DD-Net: A Dual-Domain Progressive Fusion Network Enhanced by High–Low Frequency Separation for Lung Nodule Segmentation
Precise segmentation of pulmonary nodules is critical for distinguishing benign from malignant lesions and monitoring progression. However, substantial variability in nodule size, morphology, and texture often leads to e
Introduction Precise segmentation of pulmonary nodules plays a vital role in differentiating benign from malignant lesions and in tracking their progression over time. However, the substantial variability of nodules in size, morphology, and texture frequently results in either excessive or insufficient segmentation. This challenge necessitates advanced computational approaches capable of capturing both fine-grained boundary details and broader structural context.
Methods This paper proposes the spatial-frequency dual-domain collaborative segmentation network (DD-Net), which integrates the spatial and frequency domains to enhance the ability to characterize the diverse features of nodules. First, a frequency separation mechanism is introduced during the frequency-domain encoding process to model high- and low-frequency components differentially, balancing the representation of boundary details with overall structural information. Through frequency-domain filtering and learnable weight modulation, it obtains more discriminative frequency representations. Subsequently, an incremental attention fusion process is integrated into the high-level semantic feature interaction stage, enabling gradual alignment and reweighting of spatial and frequency domain features across channels and spatial dimensions. Finally, the traditional feature propagation path is restructured by a frequency self-attention mechanism, which models correlations between different components in the frequency space to help amplify key responses.
Results The effectiveness of the proposed model was evaluated using the publicly available LIDC-IDRI dataset together with a meticulously annotated private dataset. Experimental results indicate that the proposed approach outperforms multiple state-of-the-art methods on both datasets, achieving IoU/Dice scores of 88.02%/92.51% on the public LIDC-IDRI dataset and 65.03%/77.08% on the highly challenging private dataset.
Conclusion The proposed DD-Net employs frequency separation and attention fusion mechanisms for accurate lung nodule segmentation, demonstrating great potential for early lung cancer screening.