Multi-window uncertainty-guided network for lung nodule CT segmentation

Accurate segmentation of lung nodules from computed tomography (CT) scans is crucial for lung cancer prevention and early diagnosis. However, this task is challenging due to the varying sizes, irregular shapes, inconsistent densities of lung nodules, and the visual similarity between nodules and surrounding tissues. To overcome these challenges, we propose a novel Multi-Window Uncertainty-Guided Segmentation Network (MUS-Net). This network learns diverse features of lung nodules from CT images under different window settings, enhancing segmentation accuracy. Specifically, we introduce an uncertainty indication module that provides Multi-Window Mask (MWM), effectively combining segmentation predictions from lung and mediastinal window settings. This innovative approach allows the network to identify regions of varying segmentation uncertainty, thereby enhancing the model’s sensitivity to complex nodule boundaries and density variations. Furthermore, we incorporate a feature filter module to specifically enhance the learning of features from images under different window settings. Our extensive experiments on the LIDC-IDRI dataset demonstrate that our MUS-Net outperforms existing methods, significantly improving segmentation accuracy and robustness against nodule variability. • We explore the uncertainty in lung nodule segmentation under multiple CT window settings. • We introduce Lung and mediastinal window inputs for better segmentation. • Our method uses a single annotation, reducing radiologists’ workload in labeling. • We propose MUS-Net to address segmentation uncertainty, significantly improving segmentation performance.