Abdominal multi-organ segmentation poses a significant challenge in medical image analysis; nevertheless, conventional supervised learning methods face substantial limitations due to the intricate distribution of organs and the scarcity of annotated data. In contrast, self-supervised learning (SSL) leverages the inherent feature-representation capabilities of unlabeled data and demonstrates remarkable potential in mitigating the generalization gap caused by insufficient annotations. Yet, existing SSL approaches still struggle to capture fine-grained, local features in medical images, thereby constraining their applicability in segmentation tasks. To address this limitation, this study proposes a novel network based on a momentum contrast learning framework: Momentum Gaze Contrastive Learning (MoGaze). This framework simulates the “focus and search” mechanisms of the human visual system, incorporating global-local feature interaction modeling and dynamic attention mechanisms to significantly enhance representations of critical anatomical regions. Unlike existing approaches, MoGaze decouples features from global and local views of the image and applies contrastive learning with mutual information maximization constraints, achieving fine-grained modeling of tissue morphology and boundary features. Experimental results demonstrate that MoGaze significantly improves segmentation performance across multiple medical image datasets, showing particular advantages in extracting fine anatomical details of small organs. In tasks such as abdominal multi-organ segmentation, MoGaze outperforms other state-of-the-art pretrained models, demonstrating superior adaptability and robustness.