Uncertainty-Aware Brain Tumor Segmentation using UNet++ and Monte Carlo Dropout

Abstract

Brain tumors must be precisely segmented using multi-modal magnetic resonance imaging (MRI) for clinical diagnosis, treatment planing, and out-of-the-box evaluations. Although deep learning–based segmentation models have achieved high accuracy, most existing approaches generate deterministic prediction without conveying confidence information. This limits the clinical reliability. To address this limitation, an uncertainty-aware brain tumor segmentation framework based on the UNet++ architecture is presented in this work. This method employs a 2D UNet++ model with nested and dense skip connections to effectively fuse multi-scale spatial and semantic features from the multi-modal MRI sequence, including FLAIR, T1, T1ce and T2. Experiments are conducted using the BraTS 2020 dataset for multi-class segmentation of background, whole tumor, tumor core, and enhancing tumor regions. A hybrid loss function is adopted to alleviate the class imbalance and improve the delineation of small tumor subregions. To estimate predictive uncertainty, Monte Carlo Dropout is activated during inference, allowing multiple stochastic forward passes. The resulting voxel-wise uncertainty maps highlight regions of ambiguous predictions, particularly around tumor boundaries and complex structures. Visual results demonstrate close agreement between predicted masks and ground truth, while uncertainty maps provide valuable insights into model confidence. The proposed uncertainty-aware segmentation approach improves robustness and trustworthiness, making it well-suited for the clinical decision-support system.