GOAL: Gist-set online active learning for efficient chest X-ray image annotation

Abstract

Deep learning in medical image analysis often requires extensive high-quality labeled data to achieve human-level accuracy. We propose Gist-set Online Active Learning (GOAL), a novel solution to the challenge of limited high-quality labeled data in medical imaging. Our approach advances existing active learning methods in three key ways.

Firstly, we enhance classification performance with fewer manual annotations through a sample selection strategy called gist set selection. Secondly, instead of focusing solely on random uncertain samples with low prediction confidence, we select only informative uncertain samples for human annotation. Lastly, we implement an online learning application where high-confidence samples are automatically selected, iteratively assigned, and pseudo-labels are updated.

We validated GOAL on two private datasets and one public dataset. Experimental results demonstrate that our approach can reduce the required labeled data by up to 88% while maintaining the same F1 scores as models trained on full datasets.