Related Experiment Video
Updated: Jul 3, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Uncertainty-aware feature mapping and adaptive inference for brain tumor segmentation with missing contrast-enhanced
Weidong Liu1, Zhuyin Zhang1, Fangfang Deng1
1Guangzhou University of Chinese Medicine, Guangzhou, China.
Frontiers in Neuroscience
|July 2, 2026
Summary
This study introduces UAF-AIMM, a novel framework for brain tumor segmentation when contrast-enhanced T1-weighted imaging (T1ce) is missing. The method achieves high accuracy, improving segmentation of contrast-dependent enhancing tumor subregions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Multi-parametric magnetic resonance imaging (mpMRI) is crucial for brain tumor assessment.
- The contrast-enhanced T1-weighted (T1ce) sequence is often unavailable, hindering segmentation accuracy.
- Existing methods struggle with missing modalities, necessitating robust frameworks.
Purpose of the Study:
- To develop a 3D brain tumor segmentation framework for scenarios with missing T1ce sequences.
- To improve the robustness of segmentation when key MRI modalities are absent.
Main Methods:
- UAF-AIMM (Uncertainty-Aware Feature Mapping and Adaptive Inference) framework proposed.
- Incorporates MR-Mapper for cross-modal feature recovery.
- Utilizes Uncertainty-Aware Attention (UAA) for reliable feature fusion.
- Employs Single-Image Test-Time Adaptation (SITA) for inference-time calibration.
Main Results:
- Achieved Dice Similarity Coefficient of 89.20% and HD95 of 5.50 mm on the BraTS 2021 dataset (missing-T1ce protocol).
- Outperformed leading baseline methods like 3D U-Net, Swin UNETR, mmFormer, and ReCoSeg.
- Demonstrated significant improvement in segmenting the contrast-dependent enhancing tumor (ET) subregion.
Conclusions:
- Feature-space compensation with reliability control and test-time calibration offers effective brain tumor segmentation.
- The approach is practical for incomplete multi-modal MRI protocols.
- Potential for translation to diverse clinical settings with heterogeneous MRI data.