Objective To address the clinical challenges of low tissue contrast, difficult segmentation, and limited performance of general segmentation models for esophagus on non-contrast CT (NCCT), this study aimed to develop a 3D nnU-Net model guided by dual-layer CT (DLCT) annotation to improve the accuracy and generalization ability of esophagus segmentation on NCCT, and to compare its performance with the Total Seg mentator general model. Methods A total of 300 patients who underwent chest DLCT examinations at the Fourth Affiliated Hospital of Nanjing Medical University were retrospectively enrolled. Manual gold-standard segmentation of the esophagus was performed by fusing virtual non-contrast (VNC) and enhanced images. A 3D nnU-Net model was trained using VNC images, with true non-contrast (TNC) images as the internal validation set. Additionally, 100 NCCT cases from Zhongda Hospital, Southeast University, and 48 NCCT cases from the China National Center for Bioinformation were included as two external test sets. The Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), consistency analysis, and 5-point Likert subjective scoring were used to compare the segmentation performance of the two models. Results In the internal validation set and the first external test set, the proposed model achieved a DSC of 0.756–0.798 and an HD95 of 4.00–4.10 mm, which were significantly superior to those of the Total Segmentator model (P<0.001). Its consistency with the gold standard (r?=0.821–0.887) and subjective scores (5.00–5.00) were also significantly higher than those of the control model. Conclusion The 3D nnU-Net model based on DLCT-guided annotation demonstrates superior segmentation performance and generalization ability in esophagus segmentation, providing a novel technical strategy for automated NCCT segmentation.