Accurate segmentation of the optic nerve and surrounding structures in transorbital ultrasound (TOS) imaging is essential for non-invasive assessment of elevated intracranial pressure and other neuro-ophthalmic conditions. However, inherent low contrast, speckle noise, and ambiguous anatomical boundaries in ultrasound images pose significant challenges. In this work, we propose UncEGA-Net, a novel encoder-decoder architecture designed to robustly segment the optic nerve (ON) and optic nerve sheath (ONS) from TOS images. The backbone employs a ConvNeXtV2-Tiny encoder to extract high-level semantic features while maintaining computational efficiency. Most importantly, we introduce the EGA-U module (Edge-Guided Attention with Uncertainty), which exploits a Bernoulli-based proxy approximation to predict uncertainty, thus dynamically regulating the inverse attention and enhance the feature representation near the uncertainty boundary. Extensive experiments on a public multicenter dataset demonstrate that UncEGA-Net achieves state-of-the-art performance, outperforming strong baselines in terms of Dice coefficient and other boundary-based metrics. Furthermore, we validate the model’s clinical utility by estimating optic nerve diameters that show a moderate correlation with ground-truth measurements. Our approach highlights the value of integrating uncertainty modelling with edge-guided attention for precise and reliable segmentation in challenging ultrasound settings.