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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate segmentation of glioma subregions in multi-parametric MRI (MP-MRI) is crucial for patient management.
• Tumor heterogeneity and ambiguous boundaries present significant challenges in current segmentation techniques.

Purpose of the Study:

• To develop an uncertainty-guided hybrid segmentation framework integrating 2D and 3D deep learning for enhanced glioma segmentation fidelity.
• To leverage prediction variance for quantifying voxel-level uncertainty and guiding localized 3D refinement.

Main Methods:

• A hybrid framework using a 2D nnU-Net with spherical projection deformation for initial predictions and uncertainty quantification.
• Localized 3D refinement of high-uncertainty regions using a dedicated 3D nnU-Net.
• Adaptive fusion of 2D and 3D predictions optimized via Particle Swarm Optimization on the BraTS 2020 dataset.

Main Results:

• The proposed hybrid method significantly outperformed standalone 2D and 3D nnU-Net baselines in segmenting enhancing tumor (ET), tumor core (TC), and whole tumor (WT).
• Achieved superior Dice Similarity Coefficients (DSC) for ET (0.8124), TC (0.7499), and WT (0.9055), with consistent improvements in HD95 and sensitivity.
• Demonstrated enhanced spatial coherence and boundary preservation, particularly in complex tumor regions.

Conclusions:

• The uncertainty-guided hybrid framework effectively combines 2D efficiency with 3D contextual accuracy for robust automated glioma segmentation.
• Interpretable uncertainty maps serve as a spatial attention mechanism, dynamically focusing computational resources on ambiguous areas.
• This approach offers a promising solution for improving the clinical management of gliomas through precise segmentation.