FOC-Net: A lightweight network combining full 1 × 1 convolutions with wavelet and attention mechanisms for lung nodule segmentation

Abstract

Accurate segmentation and analysis of lung nodules are essential for the formulation of effective treatment strategies. However, existing high-performance segmentation algorithms generally depend on substantial computational resources, presenting significant challenges for resource-constrained medical devices, especially edge devices. This study proposes a lightweight and efficient segmentation network called FOC-Net, which replaces large convolutional kernels by using 1 $\times$ 1 convolutions combined with spatial shifting operations. FOC-Net builds a U-shaped encoder–decoder architecture based on the shift convolution residual block (SCR-Block) and three other key modules: the wavelet-based downsampling (WBD) module, which preserves detailed information and suppresses noise, thereby reducing information loss during the downsampling process; the channel-prior spatial attention (CPSA) module, which makes the model focus on lung nodule regions; and the weight-aware feature fusion (WAFF) module, which augments the model’s ability to capture contextual information. Experiments conducted on the LIDC-IDRI dataset demonstrate that the proposed model outperforms other state-of-the-art methods in lung nodule segmentation tasks, achieving a Dice similarity coefficient (DSC) of 92.06% and a Jaccard index (JI) of 85.37%, while maintaining a parameter count of only 0.64 million with GFLOPs of 3.15. Further experiments on the ISIC-2018 skin disease dataset validate the model’s generalization capability, with similar results: a DSC of 89.36% and a JI of 80.77%, still outperforming other state-of-the-art methods.