Scholarlink

Gliomas are the most aggressive primary brain tumors and represent a major public health challenge. In sub-Saharan Africa (SSA), the scarcity of radiologists, the degraded quality of MRI scans, and the lack of infrastructure considerably limit early detection. We present TerangaNet, a 3D UNet architecture optimized for low-resource environments. Designed with a symmetrical encoder-decoder, InstanceNorm normalization, and a robust TorchIO-based augmentation pipeline, TerangaNet delivers accurate segmentation despite the noise and variability of African MRIs. On the BraTS-Africa 2023–2024 dataset (n = 95), TerangaNet achieved an average Dice score of 81.9%, significantly outperforming conventional UNet (77.1%, p < 0.001). Our visual results show an ability to better detect complex tumor areas, including edema and necrosis. The ablation study confirms that the integration of augmentations and a combined loss function improves performance by + 4.8 Dice. The model is 2.4 × lighter than SwinUNETR and compatible with standard hardware (8–12 GB GPU). All code, pre-trained weights, and reproduction scripts are publicly available. TerangaNet represents a breakthrough for equitable artificial intelligence, deployable in under-equipped hospitals in the Global South. All code, pre-trained weights, and reproduction scripts are publicly available at: TerangaNet – Team Teranga (SPARK GitHub).

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

TerangaNet: An Optimized 3D U-Net for Brain Tumor Segmentation in Sub-Saharan African MRI Volumes

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kéba Faye,
  • Abdourahmane Balde,
  • Racky Barro Diatta,
  • Abdoul Wahab Soumare,
  • Penda Ka,
  • Mohameth DIA,
  • Khoudia Sow,
  • Doudou Mohamet Gaye,
  • Mohamadou Bamba Diop,
  • Magatte Diouf,
  • Marième Dieng Fall,
  • Guy Mbatchou,
  • Aondona Iorumbur,
  • Raymond Confidence,
  • Udunna Anazodo

摘要

Gliomas are the most aggressive primary brain tumors and represent a major public health challenge. In sub-Saharan Africa (SSA), the scarcity of radiologists, the degraded quality of MRI scans, and the lack of infrastructure considerably limit early detection. We present TerangaNet, a 3D UNet architecture optimized for low-resource environments. Designed with a symmetrical encoder-decoder, InstanceNorm normalization, and a robust TorchIO-based augmentation pipeline, TerangaNet delivers accurate segmentation despite the noise and variability of African MRIs. On the BraTS-Africa 2023–2024 dataset (n = 95), TerangaNet achieved an average Dice score of 81.9%, significantly outperforming conventional UNet (77.1%, p < 0.001). Our visual results show an ability to better detect complex tumor areas, including edema and necrosis. The ablation study confirms that the integration of augmentations and a combined loss function improves performance by + 4.8 Dice. The model is 2.4 × lighter than SwinUNETR and compatible with standard hardware (8–12 GB GPU). All code, pre-trained weights, and reproduction scripts are publicly available. TerangaNet represents a breakthrough for equitable artificial intelligence, deployable in under-equipped hospitals in the Global South. All code, pre-trained weights, and reproduction scripts are publicly available at: TerangaNet – Team Teranga (SPARK GitHub).