Cone Beam Computed Tomography (CBCT) is widely used in image-guided radiotherapy (IGRT) for treatment setup and adaptive planning. However, its clinical utility is limited by its image quality, characterized by increased noise, scatter artifacts, and poor soft tissue contrast comparing to CT, hindering accurate organ segmentation critical to adaptive workflows. Deep learning-based segmentation methods, although highly effective on high-quality CT and MRI, struggle to generalize on CBCT due to its degraded imaging conditions and the scarcity of reliable annotations. Manual contouring is both labor-intensive and inconsistent due to anatomical ambiguities in CBCT images, creating a significant bottleneck for developing robust segmentation models for CBCT. To overcome these challenges, we propose a self-supervised segmentation framework for CBCT using a Joint-Embedding Predictive Architecture (JEPA) built on Vision Transformers (ViTs). The method learns context-aware anatomical representations from unlabeled CBCT images by enforcing consistency between spatially distinct regions, eliminating the need for manual annotations. We validated the proposed framework on multi-fraction lung CBCT datasets, demonstrating that it significantly improves segmentation performance in CBCT while maintaining annotation efficiency and scalability. This study presents three major contributions: (1) a novel ViT-based self-supervised framework tailored for CBCT segmentation, (2) empirical validation of joint-embedding pretraining for enhancing downstream segmentation accuracy, and (3) an annotation-free, scalable strategy that bridges the gap between clinical CBCT data and accurate organ delineation. Our approach holds promise for enabling automated and consistent segmentation in clinical radiotherapy workflows.