The very first key to the success of a supervised deep learning model is the availability of large labeled training datasets. But collecting appropriate large-scale annotated and labeled datasets for analyzing images with deep learning techniques is a persistent challenge. Contrastive self-supervised learning techniques are powerful models for solving this issue by using unlabeled data in a semi-supervised pre-train phase to obtain a good initialization for a downstream task in a supervised fine-tuning phase with restricted annotations data to provide high performance. Contrast learning strategy is a subset of self-supervised learning method which is suitable for learning representation at the image level . In this research, we study some contrastive self-supervised learning mechanisms and their role for segmentation of medical images.
