Project

Back to overview

Neuron Reconstruction from Electron Microscopy Images

Applicant Cook Matthew
Number 149861
Funding scheme Project funding
Research institution Institut für Neuroinformatik Universität Zürich Irchel und ETH Zürich
Institution of higher education University of Zurich - ZH
Main discipline Information Technology
Start/End 01.12.2013 - 31.05.2017
Approved amount 329'800.00
Show all

Keywords (5)

Neural Reconstruction; Neural Image Segmentation; Connectomics; Electron Microscopy Image Segmentation; Biomedical Image Processing

Lay Summary (German)

Lead
Für eine tiefreichende Untersuchung der Funktionsweise des Gehirn ist es notwendig, seine Bestandteile, die Neuronen, und deren Verschaltung zu analysieren. Neuronen bilden komplexe Netzwerke mit unzähligen synaptischen Verbindungen, die so klein sind, dass sie in der klassischen Lichtmikroskopie nicht sichtbar sind. Elektron-Mikroskopie (EM) erreicht eine ausreichende Vergrösserung, allerdings resultiert selbst ein Fruchtfliegengehirn in mehreren Terrabyte an Bilddaten. Eine manuelle Analyse dieser Daten würde mehrere Jahrzehnte dauern.
Lay summary

Inhalt und Ziel des Forschunsantrages

Unser Ziel ist es, automatische Methoden zur Analyse der EM-Bilder zu entwickeln und der Neurobiologie zugänglich zu machen. Dazu gilt es einerseits, vorhandene Ansätze zu verbessern. Dies betrifft sowohl deren Genauigkeit als auch deren Fähigkeit, sehr grosse Datensätze zu verarbeiten. Andererseits streben wir eine engere Zusammenarbeit mit den Neurowissenschaften an. In die eine Richtung kann die Entwicklung automatischer Methoden von Expertenwissen profitieren. In die andere Richtung sollten die Methoden den Neurowissenschaftlern in einem einfachen Interface zur Verfügung gestellt werden.

Wissenschaftlicher und gesellschaftlicher Kontext

Wir erwarten, dass unsere Arbeit ein wichtiger Beitrag zur automatischen Bildanalyse im Allgemeinen leisten wird. Eine erfolgreiche Lösung muss in der Lage sein, mit verrauschten Bildern, fehlenden Informationen, mehrdeutigen Situationen und wenigen Annahmen über das Problem umzugehen. Dabei sollten die entwickelten Algorithmen effizient sein, um mehrere Terrabyte an Daten in kurzer Zeit verarbeiten zu können. Diese Bedingungen treffen auch auf viele andere Bildanalyseprobleme ausserhalb der Neurowissenschaften zu. Erkenntnisse in der EM Bildanalyse können daher übertragbar sein.

Direct link to Lay Summary Last update: 10.03.2014

Responsible applicant and co-applicants

Employees

Publications

Publication
TED: A Tolerant Edit Distance for segmentation evaluation
Funke Jan, Klein Jonas, Moreno-Noguer Francesc, Cardona Albert, Cook Matthew (2017), TED: A Tolerant Edit Distance for segmentation evaluation, in Methods, 115, 119-127.
Efficient convolutional neural networks for pixelwise classification on heterogeneous hardware systems
Tschopp Fabian, Martel Julien N. P., Turaga Srinivas C., Cook Matthew, Funke Jan (2016), Efficient convolutional neural networks for pixelwise classification on heterogeneous hardware systems, in ISBI, 1225-1228, 1225-1228.
Structured learning of assignment models for neuron reconstruction to minimize topological errors
Funke Jan, Klein Jonas, Moreno-Noguer Francesc, Cardona Albert, Cook Matthew (2016), Structured learning of assignment models for neuron reconstruction to minimize topological errors, in Biomedical Imaging (ISBI), 2016 IEEE 13th International Symposium on, 607-611, 607-611.
Tracking of microtubules in anisotropic volumes of neural tissue
Buhmann Julia M, Gerhard Stephan, Cook Matthew, Funke Jan (2016), Tracking of microtubules in anisotropic volumes of neural tissue, in Biomedical Imaging (ISBI), 2016 IEEE 13th International Symposium on, 326-329, 326-329.
Learning to segment: training hierarchical segmentation under a topological loss
Funke Jan, Hamprecht Fred A, Zhang Chong (2015), Learning to segment: training hierarchical segmentation under a topological loss, in International Conference on Medical Image Computing and Computer-Assisted Intervention, 268-275, 268-275.
Candidate sampling for neuron reconstruction from anisotropic electron microscopy volumes
Funke Jan, Martel Julien NP, Gerhard Stephan, Andres Bjoern, Ciresan Dan C, Giusti Alessandro, Gambardella Luca M, Schmidhuber Juergen, Pfister Hanspeter, Cardona Albert, Cook Matthew (2014), Candidate sampling for neuron reconstruction from anisotropic electron microscopy volumes, in International Conference on Medical Image Computing and Computer-Assisted Intervention, 17-24, 17-24.

Collaboration

Group / person Country
Types of collaboration
Stephan Saalfeld, Janelia Research Campus United States of America (North America)
- in-depth/constructive exchanges on approaches, methods or results
Gene Myers, Max Planck Institute for Molecular Cell Biology and Genetics Germany (Europe)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
Albert Cardona, Janelia Research Campus United States of America (North America)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
- Exchange of personnel

Use-inspired outputs

Software

Name Year
TED (tolerant edit distance) 2017
bigcat 2016
caffe extension for heterogeneous hardware systems 2016


Associated projects

Number Title Start Funding scheme
160133 A Higher Level for Neuron Reconstruction 01.07.2016 Project funding

Abstract

Neuroscientists are currently imaging multi-terabyte volumes of neural tissue for the purpose of reconstructing the neuronal circuitry [5,7,8,18], but this unfortunately requires an impractical amount of time to be spent on manually labeling these images. This young field, often termed "connectomics" [26], challenges the computer vision and machine learning community to develop accurate and efficient techniques for automated neuronal reconstruction. Despite being an active field of computer vision research, there is currently no application of automated neuron reconstruction methods to real biological questions. From the published methods and from our own experience we believe that this is due to two key problems. The first problem is that current algorithms are not practical. The accuracy of state-of-the-art methods is not good enough, meaning that every automatic reconstruction has to be reviewed manually, which is hardly faster than a manual reconstruction in the first place. Furthermore, current methods scale poorly with the size of the data set. At present, they can only process datasets tens of thousands of times smaller than what is needed in neuroscience. The second problem is that there is a significant gap between the neuroscience side and the computer science side of neuronal reconstruction. Expert knowledge from neuroscientists is barely included in automatic or interactive reconstruction systems. In the other direction, automatic reconstruction algorithms from computer science are barely used by neuroscientists because they are not made available as usable and publicly available tools. Despite coming to the field relatively recently, our method for the automatic reconstruction of neurons from anisotropic EM-image stacks is at present the most accurate one [14]. Furthermore, we have integrated some of our methods into the open source project CATMAID [28] to make our algorithms available to the wider community. Building on our areas of expertise, we propose to address the two key problems of our field through the following four subprojects. We will undoubtedly find more opportunities similar to these subprojects over the course of this project. First, the development of a simple visual language will allow us to incorporate expert knowledge as priors in the reconstruction process. Second, by optimizing low-level classifiers with respect to high-level objectives (including expert knowledge), we will unify stages of the reconstruction process that have heretofore been deleteriously separate. Third, the development of a problem decomposition algorithm with global optimality guarantees will allow us to scale up to large image stacks, allowing us to use experts’ large-scale prior knowledge and avoid small-scale approximation artifacts. Fourth, the development of an interactive user interface in CATMAID will get the expert into the reconstruction loop, helping with cases that the automated system finds difficult, which in turn can be used to improve the automated system. These improvements over the current state of the art can be seen to directly contribute not only to accuracy and scalability, but also to closing the gap between the neuroscience and computer science approaches to the reconstruction challenge.
-