Scholarlink

NeuroGleam is a deep-learning (DL) pipeline for fully automated segmentation of white-matter hyperintensities (WMH) on single-modal, low-resolution T2-FLAIR which is one of the most common MRI series in clinical care. Our preliminary investigation involved benchmarking of six DL architectures under four loss objectives. We found HRNet and its variants to be the best performing architecture and introduced a hyper-parameterized HRNet which was tuned using Bayesian optimization. Experiments span the Medical Image Computing and Computer Assisted Intervention (MICCAI) WMH Challenge dataset, and a 69-subject in-house clinical cohort called Assessing Population-based Radiological Brain Health in Stroke Epidemiology (APRISE). Metrics include Dice, Hausdorff95, lesion-wise sensitivity/F1, and average volume difference. The best HRNet achieves Dice 0.742 (MICCAI) and 0.651 (APRISE) with Hausdorff95 6.24–7.75 mm. Cross-dataset testing drops to Dice 0.523, underscoring domain-shift limits. We discuss practical design choices and outline ongoing work including transfer learning, domain adaptation and out-of-distribution detection to close the generalization gap and enable robust deployment.

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

NeuroGleam: Illuminating Small Vessel Disease Detection Through Deep Learning Based Segmentation of Brain MRI White Matter Hyperintensities

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Balaji Iyer,
  • Brady Williamson,
  • V. B. Surya Prasath,
  • Bruce J. Aronow,
  • Pooja Khatri,
  • Heidi Sucharew,
  • Vivek Khandwala,
  • Joseph LaPorta,
  • Lily Wang,
  • Rebecca Cornelius,
  • Mary Gaskill-Shipley,
  • Thomas Tomsick,
  • David Wang,
  • Thomas Maloney,
  • Paul S. Horn,
  • Janice Carrozzella,
  • Brett M. Kissela,
  • Achala Vagal

摘要

NeuroGleam is a deep-learning (DL) pipeline for fully automated segmentation of white-matter hyperintensities (WMH) on single-modal, low-resolution T2-FLAIR which is one of the most common MRI series in clinical care. Our preliminary investigation involved benchmarking of six DL architectures under four loss objectives. We found HRNet and its variants to be the best performing architecture and introduced a hyper-parameterized HRNet which was tuned using Bayesian optimization. Experiments span the Medical Image Computing and Computer Assisted Intervention (MICCAI) WMH Challenge dataset, and a 69-subject in-house clinical cohort called Assessing Population-based Radiological Brain Health in Stroke Epidemiology (APRISE). Metrics include Dice, Hausdorff95, lesion-wise sensitivity/F1, and average volume difference. The best HRNet achieves Dice 0.742 (MICCAI) and 0.651 (APRISE) with Hausdorff95 6.24–7.75 mm. Cross-dataset testing drops to Dice 0.523, underscoring domain-shift limits. We discuss practical design choices and outline ongoing work including transfer learning, domain adaptation and out-of-distribution detection to close the generalization gap and enable robust deployment.