Melanoma classification from dermatoscopy images using knowledge distillation for highly imbalanced data.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melanoma is a dangerous skin cancer requiring early detection for better patient outcomes.
• Current large-scale screening is limited by human error and expert availability.
• Automatic melanoma classification from dermoscopy images faces challenges like class imbalance and similar-looking lesions.

Purpose of the Study:

• To develop a robust and accurate deep learning framework for melanoma classification.
• To address challenges of class imbalance, high inter-class, and low intra-class similarity in dermoscopy images.
• To improve sensitivity scores for early melanoma detection.

Main Methods:

• A knowledge-distilled lightweight Deep-CNN framework was proposed.
• Cost-Sensitive Learning with Focal Loss was used to handle class imbalance.
• Novel CutOut augmentation variants and an in-painting algorithm were employed for regularization and artifact removal.

Main Results:

• The EfficientNet-B2 student model achieved an Area Under the Curve (AUC) of 0.9295.
• A state-of-the-art sensitivity of 0.8087 was reached on the ISIC-2020 dataset.
• This represents a 49.48% increase in sensitivity compared to standalone models.

Conclusions:

• The proposed knowledge-distilled framework effectively addresses key challenges in melanoma classification.
• The method significantly enhances diagnostic sensitivity, aiding in early melanoma detection.
• This approach offers a promising solution for large-scale, accurate melanoma screening.