Hippocampus segmentation is primarily expected to be precise and robust due to its important role in a timely and accurate diagnosis of brain-related disorders such as Alzheimer's disease. As the previous research using deep learning mostly relied on a large number of labeled data samples, here we investigate the effect of pre-training using the state-of-the-art self-supervised contrastive learning-based technique in order to leverage the performance without large amounts of labeled data. We thus develop a new framework for the task of hippocampus segmentation based on learning local-level discriminative features for better generalization of structural information of MRI brain images. The comparative results of downstream training with different labeled data fractions reveal that pre-training without labels provides a significant margin of improvement. Moreover, we also evaluate and validate the robustness and generalizability through downstream training using a different dataset.