Related Experiment Videos
Region-Wise Bézier Intensity Augmentation for Domain-Generalized Brain Tumor Segmentation with a Mamba U-Net
Mustafa Yurdakul1, Merve Ersoy2, Faruk Özger3
1Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Kırıkkale University, Kırıkkale 71450, Türkiye.
Journal of Clinical Medicine
|July 28, 2026
Summary
This study introduces BA-SwinMamba, a novel framework for brain tumor segmentation that improves robustness against MRI variations. The region-wise Bézier intensity augmentation significantly enhances cross-dataset performance without adding inference costs.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
- Neuro-oncology imaging
Background:
- Brain tumor segmentation on MRI is challenged by intensity variations and limited data.
- Existing methods struggle with scanner-dependent shifts and generalization across datasets.
- Robust segmentation is crucial for accurate diagnosis and treatment planning.
Purpose of the Study:
- To develop a novel augmentation framework, BA-SwinMamba, for robust brain tumor segmentation.
- To improve the generalization capability of segmentation models across different MRI datasets.
- To address limitations of intensity shifts and scarce annotations in medical image segmentation.
Main Methods:
- Proposed BA-SwinMamba, integrating region-wise Bézier intensity augmentation with Swin-UMamba, a selective state-space U-Net.
- Applied Bézier transfer functions to perturb lesion-to-background contrast during training.
- Evaluated 14 models on the Cheng dataset using a patient-level five-fold cross-validation and tested for single-source domain generalization.
Main Results:
- BA-SwinMamba achieved 89.6% Dice and 82.0% IoU on the source domain, outperforming Swin-UMamba by 1.7 Dice points.
- Mean target-domain Dice improved from 72.7% (Swin-UMamba) to 78.3% with BA-SwinMamba under domain shift.
- Region-wise augmentation contributed an additional 1.5 Dice points compared to global augmentation.
Conclusions:
- BA-SwinMamba enhances cross-dataset robustness for Mamba-based 2D brain tumor segmentation through lesion-aware intensity perturbation.
- The augmentation framework incurs no inference-time cost and does not require target-domain labels.
- Further validation with volumetric and multi-institutional data is warranted.