Related Experiment Video
Updated: Jun 17, 2026
10:48
PET and MRI Guided Irradiation of a Glioblastoma Rat Model Using a Micro-irradiator
Published on: December 28, 2017
ROI-aware uncertainty fusion for label-efficient glioma MRI segmentation
Saher Mohamed1, Ghada Khoriba1, Essam A Rashed2,3
1Center for Informatics Science, School of Information Technology and Computer Science (ITCS), Nile University, Giza, Egypt.
Physics in Medicine and Biology
|June 10, 2026
Summary
This study introduces an active learning framework for glioma MRI segmentation, significantly reducing annotation costs by intelligently selecting crucial data slices. This approach achieves high segmentation accuracy with substantially less labeled data, improving efficiency in clinical workflows.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate glioma segmentation on MRI is vital for treatment planning but is hindered by costly and time-consuming manual annotation.
- Existing methods often require extensive labeled data, limiting their practical application.
Purpose of the Study:
- To develop an efficient active learning framework for glioma MRI segmentation that minimizes annotation cost while maximizing segmentation quality.
- To integrate tumor-specific context, diverse uncertainty estimation, and robust fusion techniques for optimal data selection.
Main Methods:
- A tumor-focused active learning framework utilizing a 40-slice window for context and computational efficiency.
- A diverse uncertainty committee including softmax entropy, margin, least confidence, variation ratio, and Bayesian Active Learning by Disagreement (BALD), with BALD and variation ratio estimated via Monte Carlo (MC) dropout.
- Multi-criteria rank-level fusion using Borda count to aggregate slice-level and region-of-interest (ROI)-level uncertainties for robust data selection.
Main Results:
- The framework achieved near-full performance on UCSF-PDGM with only ~16% of the training data.
- On BraTS 2019, ~32% of the data yielded performance comparable to, and in some cases surpassing, models trained on the entire dataset.
- Demonstrated strong label efficiency and robustness across different datasets under realistic annotation budgets.
Conclusions:
- The proposed framework effectively combines tumor-focused processing, MC-dropout-enhanced uncertainty estimation, ROI-aware analysis, and Borda fusion.
- This approach significantly reduces annotation requirements for accurate glioma MRI segmentation, aligning with practical clinical data curation workflows.
