Region-Wise Bézier Intensity Augmentation for Domain-Generalized Brain Tumor Segmentation with a Mamba U-Net.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Kırıkkale University, Kırıkkale 71450, Türkiye.
- Department of Software Engineering, Faculty of Engineering and Natural Sciences, Istanbul Topkapı University, Istanbul 34087, Türkiye.
- Department of Computer Engineering, Faculty of Engineering, Iğdır University, Iğdır 76000, Türkiye.
- Department of Electronics and Information Technologies, Faculty of Architecture and Engineering, Nakhchivan State University, Nakhchivan AZ 7012, Azerbaijan.
- Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Fenerbahce University, Istanbul 34758, Türkiye.
Abstract
<b>Background/Objectives:</b> Robust brain-tumor segmentation on contrast-enhanced MRI remains limited by scanner-dependent intensity shifts, scarce annotations, and evaluation protocols that may leak patient-specific information. We propose BA-SwinMamba, a region-wise Bézier intensity augmentation framework built on Swin-UMamba, a selective state-space U-Net that combines hierarchical Swin-style visual modeling with Mamba's linear-complexity long-range sequence representation. <b>Materials and Methods:</b> During training, independent monotonic or non-monotonic Bézier transfer functions are sampled for tumor and background regions, perturbing lesion-to-background contrast while preserving the binary mask geometry. Fourteen convolutional, transformer-based, and state-space segmentation models were evaluated on the Cheng brain-tumor dataset, comprising 3064 contrast-enhanced T1-weighted slices from 233 patients, using a strictly patient-level five-fold protocol. Single-source domain generalization was assessed by training only on Cheng and testing, without fine-tuning, on two independent target datasets. <b>Results:</b> BA-SwinMamba achieved 89.6% Dice, 82.0% IoU, and 5.9-pixel HD95 on the source domain, outperforming the plain Swin-UMamba backbone by 1.7 Dice points. The benefit was larger under domain shift: mean target-domain Dice increased from 72.7% with Swin-UMamba to 78.3% with BA-SwinMamba. Ablation analysis showed that replacing global Bézier augmentation with the proposed region-wise formulation added 1.5 Dice points. <b>Conclusions:</b> The method introduces no inference-time cost because augmentation is disabled after training, without modifying the deployed network or requiring target-domain labels during model optimization or tuning. The results indicate that lesion-aware intensity perturbation can improve cross-dataset robustness of Mamba-based 2D brain-tumor segmentation, while wider volumetric and multi-institutional validation remains necessary.