A Multi-model Deep Learning Architecture for Diagnosing Multi-class Skin Diseases.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin diseases are a major global health issue, impacting a large percentage of the population.
• Accurate and prompt diagnosis is essential for effective treatment and better patient outcomes.
• Existing diagnostic methods can be limited, necessitating advanced tools.

Purpose of the Study:

• To develop and evaluate a novel deep-learning multi-model architecture for high-precision skin disease diagnosis.
• To classify skin lesions into specific categories including Atopic Dermatitis, Acne and Rosacea, Skin Cancer, and Bullous conditions.
• To enhance diagnostic accuracy through transfer learning for specialized classification models.

Main Methods:

• A five-category Xception model was employed for initial classification of skin lesions.
• The model was trained on a dataset of 25,010 images.
• Transfer learning was utilized to create specialized models for improved accuracy across 40 distinct skin conditions.

Main Results:

• The initial multi-model achieved 95% accuracy and 99.4% AUROC.
• Specialized models demonstrated high performance: Skin Cancer (94.0% accuracy, 99.5% AUROC), Atopic Dermatitis (91.8% accuracy, 98.8% AUROC), Acne and Rosacea (90.0% accuracy, 99.0% AUROC), and Bullous (90.0% accuracy, 98.9% AUROC).
• The developed approach surpasses previous studies in diagnostic comprehensiveness.

Conclusions:

• The deep-learning multi-model architecture offers a significant advancement in skin disease diagnosis.
• The system provides high accuracy and reliability for identifying various dermatological conditions.
• The study's code is publicly available on GitHub to ensure reproducibility and further research.