Contrastive learning of global and local features for medical image segmentation with limited annotations
Published in NeurIPS 2020 (Oral presentation) (34th Conference on Neural Information Processing Systems), 2020
In this work, we propose contrasting strategies to leverage domain-specific cues in defining positive and negative pairs to leverage structural similarity across medical volumes. We also devise a local contrastive loss (problem-specific cue) lo learn distinctive local-level representations useful for segmentation tasks. The proposed pre-training lead to substantial improvements compared to baseline, pre-training with contrastive loss without domain cues, compared self-supervision and semi-supervised learning approaches.
Download here