LightDPH: Lightweight Dual-Projection-Head Hierarchical Contrastive Learning for Skin Lesion Classification

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Effective skin cancer detection relies on accurate skin lesion classification.
• Existing models often lack efficiency for deployment on portable devices.
• Major-type misclassification remains a challenge in real-world applications.

Purpose of the Study:

• To propose a novel lightweight method for skin lesion classification that addresses major-type misclassification.
• To develop a model suitable for resource-constrained environments and portable applications.
• To enhance the sensitivity and efficiency of skin lesion detection models.

Main Methods:

• Introduced the Lightweight Dual Projection-Head Hierarchical contrastive learning (LightDPH) method.
• Utilized a dual projection-head mechanism with a multi-level contrastive loss (MultiCon Loss).
• Implemented a distance-based weight (DBW) function to adjust losses based on hierarchical levels.

Main Results:

• LightDPH achieved significant reductions in parameters (52.6%) and computational complexity (29.9% GFLOPs).
• The model maintained high classification performance comparable to state-of-the-art methods.
• A new metric, model efficiency score (MES), was proposed to evaluate cost-effectiveness.

Conclusions:

• The proposed LightDPH method effectively mitigates major-type misclassification in skin lesion detection.
• LightDPH offers a resource-efficient solution suitable for clinical applications in constrained environments.
• This study presents the first lightweight hierarchical classification model for skin lesion detection.