Scholarlink

Neuronal reconstruction–a process that transforms image volumes into 3D geometries and skeletons of cells–bottlenecks the study of brain function, connectomics and pathology. Domain scientists need exact and complete segmentations to study subtle topological differences. Existing methods are disk-bound, dense-access, coupled, single-threaded, algorithmically unscalable and require manual cropping of small windows and proofreading of skeletons due to low topological accuracy. Designing a data-intensive parallel solution suited to a neurons’ shape, topology and far-ranging connectivity is particularly challenging due to I/O and load-balance, yet by abstracting these vision tasks into strategically ordered specializations of search, we progressively lower memory by 4 orders of magnitude. This enables 1 mouse brain to be fully processed in-memory on a single server, at 67 \(\times \) the scale with 870 \(\times \) less memory while having 78% higher automated yield than APP2, the previous state of the art in performant reconstruction.

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

Gossamer: Scaling Image Processing and Reconstruction to Whole Brains

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Karl Marrett,
  • Keivan Moradi,
  • Chris Sin Park,
  • Ming Yan,
  • Chris Choi,
  • Muye Zhu,
  • Masood Akram,
  • Sumit Nanda,
  • Qing Xue,
  • Hyun-Seung Mun,
  • Adriana E. Gutierrez,
  • Mitchell Rudd,
  • Brian Zingg,
  • Gabrielle Magat,
  • Kathleen Wijaya,
  • Hongwei Dong,
  • X. William Yang,
  • Jason Cong

摘要

Neuronal reconstruction–a process that transforms image volumes into 3D geometries and skeletons of cells–bottlenecks the study of brain function, connectomics and pathology. Domain scientists need exact and complete segmentations to study subtle topological differences. Existing methods are disk-bound, dense-access, coupled, single-threaded, algorithmically unscalable and require manual cropping of small windows and proofreading of skeletons due to low topological accuracy. Designing a data-intensive parallel solution suited to a neurons’ shape, topology and far-ranging connectivity is particularly challenging due to I/O and load-balance, yet by abstracting these vision tasks into strategically ordered specializations of search, we progressively lower memory by 4 orders of magnitude. This enables 1 mouse brain to be fully processed in-memory on a single server, at 67 \(\times \) the scale with 870 \(\times \) less memory while having 78% higher automated yield than APP2, the previous state of the art in performant reconstruction.