TY  - GEN
A1  - Hanslovsky, Philipp
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/27375/
N2  - Neuroscientists are increasingly convinced that it is necessary to reconstruct
the precise wiring and synaptic connectivity of biological nervous systems to
eventually decipher their function. The urge to reconstruct ever larger and more
complete synaptic wiring diagrams of animal brains has created an entire new
subfield of neuroscience: Connectomics. The reconstruction of connectomes is
difficult because neurons are both large and small. They project across distances
of many millimeters but each individual neurite can be as thin as a few tens of
nanomaters. In order to reconstruct all neurites in densely packed neural tissues,
it is necessary to image this tissue at nanometer resolution which, today, is only
possible with 3D electron microscopy (3D-EM).
Over the last decade, 3D-EM has become significantly more reliable than ever
before. Today, it is possible to routinely image volumes of up to a cubic millimeter,
covering the entire brain of small model organisms such as that of the fruit fly
Drosophila melanogaster. These volumes contain tens or hundreds of tera-voxels
and cannot be analyzed manually. Efficient computational methods and tools
are needed for all stages of connectome reconstruction: (1) assembling distortion
and artifact free volumes from serial section EM, (2) precise automatic recon-
struction of neurons and synapses, and (3) efficient and user-friendly solutions
for visualization and interactive proofreading. In this dissertation, I present new
computational methods and tools that I developed to address previously unsolved
problems covering all of the above mentioned aspects of EM connectomics.
In chapter 2, I present a new method to correct for planar and non-planar axial
distortion and to sort unordered section series. This method was instrumental for
the first ever acquisition of a complete brain of an adult Drosophila melanogaster
imaged with 3D-EM.
Machine learning, in particular deep learning, and the availability of public
training and test data has had tremendous impact on the automatic reconstruction
of neurons and synapses from 3D-EM. In chapter 3, I present a novel artificial
neural network architecture that predicts neuron boundaries at quasi-isotropic
resolution from non-isotropic 3D-EM. The goal is to create a high-quality over-
segmentation with large three-dimensional fragments for faster manual proof-
reading.
In chapter 4, I present software libraries and tools that I developed to support
the processing, visualization, and analysis of large 3D-EM data and connectome
reconstructions. Using this software, we generated the largest currently existing
training and test data for connectome reconstruction from non-isotropic 3D-EM.
I will particularly emphasize my flexible interactive proof-reading tool Paintera
that I built on top of the libraries and tools that I have developed over the last
four years.
CY  - Heidelberg
TI  - Isotropic Reconstruction of Neural Morphology
from Large Non-Isotropic 3D Electron Microscopy
AV  - public
ID  - heidok27375
KW  - Computer Vision
Machine Learning
AI
Neuron Reconstruction
Segmentation
Big Data
Neuroscience
Connectomics
Y1  - 2019///
ER  -