Related Experiment Video
Updated: Jun 13, 2026
04:25
Manual Segmentation of the Human Choroid Plexus Using Brain MRI
Published on: December 15, 2023
Bridging Annotation Gaps: Hierarchical Self-Support Learning for Brain Tumor Segmentation
Saqib Qamar1,2, Mohd Fazil3, Zubair Ashraf4
1Department of Intelligent Systems, KTH Royal Institute of Technology, 10044 Stockholm, Sweden.
Diagnostics (Basel, Switzerland)
|June 12, 2026
Summary
Hierarchical Adaptive Group Self-Support Learning (HAGSS) improves brain tumor segmentation from incomplete MRI data. This plug-in framework enhances accuracy by adaptively forming groups and focusing on tumor boundaries, without increasing computational cost.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computational Biology
Background:
- Accurate brain tumor segmentation using Magnetic Resonance Imaging (MRI) relies on fusing multiple data modalities.
- Clinical MRI often presents incomplete datasets due to acquisition issues, patient limitations, or varying protocols.
- Existing segmentation methods struggle with incomplete data, either isolating modality processing or using costly teacher networks.
Purpose of the Study:
- Introduce Hierarchical Adaptive Group Self-Support Learning (HAGSS) to address limitations in existing group self-support methods for incomplete multi-modal brain tumor segmentation.
- Overcome static group formation, uniform voxel treatment, and distribution mismatch issues present in current approaches.
- Enhance the robustness and accuracy of brain tumor segmentation when dealing with missing MRI modalities.
Main Methods:
- Implement a hierarchical adaptive group formation mechanism using voxel-level prediction confidence scores for dynamic group reassignment.
- Introduce a boundary-aware calibration module for spatially varied distillation weights, prioritizing tumor boundary regions.
- Incorporate a cross-scale consistency regularization term to ensure agreement between multi-resolution predictions and stabilize self-support targets.
Main Results:
- HAGSS demonstrated consistent performance improvements over state-of-the-art methods on BraTS2020, BraTS2018, and BraTS2021 datasets.
- Achieved average Dice score gains of 1.30% (BraTS2020) and 1.61% (BraTS2021) across whole tumor, tumor core, and enhancing tumor regions.
- Observed statistically significant improvements (p<0.05) across all tested datasets and segmentation regions.
Conclusions:
- HAGSS functions solely during the training phase, introducing no additional parameters or inference costs.
- The framework is designed as a plug-in module, compatible with various multi-encoder incomplete multi-modal segmentation architectures.
- Publicly available code facilitates the adoption and further development of this approach for brain tumor segmentation research.
