.png){kind=icon}
.png){kind=icon}
Return
NestStruct-Net: A structure-aware 3D segmentation network for nested tumor subregions
.png){kind=icon}
PRE
AI
B
T
S
Y
L
DOI：10.1016/j.eswa.2026.133868.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate segmentation of nested tumor subregions is challenging because inner lesions often exhibit small volumes, irregular morphologies, and ambiguous boundaries while remaining hierarchically coupled with larger surrounding regions. Existing 3D segmentation methods commonly rely on fixed receptive fields and discrete voxel-grid representations. As a result, incomplete representations of small inner lesions may be further degraded by repeated feature aggregation, which attenuates high-frequency boundary cues and introduces quantization artifacts during fine-grained reconstruction. To address this progressive chain of challenges, we propose NestStruct-Net, a structure-aware 3D segmentation framework organized as a coarse-to-fine process of adaptive perception, boundary preservation, and continuous-space refinement. First, Dynamic Receptive Field Attention adaptively adjusts sampling locations and receptive-field distributions to capture heterogeneous tumor morphologies. Second, Differentiable Dual-Edge Guided Feature Enhancement progressively integrates first-order Canny and second-order Laplacian operators into multi-scale features to reinforce ambiguous inter-subregion boundaries. Finally, Class-Aware Edge-Guided Implicit Feature Refinement combines structure-aware coordinate encoding, class-aware channel routing, and uncertainty-guided high-frequency correction to prioritize small-volume and unstable nested regions. Experiments on the BraTS benchmark demonstrate that NestStruct-Net achieves the highest average DSC of 84.01% and the lowest average HD95 of 3.06 mm among the compared methods, while requiring only 13.11M parameters and 46.99G FLOPs. Evaluations on Synapse and ACDC further demonstrate its applicability across different anatomical structures and imaging modalities: NS-Net achieves the highest average DSC on both datasets, together with the second-best average HD95 on Synapse.
Keywords:
3D medical image segmentation
Brain tumor segmentation
Nested subregions
Structure-aware,
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
Cited Papers