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

Abstract

In real-world clinical and biomedical settings, data distributions continuously evolve as new cases emerge in limited numbers. Traditional neural networks typically suffer severe knowledge forgetting when adapting to new classes unless retrained from scratch, which demands substantial resources. Moreover, biomedical datasets commonly exhibit imbalanced or long-tailed distributions where newly emerging classes have fewer samples, leading to classification bias toward established classes. However, the biomedical field currently lacks effective class-incremental learning methods to address this challenge of recognizing new diseases while preserving knowledge of previously learned ones. To address this gap, we introduce the first class-incremental learning method specifically designed for limited biomedical samples, featuring three key innovations: 1) a fine-grained semantic expansion module that employs diverse augmentation techniques to create compact feature distributions while accommodating new class generalization; 2) a cumulative entropy-based selection module that identifies and stores highly informative samples as exemplars for model review; and 3) a dynamic cosine classifier that mitigates classification bias from imbalanced datasets. Experiments across three datasets with different resolutions demonstrate superior performance under both imbalanced and long-tailed distributions, achieving up to 36.52% accuracy improvement over state-of-the-art methods.