MFAF-Net: Lung Lobe Segmentation in CT Images Based on Multiscale Feature and Attention Fusion Network

This is the authors' abstract. We don't add key points for this paper.

Computed tomography (CT) images are widely used as clinical aids. Accurate lung lobe segmentation in CT images provides a strong basis for the diagnosis and treatment of lung diseases, but challenges arise from incomplete fissures, unpredictable pathological deformations, indistinguishable pulmonary arteries and veins, and severe injuries to the lung airways. To tackle these issues, we propose a multiscale feature and attention fusion network (MFAF-Net), which emphasizes lung fissure representations and suppresses the influence of other structures in lung lobe segmentation. First, the network follows the traditional encoder–decoder structure, incorporating FastKANConv-based residual modules to prevent gradient vanishing and further improve the spatial feature information extraction capability. Second, the adaptive spatial feature fusion module (ASFF), the atrous spatial pyramid pooling module (ASPP), and the feature fusion strategy are fused to address the target feature information loss issue effectively. Third, a stronger spatial and channel squeezing and excitation module (scSE) is improved to extract semantic information in both the spatial and channel domains in the skip connection part. The efficacy of the presented model is validated through experiments conducted on an existing dataset, which demonstrate superior intersection over union (IoU) values and lower Hausdorff distances (HD) than other efficient segmentation networks. The experimental results substantiate that the designed framework can effectively segment lung lobes in CT images.