Cross-level wavelet aggregation and dynamic dual-domain gated attention network for lung nodule segmentation in CT images

Precise segmentation of lung nodules is critical for early diagnosis and treatment of lung cancer. However, traditional methods are prone to boundary missegmentation and missed detection owing to the irregular shapes of nodules, blurred edges of small lesions, and vascular adhesion interference in CT images. This study proposes an improved U-Net model, SWAG-Net, which enhances the segmentation accuracy of edge-complex nodules and small-sized nodules through multi-module collaboration. The model designs a bidirectional path aggregation network that fuses cross-level feature pyramids to integrate the deep semantic information from the encoder and shallow details from the decoder, thereby enhancing the feature representation of small lesions. It embeds an adaptive multi-subband wavelet convolution module, utilizing wavelet transforms to explicitly separate multi-scale features and adaptively enhance high-frequency detail subbands, thereby addressing edge blurring and adhesion interference. Additionally, it introduces a dual-domain dynamic gating attention mechanism that combines channel-domain semantic filtering and spatial-domain edge enhancement to dynamically suppress background noise and highlight nodule regions. In the preprocessing stage, coordinate system decoupling and adaptive normalization were employed to reduce data bias. Experiments on the LUNA16 dataset demonstrated that the model achieved a DSC of 93.37% and an IOU of 88.79%, representing improvements of 11.43% and 19.39%, respectively, over the baseline U-Net. These results significantly improved the segmentation performance for irregular and small nodules. The experimental results demonstrate that this model effectively reduces mis-segmentations and missed detections, providing a reliable auxiliary basis for the early diagnosis of lung cancer.