Self-paced and self-consistent co-training for semi-supervised image segmentation

Deep co-training has recently been proposed as an effective approach for image segmentation when annotated data is scarce. In this paper, we improve existing approaches semi-supervised with a self-paced and self-consistent method. To help distillate information from unlabeled images, first design learning strategy that lets jointly-trained neural networks focus on easier-to-segment regions first, then gradually consider harder ones. This achieved via end-to-end differentiable loss in the form of generalized Jensen Shannon Divergence (JSD). Moreover, to encourage predictions different be both consistent confident, enhance JSD uncertainty regularizer based entropy. The robustness individual models further improved using self-ensembling enforces their prediction across training iterations. We demonstrate potential our method three challenging problems modalities, small fraction labeled data. Results show clear advantages terms performance compared standard baselines state-of-the-art segmentation.