diagnostic capabilities.
• This research contributes a robust computational tool for aiding dermatologists in skin tumor identification.