Multi-Class Skin Lesions Classification Using Deep Features.

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Skin cancer classification is challenging due to lesion variability.
• Current segmentation methods struggle with diverse lesion sizes and shapes.
• Automated systems are needed for efficient and accurate skin cancer diagnosis.

Purpose of the Study:

• To develop an improved automated system for skin cancer classification.
• To enhance classification accuracy and efficiency by overcoming limitations of segmentation.
• To introduce a novel approach using hybrid and optimal feature selection.

Main Methods:

• Dataset balancing using brightness, sharpening, and contrast enhancement transformations.
• Transfer learning applied to retrain Darknet53 and Inception V3 Convolutional Neural Networks (CNNs).
• Deep feature extraction from retrained CNNs, followed by optimal feature selection using Moth Flame Optimization (MFO) to address the curse of dimensionality.

Main Results:

• The proposed system achieved high classification accuracies: 95.9% (cubic SVM), 95.0% (quadratic SVM), and 95.8% (ensemble subspace discriminants).
• The hybrid and optimal feature selection approach demonstrated superior performance compared to existing state-of-the-art methods.
• The system effectively improved both the accuracy and efficiency of skin cancer classification.

Conclusions:

• The developed automated system offers a robust and accurate solution for skin cancer classification.
• The integration of hybrid feature selection and MFO effectively enhances diagnostic performance.
• This research provides a promising advancement in computer-aided diagnosis for dermatological conditions.