Self-Paced Contrastive Learning for Semi-supervised Medical Image Segmentation with Meta-labels

Pre-training a recognition model with contrastive learning on a large dataset\nof unlabeled data has shown great potential to boost the performance of a\ndownstream task, e.g., image classification. However, in domains such as\nmedical imaging, collecting unlabeled data can be challenging and expensive. In\nthis work, we propose to adapt contrastive learning to work with meta-label\nannotations, for improving the model's performance in medical image\nsegmentation even when no additional unlabeled data is available. Meta-labels\nsuch as the location of a 2D slice in a 3D MRI scan or the type of device used,\noften come for free during the acquisition process. We use the meta-labels for\npre-training the image encoder as well as to regularize a semi-supervised\ntraining, in which a reduced set of annotated data is used for training.\nFinally, to fully exploit the weak annotations, a self-paced learning approach\nis used to help the learning and discriminate useful labels from noise. Results\non three different medical image segmentation datasets show that our approach:\ni) highly boosts the performance of a model trained on a few scans, ii)\noutperforms previous contrastive and semi-supervised approaches, and iii)\nreaches close to the performance of a model trained on the full data.\n