Uncertainty-Guided Dual-Views for Semi-Supervised Volumetric Medical Image Segmentation

Deep learning has resulted in tremendous progress in the field of medical AI. However,  training deep learning models usually require large amounts of annotated data. Annotating large-scale datasets, is prone to human biases and often very laborious, especially for dense prediction tasks such as image segmentation. Inspired by semi-supervised algorithms that employ both labeled and unlabeled data for training, we propose a  dual-view framework based on adversarial learning for segmenting volumetric images. In doing so, we employ critic networks to allow each view to learn from high-confidence predictions of the other view via measuring a notion of uncertainty. Furthermore, to jointly learn the dual-views and the critics, we formulate the learning problem as a min-max problem. We analyze and contrast our proposed method against state-of-the-art baselines, both qualitatively and quantitatively, on four public datasets with multiple modalities (\eg, CT and MRI) and demonstrate that the proposed semi-supervised method substantially outperforms the competing baselines while achieving competitive performance compared to fully-supervised counterparts. Our empirical results suggest that an uncertainty-guided co-training framework can make two neural networks robust to data artifacts and have the ability to generate plausible segmentation masks that can be helpful for semi-automated segmentation processes.