Developing automated accurate and robust hippocampus segmentation is associated with the prevention of Alzheimer's disease. In this study, we devise a self-supervised learning framework for hippocampus segmentation while pre-training model without labels and transferring the pre-trained weights for downstream training with fewer labeled data. Results indicate competitive segmentation performance in fewer labeled training, especially in 10% and 20% label fractions, as well as robustness when trained for segmentation on another dataset.