DecX-Net: a dual-path feature decoupling network with layer-wise feature alignment for subsolid pulmonary nodule...

Objective.Subsolid pulmonary nodules are critical radiological indicators of early lung cancer in low-dose CT screening. Accurate component-level segmentation, particularly of solid and ground-glass components, is essential for malignancy risk assessment, treatment planning, and prognosis evaluation. However, precise segmentation of solid components in part-solid nodules remains challenging due to blurred boundaries, severe class imbalance, and insufficient multi-scale feature fusion in existing methods.Approach.We propose a dual-path feature decoupling network, termed DecX-Net, with layer-wise feature alignment for component-level segmentation of subsolid pulmonary nodules. Parallel dual-branch encoders are designed to separately capture local solid textures and global nodule structures, enabling effective feature decoupling. A cross-branch layer-wise alignment and mutual guidance module (CLAMG) enhances multi-scale semantic fusion, while an enhanced position attention module (E-PAM) further improves boundary localization and heterogeneous component segmentation. A multi-center clinical dataset with 463 annotated cases is constructed for evaluation.Main results.DecX-Net achieves superior performance over mainstream methods in Recall, Precision, MIoU, and DSC, with particularly strong capability in detecting small solid components.Significance.The proposed method enables accurate component-level analysis of subsolid pulmonary nodules and holds significant potential for improving clinical risk assessment and decision-making in early lung cancer.