Uncertainty-Aware Multi-Class Brain Tumor Segmentation Using Bayesian U-Net Variants.
Authors
Affiliations (3)
Affiliations (3)
- Department of Mathematics, UPES, UPES, Bidholi Campus, Via Prem Nagar, Dehradun, Dehradun, Uttarakhand, 248007, India.
- Data Science Cluster, UPES, School of Computer Science, UPES, Bidholi, Via Prem Nagar, Dehradun, Uttarakhand, 248002, India.
- Department of Mathematics, Indian Institute of Technology Jodhpur, IIT Jodhpur, Jodhpur, Jodhpur, Rajashthan, 342011, India.
Abstract
Accurate segmentation of brain tumors in magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, and surgical guidance. Although deep learning-based segmentation methods have achieved strong performance, their clinical adoption remains limited due to the lack of reliable uncertainty estimation. To address this challenge, we propose an uncertainty-aware framework for multi-class brain tumor sub-region segmentation and evaluate it across several U-Net variants. The study investigates how architectural design choices influence segmentation accuracy and prediction reliability. Specifically, we analyze Attention U-Net, Residual U-Net, Squeeze-Attention U-Net, and CBAM U-Net, each incorporating distinct attention mechanisms and feature reuse strategies. All architectures are integrated within a unified Bayesian inference framework using Monte Carlo (MC) dropout at test time to approximate posterior sampling. This approach enables pixel-wise and class-specific estimation of epistemic uncertainty along with segmentation outputs. Experiments conducted on the BraTS 2020 dataset show that segmentation accuracy across U-Net variants remains competitive for whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions. However, the generated uncertainty maps reveal distinct spatial patterns, particularly near ambiguous tumor boundaries and anatomically complex regions. These findings demonstrate that uncertainty-aware evaluation provides complementary insights beyond conventional accuracy metrics, improving model interpretability and reliability. Clinically, the integration of Bayesian uncertainty estimation with multiple U-Net backbones produces interpretable segmentation results accompanied by pixel-wise confidence information. Such uncertainty maps can assist radiologists in identifying low-confidence regions, supporting more informed decision-making, reducing inter-observer variability, and increasing trust in automated brain tumor segmentation systems.