A hierarchical three-step superpixels and deep learning framework for skin lesion classification

Area of Science:

• Dermatology and Computational Pathology
• Artificial Intelligence in Medical Imaging

Background:

• Skin cancer, particularly malignant melanoma, poses a significant global health risk with high mortality rates.
• Existing computerized methods for skin lesion diagnosis often lack sufficient accuracy.
• Accurate and early diagnosis of skin cancer is crucial for effective treatment and improved patient outcomes.

Purpose of the Study:

• To develop and evaluate a novel hierarchical framework for enhanced skin cancer classification.
• To improve the diagnostic accuracy of automated skin lesion analysis using deep learning and image processing techniques.

Main Methods:

• A hierarchical framework combining contrast enhancement of dermoscopy images, three-step superpixel lesion segmentation, and deep learning (ResNet-50) was proposed.
• Feature extraction using ResNet-50 was optimized with an improved grasshopper optimization algorithm.
• Classification was performed using a Naïve Bayes classifier on the optimized features.

Main Results:

• The proposed method achieved high classification accuracies on three diverse datasets (Ph2, ISBI2016, HAM1000): 95.40%, 91.1%, and 85.50%, respectively.
• The framework demonstrated effectiveness across datasets with varying numbers of skin cancer classes (3, 2, and 7 classes).

Conclusions:

• The developed hierarchical framework significantly improves the accuracy of skin cancer classification from dermoscopy images.
• This approach offers a promising tool for aiding dermatologists in the accurate diagnosis of skin lesions.