Scholarlink

Brain tumor segmentation remains a critical yet challenging task, particularly in low-resource settings where imaging data are often acquired with low-field MRI scanners. The BraTS-Africa 2025 Sub-Saharan Challenge provides a unique opportunity to assess the robustness of segmentation models on heterogeneous, real-world data from African clinical environments. In this study, we investigated the performance of three state-of-the-art architectures; nnU-Net, MedNeXt, and SegMamba for sub-compartmental glioma segmentation. While each model embodies a distinct design philosophy, ablation experiments revealed that nnU-Net consistently outperformed the others as well as their ensemble. Specifically, nnU-Net trained on the entire dataset achieved Dice scores of 0.870 (ET), 0.852 (TC), and 0.887 (WT) on the validation set, surpassing both the alternative models and a 5-fold cross-validated nnU-Net baseline. These results suggest that training on all available cases provides advantages over cross-validation in smaller datasets. This work underscores the effectiveness of nnU-Net in this setting and contributes to the broader goal of developing reliable AI tools for neuro-oncology in underrepresented global contexts.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reassessing Glioma Segmentation Strategies: nnU-Net as a Strong Baseline on Limited Sub-Saharan MRI Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Uwimana Lowami,
  • Andrew Blayama Stephen,
  • Confidence Raymond,
  • Dong Zhang,
  • Maruf Adewole,
  • Udunna C. Anazodo,
  • Mehmet Kurt,
  • Damilare Olatunji,
  • Bernes Lorier Atabonfack

摘要

Brain tumor segmentation remains a critical yet challenging task, particularly in low-resource settings where imaging data are often acquired with low-field MRI scanners. The BraTS-Africa 2025 Sub-Saharan Challenge provides a unique opportunity to assess the robustness of segmentation models on heterogeneous, real-world data from African clinical environments. In this study, we investigated the performance of three state-of-the-art architectures; nnU-Net, MedNeXt, and SegMamba for sub-compartmental glioma segmentation. While each model embodies a distinct design philosophy, ablation experiments revealed that nnU-Net consistently outperformed the others as well as their ensemble. Specifically, nnU-Net trained on the entire dataset achieved Dice scores of 0.870 (ET), 0.852 (TC), and 0.887 (WT) on the validation set, surpassing both the alternative models and a 5-fold cross-validated nnU-Net baseline. These results suggest that training on all available cases provides advantages over cross-validation in smaller datasets. This work underscores the effectiveness of nnU-Net in this setting and contributes to the broader goal of developing reliable AI tools for neuro-oncology in underrepresented global contexts.