Automated analysis of synaptic connections' ultrastructure in very large stacks of Electron Microscopy images

• Applicant: Anna Kreshuk, Ph.D.
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Human Cognitive and Systems Neuroscience
• Term: from 2014 to 2018
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 252629829

Connectomics is an emerging domain of neuroscience, which aims to establish the relation between physical anatomy and function of neural circuits. While multiple biological experiments allowed for understanding of the function of most body organs, this goal remains elusive for the most complex of all --- the brain. In essence, the neural circuits are composed of functional units --- the neurons --- and connections between them --- the synapses or synaptic junctions. Two competing imaging requirements make this task difficult: detailed analysis of synapse ultrastructure requires nanometer resolution, while the size of the neural circuits requires a large 3D field of view. Current connectomics experiments therefore have to acquire very large stacks of high-resolution Electron Microscopy images. Synaptic connections' analysis in these images relies on the localization of comparatively small structures of interest in Terabyte-scale volumes of image data. This task is difficult, very time-consuming and error-prone even for smaller image stacks and downright prohibitively complex for image stacks large enough to contain even a short neural circuit. The necessity of manual annotation of the images has now become the main bottleneck on the way to comprehensive analysis of neural circuits. In recent years, several algorithms were proposed to automate the two main neural image analysis tasks: segmentation of neural cells and detection and segmentation of synaptic contact sites. However, while a lot of progress has been made, segmentation of neural cells is still not possible by fully automated means. Synaptic contact sites can now be detected reliably in image stacks with isotropic resolution, but no methods exist for the segmentation of synaptic contacts' ultrastructure. For image stacks with anisotropic resolution, even synaptic contact detection can not be performed in automated manner.The proposed project aims alleviate this bottleneck by developing an algorithm to localize the synaptic connections and extract their quantitative characteristics. A supervised machine learning algorithm will learn to recognize the main elements of the synaptic contact ultrastructure, based on expert annotations on a tiny sub-volume of the data. It will first detect the different ultrastructure elements independently and then iteratively bind them together in a probabilistic graphical model. The algorithm will be made part of a user-friendly software toolkit, which will allow the neuroscientists to perform the annotation interactively, to process large data volumes offline and to export synaptic contact characteristics for follow-up analysis. We expect the proposed algorithm to fully automate one the major tasks of neural image analysis --- synaptic contact segmentation --- and assist with the other --- segmentation of neural cells, by allowing the segmentation efforts to concentrate on the relevant parts of the data volume.

• DFG Programme: Research Grants