The delineation of Lymph Nodes (LNs) is pivotal in pinpointing therapeutic targets for radiotherapy in head and neck malignancies. Nevertheless, this endeavor poses a formidable challenge, primarily stemming from the suboptimal contrast against adjacent tissues. This investigation introduces a deep learning methodology aimed at automating the segmentation of LNs within CT scans, offering the following contributions: (1) Expanding upon the 3D Unet model, we incorporate a parallel block consisting of attention gate and squeeze & excitation modules. We extensively evaluate various versions of this parallel block and achieve favorable performance. (2) To address the slow decrease in Dice loss, we introduce a lightweight boundary refinement module. Our proposed method is assessed on a dataset comprising 103 patients and 603 Lymph Nodes (LNs), with 452 nodes used for training and 151 nodes for testing. The node-level Dice similarity coefficient achieved by our method reaches an impressive 0.833.