In this study, we propose a 3D dense network with embedded capsule modules (3D-DUCaps) for automatically delineating the clinical target volume of esophageal cancer, addressing the spatial dependence issue between parts and the whole that cannot be effectively captured by 2D networks. The network integrates capsule modules into the encoding layers of the U-Net to enhance feature learning capabilities and preserve more information, enabling the inference of poses and learning the relationship between parts and the whole. Additionally, dense connections are introduced to further promote the fusion of high-level semantic information and low-level feature information, enhancing the network's information propagation capabilities. Compared with traditional 2D deep learning networks, the proposed 3D deep learning network demonstrates stronger spatial awareness and superior boundary delineation capabilities, resulting in better delineation of the clinical target volume of esophageal cancer. Experimental results indicate that the 3D-DUCaps network achieves a 2.4% improvement in the Dice Similarity Coefficient metric compared with the classical 3D-UNet network.