From Uncertainty to Clarity: Uncertainty-Guided Class-Incremental Learning for Limited Biomedical Samples via...

Area of Science:

• Biomedical informatics
• Machine learning
• Artificial intelligence in healthcare

Background:

• Real-world biomedical data distributions evolve, with new cases emerging in limited numbers.
• Traditional neural networks face knowledge forgetting and require retraining from scratch for new data, which is resource-intensive.
• Biomedical datasets often have imbalanced or long-tailed distributions, causing bias towards established classes and hindering the recognition of new diseases.

Purpose of the Study:

• To develop the first class-incremental learning method specifically for limited biomedical samples.
• To address the challenge of recognizing new diseases while retaining knowledge of previously learned ones.
• To overcome limitations of existing methods in handling imbalanced and long-tailed biomedical data.

Main Methods:

• Introduced a fine-grained semantic expansion module using diverse augmentation techniques for compact feature distributions and new class generalization.
• Implemented a cumulative entropy-based selection module to identify and store informative samples as exemplars.
• Developed a dynamic cosine classifier to mitigate classification bias caused by imbalanced datasets.

Main Results:

• Demonstrated superior performance across three datasets with varying resolutions under imbalanced and long-tailed distributions.
• Achieved up to 36.52% accuracy improvement compared to state-of-the-art methods.
• Successfully addressed knowledge forgetting and classification bias in incremental learning scenarios.

Conclusions:

• The proposed class-incremental learning method is effective for evolving biomedical data with limited samples.
• The innovations in semantic expansion, exemplar selection, and classification mitigate bias and improve accuracy.
• This work provides a significant advancement for applying machine learning to dynamic biomedical challenges.