A voxel-wise uncertainty-guided framework for glioma segmentation using spherical projection-based U-Net and localized refinement.

BACKGROUND: Accurate segmentation of glioma subregions from multi-parametric MRI (MP-MRI) is critical for clinical management but remains challenging due to tumor heterogeneity and ambiguous tissue boundaries. PURPOSE: This study proposes an uncertainty-guided hybrid segmentation framework that integrates spherical projection-based 2D modeling with localized 3D refinement to improve segmentation fidelity. METHODS: The framework was validated on the BraTS 2020 dataset (N = 369). First, a 2D nnU-Net with spherical projection deformation was employed to generate initial slice-wise predictions. Crucially, prediction variance across multiple spherical projections was utilized to quantify voxel-level uncertainty, highlighting regions of low model confidence. A kernel-based sliding window algorithm then spatially localized 3D subvolumes with high cumulative uncertainty. These targeted regions were subsequently fed into a dedicated 3D nnU-Net for volumetric refinement. Finally, the global 2D predictions and local 3D refinements were adaptively fused using weights optimized via Particle Swarm Optimization. The proposed method was implemented to segment the enhancing tumor (ET), tumor core (TC) and whole tumor (WT). Performance was evaluated against standalone 2D and 3D nnU-Net baselines using the Dice Similarity Coefficient (DSC), HD95, sensitivity, and specificity. RESULTS: The proposed method significantly outperformed 2D and 3D baselines across ET, TC, and WT targets. Notably, it achieved a DSC of 0.8124 for ET (vs. 0.7527 for 2D and 0.7530 for 3D), 0.7499 for TC (vs. 0.7002 for 2D and 0.7027 for 3D), 0.9055 for WT (vs. 0.8989 for 2D and 0.9038 for 3D) and demonstrated consistent gains in HD95 and sensitivity. Quantitative metrics and visualizations confirmed improved spatial coherence and boundary preservation in structurally complex regions. CONCLUSION: By utilizing interpretable uncertainty maps as a spatial attention mechanism, this approach dynamically allocates computational resources to anatomically ambiguous regions. The resulting hybrid framework successfully combines 2D efficiency with 3D contextual accuracy, offering a robust solution for automated glioma segmentation.