Dynamic Subcluster-Aware Network for Few-Shot Skin Disease Classification.

Area of Science:

• Medical Informatics
• Computer Vision
• Machine Learning

Background:

• Few-shot learning (FSL) presents challenges in classifying rare skin diseases due to limited data.
• Skin disease image datasets often contain inherent subclusters representing variations within a single disease class.

Purpose of the Study:

• To develop a novel approach, the subcluster-aware network (SCAN), to enhance few-shot skin disease classification accuracy.
• To improve the feature encoder's ability to capture subclustered representations within skin disease classes for better feature distribution characterization.

Main Methods:

• Proposed a dual-branch framework: one branch for classwise feature learning and another for preserving subclustered structures.
• Introduced a cluster loss for unsupervised clustering to learn image similarities.
• Designed a purity loss to refine clustering results, ensuring subclusters contain samples from the same class.

Main Results:

• The SCAN framework demonstrated superior performance compared to state-of-the-art methods.
• Achieved improvements of 2%-5% in sensitivity, specificity, accuracy, and F1-score on the SD-198 and Derm7pt datasets.

Conclusions:

• The proposed SCAN method effectively addresses the challenges of few-shot skin disease classification.
• SCAN's ability to learn subclustered representations significantly enhances diagnostic performance for rare skin conditions.