Model Pipeline

Electron microscopy is used to reveal the structure of brain tissue at the micron and submicron scales. However, the ability to analyze the spatial relationships between various cellular structures as well as the arrangement of organelles within cells is limited in two-dimensional electron microscopy images. Reconstruction from serial section transmission electron microscopy (ssTEM) offers the possibility of recreating neuronal and glial processes and their organelles in spatially realistic three-dimensional models. Our EM data comes from area CA1 of the hippocampus (courtesy of Dr. Kristen Harris).

We outline the overall computational framework that we have been developing for converting imaging data to spatially realistic meshed models above. There are four major stages: 2-D image processing (white box), 2-D geometry processing (blue box), 2-D to 3-D reconstruction (yellow box), and 3-D geometry processing (green box). An additional set of procedures, the reducing procedures (orange box), are used to generate 1-D multi-compartment models suitable for NEURON from 3-D reconstructions. The red boxes in Fig. 2 represent components of ongoing development.

The image above shows several stages of our processing pipeline. The boundaries of all cellular elements (dendrites, axons, and glia) are traced and classified in each 2-D image slice (Fig. a). These contours undergo further processing steps (referred to as 2-D curation procedures) that accomplish such goals as removing boundary overlaps and resampling by spline-fitting. The 2-D contours are tiled into 3-D objects (Fig. d) using our ContourTiler algorithm (implemented in our VolRover software package). This provides us with 3-D spatially realistic reconstructions of dendrites (Fig. e), axons, and glial processes within a reconstructed neuropil. However, the quality of the initial, tiled surface mesh (Fig. f) is not sufficient to be useful for computational simulations. Our set of 3-D curation procedures removes boundary overlaps in 3-D and applies mesh quality improvement algorithms, such as smoothing, decimation (reducing the number of triangles), and fitting by algebraic spline functions, in order to generate a forest of non-intersecting, high quality-meshed neuronal (Fig. 3g) and glial processes. These reconstructions are then ready for use in simulations of neuronal activity.