Because synapses are sub-micron objects they can be unequivocally detected only with EM. However, describing synapses of even a single neuron requires reconstructing neuropil over several hundred microns. In principle, this range of scales can be covered by serial section EM. In practice, however, neuronal circuit reconstruc-tions are rare because this technique is laborious and time-consuming. The most extensive reconstruction, that of the C. elegans nervous system (White et al. 1986) relied on ~10⁴ serial sections and took about ten years. Even this reconstruction was not completed – the exact synaptic connectivity of the ventral cord has never been published. We are finalizing the C. elegans reconstruction by using White’s EM photographs and lab notebooks, as well as newly obtained EM photographs (collaboration with D. Hall, AECOM). This reconstruction is all but complete (Figure 1) and provides the first “proof of principle” connectivity matrix of the full nervous system.

Figure 1. Synaptic connectivity matrix for 280 non-pharyngeal neurons in C. elegans. Each dot indicates a synaptic connection between neurons in the corresponding row and column. Chemical synapses are blue, gap junctions - red (B. Chen, D. Hall & D. B. Chklovskii, in preparation).

Larger scale circuit reconstructions cannot be done manually and require developing automated reconstruction algorithms (collaboration with A. Koulakov). A successful algorithm must produce a connectivity matrix of neuronal circuits of the kind shown in Figure 1. In addition, it must extract the shapes of neurons and the locations of individual synapses. Although this project is extremely challenging, we believe that this is the right time to do it because of recent developments in digital image processing, continuing improvements in computer performance and the existence of high level programming environments such as MATLAB. We validate every step of development on real-world data through numerous collaborations with experimentalists.

Alignment of serial sections is the first step for EM reconstructions. The need for alignment arises from irregular placement of the sections on the grid and from their physical distortion, including stretching and shear. Traditionally, such alignment is performed manually and is rather time-consuming, especially if significant distortion is present. We developed an automatic algorithm for section alignment that overcomes distortion (including non-linear). To assess the quality of alignment we image the two consecutive sections in red and green channels, Figure 2.

Figure 2. Automatic alignment of the two consecutive EM sections of Drosophila eye (collaboration with I. Meinertzhagen, Dalhouise). Red and green channels show (negative) photographs of consecutive EM sections. Quality of alignment, including non-linear transformation, is evident by the large amount of overlap (yellow).

The second step is segmentation of 2D images into cross-sections of individual neurons. Traditionally, 2D segmentation is done by manually tracing contours of the objects of interest. Our algorithm performs segmentation automatically and for the whole volume at once. One of the challenges was to overcome a non-uniform contrast across the image. Figure 3 shows a typical automatic segmentation and illustrates robustness of the algorithm towards non-uniform contrast.

The third step of the reconstruction is to assemble 2D cross-sections into 3D shapes of axons and dendrites. We have developed an automatic algorithm based on evaluating proximity and similarity of the 2D cross-sections. To validate our algorithm we compare the results of automatic reconstructions with those done manually by the experts on the same dataset, Figure 4.

	Automatic Reconstruction // Manual Reconstruction
Figure 3. Automatic segmentation of EM section from rat hippocampus (http://synapses.mcg.edu). Cross-sections belonging to different neurons are shown with different false colors. Notice robustness of the algorithm towards contrast non-uniformity common in EM photographs.	Figure 4. Dendrite segment reconstructed by our automatic algorithm compared with the manual reconstruction by the experts (J. Fiala & K. Harris, http://synapses.mcg.edu). Although the general quality of automatic reconstruction is reasonable, some discrepancies are present. We are developing manual editing tools for on-line correction of the automatic reconstruction.

Finally, we need to identify synapses and assemble the wiring diagram. We expect that fully reconstructing a nervous system of an organism, such as Drosophila, will make a major impact on neurobiology.

Research Home           Section 2