StructSAM: structure-aware prompt adaptation for robust lung cancer lesion segmentation in CT

Accurate delineation of lung lesions in computed tomography (CT) scans is critical for diagnosis, staging, and treatment planning, yet remains a challenging task. While foundation models like the Segment Anything Model (SAM) excel in natural images, they often falter in medical imaging due to low contrast, ambiguous boundaries, and a lack of 3D context. To address these limitations, we propose StructSAM, a structure-aware prompt adaptation framework designed for robust volumetric segmentation, with a primary focus on lung cancer. StructSAM injects anatomical priors into the prompt pathway, employs a 3D inter-slice aggregator for volumetric consistency, and leverages PEFT for scalability. Experiments on the LIDC-IDRI dataset demonstrate that StructSAM achieves state-of-the-art accuracy on lung nodule segmentation, outperforming both classical architectures and SAM-based adaptations. Crucially, extended cross-organ evaluations on KiTS19 and MSD Pancreas datasets reveal that StructSAM effectively generalizes to other anatomical structures, highlighting its robustness to domain shifts. These findings suggest that embedding structural priors into foundation models is a promising strategy toward generic, clinically reliable, and efficient medical image segmentation.