Dual-Stage Clean-Sample Selection for Incremental Noisy Label Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Image Analysis

Background:

• Class-incremental learning (CIL) in deep neural networks suffers from catastrophic forgetting (CF), degrading previously learned representations.
• Noisy labels in medical image datasets, due to costly annotations, severely compromise model performance.
• The combined impact of CF and noisy labels in CIL remains underexplored, particularly in medical imaging.

Purpose of the Study:

• To address the combined challenge of catastrophic forgetting and noisy labels in class-incremental learning for medical image analysis.
• To propose a novel method, Dual-Stage Clean-Sample Selection (DSCNL), integrating noise mitigation and memory rehearsal.

Main Methods:

• Developed a dual-stage clean-sample selection module to identify high-confidence samples and guide reliable representation learning.
• Implemented an experience soft-replay strategy for memory rehearsal to enhance robustness against historical noisy labels.
• Integrated these components into a unified framework to simultaneously mitigate noise and alleviate catastrophic forgetting.

Main Results:

• DSCNL consistently outperformed state-of-the-art CIL methods on public medical image datasets.
• The method boosted average accuracy by 55% and 31% on datasets with varying noise levels compared to baselines.
• Achieved an average noise reduction rate of 73% under original noise conditions.

Conclusions:

• The proposed DSCNL method effectively suppresses the adverse influence of noisy labels while alleviating catastrophic forgetting in CIL.
• Demonstrated the effectiveness and applicability of DSCNL in real-world medical imaging scenarios.
• Highlights the importance of addressing noisy labels and catastrophic forgetting concurrently for robust medical AI.