LLNS-Net: Lightweight Lung Nodule Segmentation Network With Multiscale Information Fusion and Complementarity

To address the challenge that existing lung nodule segmentation algorithms face in balancing high accuracy with a lightweight design, we propose LLNS-Net, a compact yet effective lung nodule segmentation network. In the feature-mining encoder, convolutional residual blocks operating at multiple scales extract both shallow and deep nodule information from CT images, while an efficient multiscale attention mechanism enriches semantic representations. A subsequent feature enhancement module explores and leverages correlations among the outputs of the encoder's submodules. Within this module, we introduce an enhanced mixed local channel attention (E-MLCA) mechanism and a reinforced multiscale feature module to further strengthen cross-scale feature learning. The decoder aggregates features from four decoding layers and applies subchannel enhancement to refine the segmentation map. This design improves boundary smoothness and more accurately preserves the true morphology of nodule regions. Experimental results demonstrate that, compared with mainstream methods such as HmsUnet, MSA-Unet, and H-vmunet, the intersection over union of LLNS-Net improves by 1.86%, 0.27%, and 7.62%, respectively. Additionally, the generated feature maps exhibit smoother boundaries and superior visual quality compared to those produced by leading medical image segmentation algorithms.