DSBNet : A Dual Sampling–Bottleneck Optimization Network for Efficient Pulmonary Nodule Segmentation

ABSTRACT

Accurate segmentation of pulmonary nodules is crucial for lung cancer screening and diagnosis. However, nodules often present irregular boundaries and heterogeneous morphology, making automatic segmentation challenging. U‐Net‐based methods remain limited by feature degradation during sampling, insufficient global representation at the bottleneck, and substantial computational overhead. To address these issues, we propose DSBNet, a dual‐stage optimization framework built upon Attention U‐Net. A contour‐guided attention wavelet module enhances feature fusion and boundary preservation during upsampling, an adaptive Fourier neural operator at the bottleneck enables efficient global context modeling, and a cross‐scale token encoder captures long‐range dependencies with limited overhead. Lightweight feature fusion operations further reduce parameter count and memory consumption. Experiments on LUNA16 and LIDC show that DSBNet improves IoU by 2.27% and 3.24% over the Attention U‐Net baseline, while reducing the parameter count by approximately 25%, FLOPs by approximately 40%, and memory consumption by approximately 15%. DSBNet achieves a favorable balance between segmentation accuracy and computational efficiency, demonstrating its potential for computer‐aided pulmonary nodule diagnosis. For further implementation details, please refer to the official DSBNet GitHub repository at https://github.com/mkx0mkx/DSBNet .