During the past few years, self-supervised contrastive learning has emerged as a paradigm for deep learning. In terms of representation learning, self-supervised learning can be thought of as an unsupervised learning approach in which the data act as the supervision. The key idea is to use available data in addition to or within the input dataset without annotations and use them to supervise intrinsic training datasets. This paper proposes a novel U-net-based self-supervised learning framework to obtain the nuclei images segmentation mask with limited labeled samples. Results of the experiment indicate, our self-supervised learning framework can significantly enhance the accuracy of nuclei segmentation even with lower annotated data required. Our developed model not only showed better performance in comparison to state-of-the-art segmentation models, but also it is more data-efficient and easier to be trained.