C MAL: cascaded network-guided class-balanced multi-prototype auxiliary learning for source-free domain adaptive...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Source-free domain adaptation (SFDA) is vital for medical image analysis, adapting models without labeled target data.
• Self-training methods in SFDA struggle with model instability from incorrect pseudo-labels and class imbalance.

Purpose of the Study:

• To enhance model stability in SFDA for medical imaging.
• To address challenges of pseudo-label accumulation and class imbalance in self-training SFDA.

Main Methods:

• Introduced a cascaded network-guided class-balanced multi-prototype auxiliary learning (C2MAL) framework.
• Employed a cascaded translation-segmentation network (CTS-Net) for accurate pseudo-label generation.
• Utilized a class-balanced multi-prototype auxiliary learning network (CMAL-Net) with novel loss functions for adaptation.

Main Results:

• C2MAL demonstrated superior performance over state-of-the-art methods on four benchmark fundus image datasets.
• The framework effectively improved model stability and adaptation, especially under significant domain shifts.
• Validated through extensive experiments on diverse medical imaging datasets.

Conclusions:

• The proposed C2MAL framework significantly enhances model stability and adaptation in SFDA for medical imaging.
• The novel approach effectively tackles pseudo-labeling issues and class imbalance, outperforming existing methods.
• C2MAL offers a robust solution for adapting medical image analysis models to new domains without labeled data.