Scholarlink

Neuronal labeling–a process that transforms image volumes into 3D geometries and skeletons of cells–bottlenecks the study of brain function, connectomics and pathology. Domain scientists are limited by reconstruction methods which have insufficient accuracy to be used without proofreading. Further gains in fidelity of automatic techniques may largely come from supervised methods (e.g., deep learning), however these require unique training data which is rare especially for novel imaging or experimental settings. Current protocols for label creation are fraught with issues in automation, which are exacerbated by modern data size (30 TB in this study). Furthermore for training data, quantity but not necessarily quality is paramount, which is why augmentation and weak supervision is so effective. We introduce an application that can produce adequate training labels automatically from raw lightsheet microscopy images of whole brains. This pipeline is effective at producing both weak and strong labels of two critical categories of neurons: cell bodies (somas) and whole cells.

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

High Throughput Training Label Generation from Whole Brain Images

  • 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 labeling–a process that transforms image volumes into 3D geometries and skeletons of cells–bottlenecks the study of brain function, connectomics and pathology. Domain scientists are limited by reconstruction methods which have insufficient accuracy to be used without proofreading. Further gains in fidelity of automatic techniques may largely come from supervised methods (e.g., deep learning), however these require unique training data which is rare especially for novel imaging or experimental settings. Current protocols for label creation are fraught with issues in automation, which are exacerbated by modern data size (30 TB in this study). Furthermore for training data, quantity but not necessarily quality is paramount, which is why augmentation and weak supervision is so effective. We introduce an application that can produce adequate training labels automatically from raw lightsheet microscopy images of whole brains. This pipeline is effective at producing both weak and strong labels of two critical categories of neurons: cell bodies (somas) and whole cells.