An nnU-Net-based framework with adaptive feature representation for 3D brain tumor segmentation
TL;DR
CondSEB-Net improves adaptive feature representation and boundary-related segmentation performance while preserving the self-configuring pipeline of nnU-Net v2, and provides benchmark-level evidence for the proposed framework.
Abstract
Background Accurate segmentation of glioma subregions from multimodal magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, and response assessment, but remains challenging because of boundary ambiguity, heterogeneous appearance, and small enhancing tumor (ET) components. This study aimed to develop and evaluate a controlled nnU-Net v2-based framework for three-dimensional (3D) brain tumor segmentation (BraTS) by improving adaptive feature representation and boundary-aware learning while preserving the reproducibility of the self-configuring nnU-Net pipeline. Methods We propose a conditional convolution and squeeze-and-excitation with boundary-aware learning network (CondSEB-Net), a controlled enhancement of the nnU-Net v2 framework for 3D medical image segmentation. Conditional convolution (CondConv) is introduced to improve sample-specific feature adaptation, multi-level 3D squeeze-and-excitation (SE) attention is used to recalibrate channel responses, and a boundary-aware loss based on signed distance maps (SDMs) is incorporated to strengthen contour-level supervision. The proposed method was evaluated on BraTS2020 and kidney tumor segmentation (KiTS)2019 using fixed five-fold cross-validation and compared with representative 3D segmentation baselines. Results On BraTS2020, CondSEB-Net improved the average Dice score from 0.8525±0.0217 to 0.8659±0.0184 and reduced the average 95% Hausdorff distance (HD95) from 10.57±4.01 to 9.08±3.64 mm compared with nnU-Net v2. The average symmetric surface distance (ASSD) also decreased from 3.08±0.37 to 2.65±0.30 mm. Fold-wise paired statistical analysis showed statistically supported improvements in both region overlap and boundary-related metrics. On KiTS2019, CondSEB-Net also achieved superior or competitive performance with limited additional computational overhead. Conclusions CondSEB-Net improves adaptive feature representation and boundary-related segmentation performance while preserving the self-configuring pipeline of nnU-Net v2. The current results provide benchmark-level evidence for the proposed framework, while external multi-center validation remains necessary before clinical deployment.