SM-FSL: Similarity Guided Multi-Source Few-Shot Learning for Lung Ultrasound Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Lung ultrasound (LUS) is vital in emergency medicine and critical care.
• Limited annotated data and expert availability hinder LUS deployment.
• Few-shot learning is crucial for LUS analysis in data-scarce environments.

Purpose of the Study:

• To propose a Similarity-Guided Multi-Source Few-Shot Learning (SM-FSL) paradigm for LUS analysis.
• To develop the Global-Local Multi-Source Domain Network (GLMD-Net) for robust LUS analysis.
• To address challenges of data scarcity and improve diagnostic performance in LUS.

Main Methods:

• Developed the Average Nearest Neighbor Set Distance (ANNSD) for source domain prioritization.
• Implemented a similarity-guided optimization for multi-source gradient harmonization.
• Incorporated a local feature enhancement module and self-supervised learning for robustness.
• Utilized a few-shot adaptation strategy for efficient knowledge transfer.

Main Results:

• The GLMD-Net demonstrated superior performance compared to state-of-the-art methods.
• The approach effectively enhances generalization and robustness in few-shot learning scenarios.
• Validated effectiveness on public COVID-19 ultrasound datasets.

Conclusions:

• The proposed SM-FSL paradigm and GLMD-Net effectively address data scarcity in LUS analysis.
• The method offers a robust solution for AI-driven LUS diagnostics in resource-limited settings.
• This work advances the application of few-shot learning in medical imaging.