Node-U-Net: A Lightweight Attention-Guided Network for Pulmonary Nodule Segmentation

Lung cancer survival relies heavily on early nodule detection via computed tomography (CT), yet manual segmentation remains labor-intensive and unscalable for large clinical workflows. While existing deep learning models achieve high accuracy, they face a critical trade-off with computational efficiency and frequently lack cross-dataset generalization. To overcome these barriers, we introduce Node-U-Net, an ultra-lightweight attention-guided encoder-decoder architecture. This framework integrates a MobileNetV2 backbone with Residual Dilated Blocks and Light Atrous Spatial Pyramid Pooling (Light-ASPP) to expand multi-scale contextual awareness without inflating computational costs. Additionally, spatio-channel attention mechanisms and skip-connection attention gates adaptively suppress irrelevant background noise, while deep supervision ensures precise small nodule detection. Evaluated on the LIDC-IDRI dataset, Node-U-Net achieves a 94.34% Dice similarity coefficient, 94.20% Intersection over Union, and 98.31% precision ( $p<0.001$ , 92% win rate). Crucially, the model demonstrates robust zero-shot cross-domain generalization, maintaining a 91.49% Dice score on the unseen, independent UniToChest dataset. Notably, Node-U-Net achieves this with only 4.56 million parameters and 7.94 GFLOPs-representing a $34.8\times$ computational reduction compared to U-Net++. These results position Node-U-Net as a highly scalable solution for real-time AI-assisted lung cancer screening, making it viable for integration into resource-constrained clinical environments and edge devices.