guest ::
| Home > Publications Database > Cross-Domain Adaptation of a Whole-Body MRI Attention-Based 3D U-Net for Brain Tumor Segmentation |
| Home > Publications Database > Cross-Domain Adaptation of a Whole-Body MRI Attention-Based 3D U-Net for Brain Tumor Segmentation |
| Contribution to a conference proceedings/Contribution to a book | DZNE-2026-00760 |
; ; ; ; ;
2026
IEEE
This record in other databases:
Please use a persistent id in citations: doi:10.1109/DIPCA70202.2026.11566054
Abstract: The segmentation of brain tumors using magnetic resonance imaging (MRI) data is fundamental for precise diagnosis, effective treatment planning, and continuous monitoring of patient outcomes, and poses significant challenges as a result of variations in tumor morphology, intensity, variable size, and complex structure of tumor regions. Most existing deep learning approaches rely on task-specific architectures, requiring substantial effort to redesign and optimize models for each new imaging application. In this study, we investigate the possibility of adapting a 3D attention-based U-Net, originally developed for whole-body MRI analysis, for brain tumor segmentation. The network architecture was adapted, trained, and optimized using brain MRI datasets and evaluated on previously unseen cases. The proposed approach achieved Dice scores of 0.88,0.76, and 0.81 for whole tumor (WT), enhancing tumor (ET), and tumor core (TC), respectively, with corresponding sensitivity values of 0.95,0.84, and 0.71, and specificity exceeding 0.98 across all tumor regions. These results demonstrate that cross-domain reuse of a robust network architecture can reduce model development effort while maintaining high performance in complex medical imaging tasks. Our findings highlight the potential of attentionbased 3D networks as flexible tools for automated brain tumor analysis, offering a promising direction toward faster deployment of AI-assisted clinical applications.
|
The record appears in these collections: |
PDF
PDF (PDFA)