Information Gain Sampling for Active Learning in Medical Image Classification

Large, annotated datasets are not widely available in medical image analysis due to the prohibitive time, costs, and challenges associated with labelling large datasets. Unlabelled easier obtain, many contexts, it would be feasible for an expert provide labels a small subset of images. This work presents information-theoretic active learning framework that guides optimal selection images from unlabelled pool labeled based on maximizing expected information gain (EIG) evaluation dataset. Experiments performed two different classification datasets: multi-class diabetic retinopathy disease scale skin lesion classification. Results indicate by adapting EIG account class-imbalances, our proposed Adapted Expected Information Gain (AEIG) outperforms several popular baselines including diversity CoreSet uncertainty maximum entropy sampling. Specifically, AEIG achieves $${\sim }95\%$$ overall performance only 19% training data, while other approaches require around 25%. We show that, careful design choices, model can integrated into existing deep classifiers.