Few-shot learning for rare skin disease classification via adaptive distribution calibration

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Image Analysis

Background:

• Rare skin disease classification is hindered by data scarcity, impacting deep learning model training.
• Few-shot learning offers a solution by enabling novel disease identification with limited samples.
• Existing methods often fail to fully utilize base class information for few-shot class distribution calibration.

Purpose of the Study:

• To address data scarcity in rare skin disease classification.
• To propose a novel few-shot learning framework, SADC (skin disease classification via adaptive distribution calibration).
• To improve the accuracy of classifying rare skin diseases using limited data.

Main Methods:

• Developed a multi-scale feature extraction strategy using feature descriptor matrices and composite metrics for precise similarity computation.
• Implemented an adaptive sample calibration strategy to construct weight matrices for optimal base-class sample selection and distribution calibration.
• Utilized SADC framework incorporating both strategies for distribution-aware few-shot learning.

Main Results:

• SADC achieved state-of-the-art performance on three public dermatology datasets (ISIC2018, Derm7pt, and SD198).
• Demonstrated significant performance improvements over existing few-shot learning methods.
• Validated the effectiveness of the dual-strategy approach in advancing data-efficient medical image analysis.

Conclusions:

• SADC effectively overcomes data scarcity challenges in rare skin disease classification.
• The proposed framework enhances the accuracy and efficiency of medical image analysis.
• SADC represents a significant advancement in distribution-aware few-shot learning for dermatology.