FEGGNN: Feature-Enhanced Gated Graph Neural Network for robust few-shot skin disease classification.

Area of Science:

• Dermatology
• Computer Vision
• Machine Learning

Background:

• Accurate skin disease classification is crucial but hindered by limited annotated images, especially for rare conditions.
• Few-shot learning (FSL) methods in computer-aided diagnosis (CAD) reduce data dependency but struggle with catastrophic forgetting in sequential tasks.

Purpose of the Study:

• To propose a Feature Enhanced Gated Graph Neural Network (FEGGNN) framework to enhance few-shot skin disease classification.
• To address the challenges of limited data and catastrophic forgetting in FSL for dermatology.

Main Methods:

• FEGGNN utilizes an Asymmetric Convolutional Network (ACNet) for feature extraction from skin lesion images.
• A Graph Neural Network (GNN) framework refines features, incorporating Gated Recurrent Units (GRUs) to manage task dependencies and mitigate forgetting.
• An Efficient Channel Attention (ECA) mechanism optimizes edge feature updates for improved graph representation.

Main Results:

• FEGGNN demonstrated state-of-the-art performance in few-shot skin disease classification.
• Achieved 84.90% accuracy on Derm7pt and 95.19% on SD-198 datasets in 2-way 5-shot settings.
• Showcased superior generalization to unseen classes, effectively overcoming data scarcity and catastrophic forgetting.

Conclusions:

• The FEGGNN framework offers a robust solution for few-shot skin disease classification, significantly improving diagnostic accuracy.
• The integration of ACNet, GNN, GRUs, and ECA effectively enhances feature representation and knowledge transfer.
• This approach holds promise for advancing CAD systems in dermatology, particularly for conditions with limited available data.