EADU-Net:A Pulmonary Nodule CT Segmentation Network Based on Local-Global Fusion and Multi-Dilated Convolutions

Abstract

Accurate pulmonary nodule segmentation is essential for early lung cancer diagnosis. However, existing CNN-based methods still face three key limitations: weak local feature representation that produces coarse boundaries or internal holes, insufficient global context modeling that causes confusion with adjacent tissues and missed lesions, and limited receptive fields that hinder detection of diffuse or low-contrast nodules. To address these issues, we propose EADU-Net. The LOEAR module strengthens local feature extraction and boundary detail recovery, while its EAR submodule uses axial self-attention to capture long-range dependencies and improve semantic consistency. In the decoder, the HDC strategy expands the receptive field and fuses multi-scale context to reduce missed segmentations. On the LIDC-IDRI dataset, EADU-Net achieves 7.32 and 5.01 percentage-point gains in IoU and Dice, respectively, over U-Net under the ROI-based segmentation setting.