Exploring dermoscopic structures for melanoma lesions' classification.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melanoma, a dangerous skin cancer, has high cure rates with early detection, but misclassification significantly reduces survival.
• Clinical variations in melanoma and benign nevi pose diagnostic challenges for dermatologists.
• Current diagnostic methods have limitations, highlighting the need for AI in dermatology.

Purpose of the Study:

• To explore dermoscopic structures for improved melanoma lesion classification using AI.
• To investigate and address the "brittleness" of AI models in distinguishing melanoma from benign lesions.
• To enhance AI model robustness against image variations in melanoma diagnosis.

Main Methods:

• Utilized datasets of expert-annotated dermoscopic images.
• Employed Transformer and Convolutional Neural Network (CNN) models for image classification.
• Focused on key dermoscopic structures: blue-white veil, dots/globules, and streaks.
• Assessed model performance and susceptibility to image variations using diverse test sets.

Main Results:

• Vision Transformers with added convolutions achieved up to 98% accuracy in melanoma classification.
• CNNs (VGG-16, DenseNet-121) showed 50-60% accuracy, performing better with non-dermoscopic features.
• Vision Transformers without convolutions demonstrated brittleness, with reduced accuracy on varied datasets.
• A data augmentation strategy successfully mitigated AI model brittleness and sustained accuracy.

Conclusions:

• A novel melanoma classification scheme using three dermoscopic structures was proposed.
• The study addressed AI model susceptibility to image variations in melanoma diagnosis.
• Future work includes collecting larger annotated datasets and automating dermoscopic feature extraction.