SPFNet : A Lung Nodule Segmentation Network Based on Superpixel Feature Fusion

ABSTRACT

The ability to accurately segment lung nodules in scans is fundamental to pathological assessment and clinical management. Yet, defining these boundaries remains difficult due to variations in nodule geometry and size, as well as the significant confounding between nodules and surrounding parenchyma. In order to mitigate these challenges, our work presents the SPFNet, a new segmentation model that leverages superpixel‐based feature fusion for precise lung nodule delineation. Firstly, the Boundary‐Enhanced Superpixel Fusion (BESF) module enhances boundary perception by fusing original CT features with superpixel features, effectively resolving issues of fuzzy borders and tissue ambiguity. Secondly, a Multi‐scale Non‐local (MSNL) component is developed. Its purpose is to captures long‐range dependencies across multiple scales, thereby enriching contextual understanding and improving the delineation of small or irregular nodules. Finally, the Bidirectional Attention Fusion (BAF) method is put forward to forge active links between low‐level spatial data and high‐level semantic insights. This mechanism aims to mitigate feature inconsistency and increase overall segmentation fidelity. The LIDC‐IDRI benchmark verified SPFNet's performance, registering a Dice of 88.67% and an IoU of 80.67%. Additionally, visualization attested to SPFNet's proficiency in feature boundary capture, thereby establishing its robust capability for the detection and initial assessment of pulmonary nodules.