Scholarlink

In many African healthcare settings, the adoption of deep learning-based medical imaging models remains limited by inadequate computational resources. This study presents a fast and lightweight adaptation of the nnU-Net framework for brain tumor segmentation, specifically designed for deployment in low-resource environments. The model integrates quantization-aware training, structured pruning, and patch size optimization to reduce computational and memory demands while maintaining high segmentation performance.Evaluation on the BraTS 2023 benchmark dataset demonstrated Dice scores of 0.8501, 0.8572, and 0.8310 for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET), respectively, indicating robust agreement with expert annotations. The lightweight configuration allows real-time inference on CPU-only systems with less than 8 GB RAM, making it feasible for hospitals and diagnostic centers across resource-limited regions.These results underscore the potential of deploying practical, accurate, and accessible AI-driven radiology tools across African clinical environments, advancing equitable access to state-of-the-art medical imaging technologies.

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

A Fast, Lightweight nnUNet-Based Brain Tumor Segmentation Model Optimized for Low-Resource African Settings

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • John Emeka,
  • Nwokoma Chidiebube,
  • Chika Ojiako

摘要

In many African healthcare settings, the adoption of deep learning-based medical imaging models remains limited by inadequate computational resources. This study presents a fast and lightweight adaptation of the nnU-Net framework for brain tumor segmentation, specifically designed for deployment in low-resource environments. The model integrates quantization-aware training, structured pruning, and patch size optimization to reduce computational and memory demands while maintaining high segmentation performance.Evaluation on the BraTS 2023 benchmark dataset demonstrated Dice scores of 0.8501, 0.8572, and 0.8310 for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET), respectively, indicating robust agreement with expert annotations. The lightweight configuration allows real-time inference on CPU-only systems with less than 8 GB RAM, making it feasible for hospitals and diagnostic centers across resource-limited regions.These results underscore the potential of deploying practical, accurate, and accessible AI-driven radiology tools across African clinical environments, advancing equitable access to state-of-the-art medical imaging technologies.