DPEA-Net: a clinically-oriented lightweight 3D CNN for glioma segmentation in multiparametric MRI

Accurate segmentation of glioma subregions, whole tumor (WT), tumor core (TC), and enhancing tumor (ET), from multiparametric MRI is essential for radiotherapy planning and longitudinal assessment.

We propose DPEA-Net, a lightweight 3D architecture that addresses these challenges through two novel components. First, the Dynamic Hierarchically Decoupled Convolution (DHDC) unit reduces parameters by 99% compared to 3D U-Net while enabling adaptive multi-scale feature extraction to handle tumor heterogeneity. Second, the Cross-Dimensional Region-Specific Enhancement Attention (CDRSEA) module explicitly models 3D spatial relationships to refine ambiguous tumor boundaries.

On the BraTS 2019 and 2020 validation sets, DPEA-Net achieves mean Dice scores of 90.43%/89.96% (WT), 85.56%/86.52% (TC), and 81.89%/80.31% (ET) with a computational footprint of only 17.48 Giga Floating-point Operations (GFLOPs), enabling sub-2-second inference on standard clinical hardware. A 1.5-fold TC weighting strategy further enhances segmentation of the clinically critical tumor core.

DPEA-Net provides an accurate and computationally efficient tool for automated glioma subregion delineation, supporting practical integration into neuro-oncological workflows.