Multi-features extraction based on deep learning for skin lesion classification.

Area of Science:

• Dermatology
• Computer Science
• Machine Learning

Background:

• Feature extraction is vital for machine learning in skin lesion analysis.
• Both handcrafted and deep learning features are utilized.
• Convolutional Neural Networks (CNNs) show promise for automated skin lesion classification.

Purpose of the Study:

• To assess the efficacy of 17 pre-trained CNN architectures as feature extractors for skin lesion classification.
• To compare the performance of 24 machine learning classifiers on two distinct skin lesion datasets.
• To identify optimal CNN-classifier combinations for accurate skin lesion diagnosis.

Main Methods:

• Utilized 17 diverse pre-trained CNN architectures (e.g., DenseNet201) for feature extraction.
• Employed 24 machine learning classifiers, including Fine KNN and Cubic SVM.
• Evaluated performance on the ISIC 2019 and PH2 skin lesion datasets.

Main Results:

• DenseNet201 combined with Fine KNN achieved 92.34% accuracy on the ISIC 2019 dataset.
• DenseNet201 combined with Cubic SVM achieved 91.71% accuracy on the ISIC 2019 dataset.
• The proposed approach reached 99% accuracy on the PH2 dataset, outperforming existing methods.

Conclusions:

• Pre-trained CNNs are effective feature extractors for skin lesion classification.
• Specific CNN-classifier combinations (DenseNet201 with Fine KNN/Cubic SVM) yield high accuracy.
• The findings support the use of deep learning for improved dermatological diagnostics.