Uncertainty-Aware Brain Tumor Segmentation Using Attention Residual U-Net with Guided Decoder and Monte Carlo Dropout

ICLR 2026 Conference Submission18378 Authors

Keywords: Brain Tumor Segmentation, Medical Image Analysis, Deep Learning, Attention Res-UNet with Guided Decoder (ARU-GD), Monte Carlo Dropout (MCD), Bayesian Uncertainty Estimation, Predictive Uncertainty Maps, MRI (Magnetic Resonance Imaging), Dice Similarity Coefficient (DSC), BraTS Dataset

TL;DR: This paper presents an uncertainty-aware deep learning model, ARU-GD+MCD, for brain tumor segmentation on MRI scans, combining high accuracy with Monte Carlo Dropout-based uncertainty maps to support reliable clinical decision-making.

Abstract: Precise brain tumor segmentation from MRI scans is essential for successful diagnosis, treatment planning, and follow-up. Through this research, we developed a new model, ARU-GD+MCD by incorporating Monte Carlo Dropout layers into the Attention Res-UNet with Guided Decoder (ARU-GD), a state-of-the-art architecture for brain tumor segmentation. These dropout layers enable the model to estimate uncertainty by capturing variability during prediction, thereby generating uncertainty maps that indicate regions of low confidence in the segmentation results. This provides doctors with informative insights into the reliability of the model’s outputs. Tested on the BraTS 2019 dataset with four MRI modalities (FLAIR, T1, T1CE, T2), the improved ARU-GD attained Dice scores of 0.886, 0.899, and 0.856, and IoU scores of 0.793, 0.818, and 0.748 for whole tumor, tumor core, and enhancing tumor regions, respectively. Our method compares favorably to baseline models such as UNet and Res-UNet not only in terms of segmentation accuracy but also through the addition of an important interpretability layer. These advances enable more confident and better-informed clinical decision-making, ultimately leading to improved patient outcomes.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18378