Efficient Selection of Rare Pathology Samples from Unlabeled Medical Data via Deep Active Learning
Ioannis S. Kafetzis, Alexander Hann, George F. Fragulis · 2025
Artificial Intelligence in medicine can improve diagnostic accuracy and efficiency, but progress is hindered by limited labeled data and class imbalances in medical datasets. Active learning offers a solution by selectively annotating the most informative samples. We propose a novel active learning approach that dynamically adjusts sample selection based on class prevalence and prediction confidence, boosting representation of rare classes. Evaluations on X-ray data show our method outperforms existing techniques, enhancing model performance with fewer annotations and supporting faster artificial intelligence deployment in clinical settings.