The brain contains a vast number of neurons that are connected with each other through synapses, thereby forming a complex anatomical network that mediates information ?ow within the brain. The brain ?wiring diagram? will be a foundational tool for elucidating the function and dysfunction of brains. Electron microscopy (EM) is widely considered to be the gold standard for neuronal level circuit recon- struction. Currently, a major and serious bottleneck in this ?eld is image segmentation and reconstruc- tion. It is estimated that the data analysis accuracy and throughput are lagging behind data acquisition by orders of magnitude. This project aims at dramatically improving the accuracy and throughput of brain EM image analysis, thereby enabling accurate and ef?cient reconstruction of neuronal level brain maps. Speci?cally, this project is built up on the recent success in deep learning methods, which are dominant tools for EM image analysis. A central and unresolved challenge of using deep learning for segmentation is how to achieve the con?icting goals of integrating suf?cient contextual features while preserving full-resolution information. This project will develop a novel residual encoder-decoder model to achieve these two goals simultaneously (Aim 1). In current deep learning segmentation methods, the labels of each pixel are predicted independently. To fully consider the brain topological structure and couple the predictions of spatially adjacent pixels, this project will develop a hybrid recurrent and convo- lutional network model (Aim 2). In this model, the recurrent network is integrated with the convolutional network to incorporate the multi-dimensional structural information. When combined with Aim 1, these methods are expected to dramatically improve the accuracy of EM image segmentation. In most cur- rent deep learning segmentation methods, the training and/or prediction stages require the extraction of patches centered on each pixel. This step forms a bottleneck that limits the overall throughput. This project will develop novel techniques to achieve whole-image training and prediction (Aim 3). These approaches will enable very ef?cient training and segmentation, thereby dramatically increasing the throughput of EM image analysis.

Public Health Relevance

Wiring diagrams of the brain circuits serve as a foundational tool for studying brain function and dys- function. This project aims at developing advanced computational methods for boosting the accuracy and throughput of brain circuit reconstruction from electron microscopy images. Comparative circuit analysis of normal and disease brains would shed light on how brain circuits go awry in psychiatric and neurological disorders.

Agency
National Institute of Health (NIH)
Institute
National Institute of Neurological Disorders and Stroke (NINDS)
Type
Exploratory/Developmental Grants (R21)
Project #
7R21NS102828-03
Application #
9864551
Study Section
Biodata Management and Analysis Study Section (BDMA)
Program Officer
Gnadt, James W
Project Start
2017-05-01
Project End
2019-04-30
Budget Start
2019-02-11
Budget End
2019-04-30
Support Year
3
Fiscal Year
2018
Total Cost
Indirect Cost
Name
Texas Engineering Experiment Station
Department
Type
DUNS #
847205572
City
College Station
State
TX
Country
United States
Zip Code
77843
Zeng, Tao; Wu, Bian; Ji, Shuiwang (2017) DeepEM3D: approaching human-level performance on 3D anisotropic EM image segmentation. Bioinformatics 33:2555-2562