Gliomas are by far the most frequent primary tumors affecting the central nervous system (CNS). In Sub-Saharan Africa (SSA), gliomas pose a significant health burden due to high mortality rates that are caused by the late presentation of the disease and limited access to advanced diagnostic imaging. The prevailing state-of-the-art brain tumor segmentation algorithms designed for rapid automated lesion characterization are largely trained on datasets from high-income countries, limiting their general application in Sub-Saharan African (SSA) populations. These models frequently encounter challenges in SSA settings due to the region's lower-quality MRI scans, which present with poor image contrast and resolution, as well as distinct disease characteristics like the late-stage presentation of tumors. In this study, we proposed a deep learning framework that shifts the focus from a model-centric approach to a data processing focused approach, for accurate segmentation of the brain glioma sub-regions. The study made use of the BraTS-Africa 2025 Challenge dataset from SSA populations. Our approach combines modality-specific data augmentation and a tissue-adaptive postprocessing technique, built upon an optimized 3D U-Net architecture. Preliminary results demonstrate promising Dice scores of 0.74 for the enhancing tumor (ET), 0.75 for the tumor core (TC), and 0.85 for the whole tumor (WT), as well as HD95 values of 13.2, 14.5, and 14.8 for the ET, TC, and WT respectively, suggesting improved robustness and spatial accuracy on SSA-specific data. This work represents a step toward reducing global health disparities and advancing precision medicine for glioma patients in SSA.

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

MAPS-Glioma: Modality-Specific Augmentation and Tissue-Adaptive Postprocessing for Robust Glioma Segmentation in Resource-Limited Settings

  • Ayomide B. Oladele,
  • Helena Machibya,
  • Mariam Kaoneka,
  • Frederick Lyimo,
  • Debora Hoza,
  • Immaculata Kafumu,
  • Idris Olalekan,
  • Jeremiah Fadugba,
  • Dong Zhang,
  • Aondona Iorumbur,
  • Raymond Confidence,
  • Nicephorus Rutabasibwa,
  • Ugumba M. Kwikima

摘要

Gliomas are by far the most frequent primary tumors affecting the central nervous system (CNS). In Sub-Saharan Africa (SSA), gliomas pose a significant health burden due to high mortality rates that are caused by the late presentation of the disease and limited access to advanced diagnostic imaging. The prevailing state-of-the-art brain tumor segmentation algorithms designed for rapid automated lesion characterization are largely trained on datasets from high-income countries, limiting their general application in Sub-Saharan African (SSA) populations. These models frequently encounter challenges in SSA settings due to the region's lower-quality MRI scans, which present with poor image contrast and resolution, as well as distinct disease characteristics like the late-stage presentation of tumors. In this study, we proposed a deep learning framework that shifts the focus from a model-centric approach to a data processing focused approach, for accurate segmentation of the brain glioma sub-regions. The study made use of the BraTS-Africa 2025 Challenge dataset from SSA populations. Our approach combines modality-specific data augmentation and a tissue-adaptive postprocessing technique, built upon an optimized 3D U-Net architecture. Preliminary results demonstrate promising Dice scores of 0.74 for the enhancing tumor (ET), 0.75 for the tumor core (TC), and 0.85 for the whole tumor (WT), as well as HD95 values of 13.2, 14.5, and 14.8 for the ET, TC, and WT respectively, suggesting improved robustness and spatial accuracy on SSA-specific data. This work represents a step toward reducing global health disparities and advancing precision medicine for glioma patients in SSA.