HC-U-Net: A Hierarchical Cascade of 3D U-Nets with Spatial-Channel Attention and Multi-Scale Context for Glioma Segmentation

Precise delineation of glioma subregions in multimodal MRI is vital for clinical decision-making, yet remains difficult because of tumor heterogeneity, pronounced class imbalance, and the very limited size of the Enhancing Tumor (ET). While 3D U-Net variants with attention and multi-scale modules have advanced the field, Most existing approaches handle the Whole Tumor (WT), Tumor Core (TC), and ET as separate categories, overlooking their natural anatomical hierarchy (WT → TC → ET). To address this, we propose HC-U-Net, a novel hierarchical cascade framework that explicitly models this anatomical dependency through a progressive three-stage segmentation strategy (WT → TC → ET). The proposed framework integrates three dedicated 3D U-Net experts, each designed for a distinct hierarchical stage: (1) A lightweight expert for robust WT localization; (2) A CBAM-3D-enhanced expert that uses the WT probability map to focus attention on refining the TC boundaries; (3) A multi-scale expert equipped with an ASPP-3D module, integrating WT and TC maps to perform the final four-class segmentation with enriched contextual reasoning. Each expert is trained in isolation to maintain stability and optimize memory usage. The key contribution of HC-U-Net is its targeted integration of attention and multi-scale mechanisms within a strict anatomical cascade, enabling progressive refinement from global to local structures. Comprehensive evaluations on the BraTS 2020 dataset show that our method establishes a new benchmark in balancing accuracy and efficiency, achieving mean Dice scores of 0.917, 0.857, and 0.812 for WT, TC, and ET, respectively. Importantly, it achieves a 10.5% relative gain in ET Dice over a conventional 3D U-Net baseline. The results validate that explicitly encoding hierarchical relationships, combined with stage-specific feature enhancement, is highly effective for precise glioma subregion segmentation.