Abstract
Purpose: :
A fundamental limitation to understanding the function of retinal circuits is the lack of complete wiring (connection) diagrams. In many nervous systems, local circuits are comprised of neurons extending several hundreds of microns. The tissue volumes that are required to reconstruct such circuits are on the order of 10^8 υm3 representing several terabytes of pixel information at the pixel size (20-30nm) needed to trace all neuronal processes. It is therefore necessary to automate both the acquisition and analysis of the data.
Methods: :
Data acquisition of 3D data with the required resolution is possible in a fully automated way by using serial block-face scanning electron microscopy (SBFSEM). Images are acquired from the face of a tissue block with slices as thin as 25 nm removed by ultra-thin sectioning with an oscillating diamond knife. This technique replaces the difficult and labor-intensive process of manually cutting, mounting, and imaging sections in a transmission electron microscope.
Results: :
The segmentation of neurons from images is dependent on a high signal contrast between structures of interest. We have developed a staining strategy that labels cell surfaces with an electron dense product. This method leaves the cytosol and organelles largely unstained and thus results in a high contrast between intracellular space and cell surfaces, allowing the profiles of neurons to be easily identified.The manual tracing of each and every neuron through a large volume is prohibitively slow, hence the tracing of neurons must be automated. A manually segmented sub-volume becomes a training dataset for machine learning algorithms. A crucial step is the validation of the automated segmentations to obtain an estimate of the reconstruction reliability. We have developed tools that allow browsing through large 3D data sets and the efficient creation of 3D skeletons by human tracers. The comparison between skeletons and automated segmentations allows the estimation of error probabilities, which are needed for the optimization of the classification and segmentation algorithms. Also, this data provides a measure for the reliability of the reconstructed circuitry. We are also attempting to correlate functional recordings with circuit structure. Proof of principle experiments have shown that bulk loading of calcium indicators in the mammalian retina allow functional signals to be imaged without significantly damaging tissue ultrastructure.
Conclusions: :
Serial block-face electron microscopy, in combination with new staining methods and image segmentation algorithms, can be used for automating the analysis of retinal neuronal circuits.
Keywords: microscopy: electron microscopy • imaging/image analysis: non-clinical • retina