3D Reconstruction of Neural Circuits from Serial EM Images

A basic requirement for reconstructing and understanding complete circuit diagrams of neuronal processing units is the availability of electron microscopic 3D data sets of large ensembles of neurons. A recently developed technique, "Serial Block Face Scanning Electron Microscopy" (SBFSEM, Denk and Horstmann 2004) allows automatic sectioning and imaging of biological tissue inside the vacuum chamber of a scanning electron microscope. Image stacks generated with this technology have a resolution sucient to distinguish different cellular compartments, including synaptic structures. Such an image stack contains thousands of images and is recorded with a voxel size of 23 nm in the x- and y-directions and 30 nm in the z-direction. Consequently a tissue block of 1 mm3 produces 63 terabytes of data. Therefore new concepts for managing large data sets and automated image processing are required. I developed an image segmentation and 3D reconstruction software, which allows precise contour tracing of cell membranes and simultaneously displays the resulting 3D structure. The software contains two stand-alone packages: Neuron2D and Neuron3D, both oering an easy-to-operate graphical user interface (GUI). The software package Neuron2D provides the following image processing functions: • Image Registration: Combination of multiple SBFSEM image tiles. • Image Preprocessing: Filtering of image stacks. Implemented are Gaussian and Non-Linear-Diusion lters in 2D and 3D. This step enhances the contrast between contour lines and image background, leading to a higher signal-to-noise ratio, thus further improving detection of membrane borders. • Image Segmentation: The implemented algorithms extract contour lines from the preceding image and automatically trace the contour lines in the following images (z-direction), taking into account the previous image segmentation. They also permit image segmentation starting at any position in the image stack. In addition, manual interaction is possible. To visualize 3D structures of neuronal circuits the additional software Neuron3D was developed. The program relies on the contour line information provided by Neuron2D to implement a surface reconstruction algorithm based on dynamic time warping. Additional rendering techniques, such as shading and texture mapping, are provided. The detailed anatomical reconstruction provides a framework for computational models of neuronal circuits. For example in ies, where moving retinal images lead to appropriate course control signals, the circuit reconstruction of motion-sensitive neurons can help to further understand the neural processing of visual motion in ies.