Accurate segmentation of retinal layers in optical coherence tomography (OCT) is crucial in the diagnosis of various retinal diseases, including the early detection of age-related macular degeneration (AMD). The ellipsoid zone (EZ), a hyperreflective layer associated with photoreceptor integrity, plays a critical role in disease assessment but poses challenges for automated analysis, as degeneration in pathologic eyes leads to variable visibility and structural alterations. This study introduces a fully automated method for segmenting the anterior EZ boundary using a Transformer-based architecture derived from UNETR. The proposed network produces two outputs (1) a sub-pixel depth position, indicating the location of the anterior EZ boundary and (2) a presence prediction, representing the probability that the EZ is present at that position. The model is trained only on the prediction of the depth position, while the presence prediction is derived from the 3D output feature map. Trained on 68 high-resolution volumetric OCT scans with an axial resolution of 2.7 μm, the proposed method achieved a mean Median Absolute Error of 1.11 μm and a 12% failure rate at a 4 μm threshold for subpixel EZ boundary localization. For EZ presence prediction, it reached 0.96 accuracy and an ROC AUC of 0.86. Qualitative evaluation confirmed anatomically consistent predictions across healthy and diseased retinas. Furthermore, a derived EZ-based biomarker showed a gradual increase with disease severity in AMD, highlighting its potential for both early detection and disease staging. These findings demonstrate the robustness and clinical relevance of the proposed framework.