Show abstract - PDF - Reviews

Active learning (AL) selects informative samples for annotation. This is becoming increasingly crucial to medical image segmentation since image annotation is hardly scalable to full pixel-level labeling of large datasets. However, most research focuses on classification or natural image segmentation. Uncertainty-based AL methods tend to have sub-optimal batch-query strategies, and diversity-based methods are computationally expensive. This work improves uncertainty-based AL for medical image segmentation using stochastic batches during sampling, computing uncertainty at the batch-level. Experiments on MRI prostate imaging show this approach’s effectiveness and robustness under various conditions.
Hide abstract