Comparative analysis of active learning methods in biomedical image classification with MedMNIST benchmark

Deep learning has achieved tremendous results in computer vision tasks, setting new benchmarks in domains such as medical imaging. However, such performance often requires extensive human-labeled data, thereby increasing costs with task complexity. Active learning mitigates labeling expenses by selecting the most informative samples from a large pool of unlabeled data to construct a more compact training dataset. Based on the approach to data selection, the active learning methods are broadly categorized into two: uncertainty- and diversity-based methods. Uncertainty-based methods select samples with uncertain model predictions from the large pool of unlabeled data. In contrast, diversity-based methods select samples that reflect the overall distribution of the unlabeled dataset. While both strategies are effective, each has limitations which necessitate hybrid methods that combine uncertainty and diversity criteria. Active learning has been acknowledged for its potential with complex datasets such as medical images, and some studies have indeed tested its efficacy utilizing one or two types of medical data. However, few studies have evaluated its effectiveness across the full spectrum of medical imaging data. This study addresses that gap by offering a comprehensive review of the state-of-the-art active learning methods in medical imaging tasks on the MedMNIST v2 dataset.