Hyperbolic Geometry-Driven Robustness Enhancement for Rare Skin Disease Diagnosis

Area of Science:

• Dermatology
• Computer Science
• Machine Learning

Background:

• Automated diagnosis of rare skin diseases using dermoscopy images is a challenging few-shot learning (FSL) problem.
• Traditional FSL methods often overlook the hierarchical nature of rare diseases and inherent data uncertainty.

Purpose of the Study:

• To develop a novel framework for rare skin disease diagnosis that leverages hyperbolic geometry to address limitations in FSL.
• To improve the representation of rare disease features and enhance uncertainty measurement in diagnostic models.

Main Methods:

• Proposed a Hyperbolic Geometry-driven Robustness Enhancement (HGRE) framework for rare skin disease diagnosis.
• Utilized hyperbolic space to implicitly capture class hierarchies and measure data uncertainty.
• Incorporated an Adversarial Proxy Construction (APC) module to handle data uncertainty by using distance to origin as an uncertainty indicator for adversarial robust training.

Main Results:

• The HGRE framework effectively addressed limitations in hierarchical relation utilization and data uncertainty in FSL.
• Empirical validation on two skin lesion datasets demonstrated HGRE's superior performance over state-of-the-art FSL methods.
• The model's robustness in training was significantly enhanced.

Conclusions:

• Hyperbolic geometry offers a promising approach for enhancing FSL in rare skin disease diagnosis.
• The HGRE framework provides a robust and effective solution for improving diagnostic accuracy and reliability.
• This work advances automated dermatological diagnostics by integrating advanced geometric deep learning techniques.