Computer simulations of neuronal morphology and automated digital tracing:

A win/win image-data-model reconstruction loop.


Giorgio A. Ascoli


The automated digital tracing of neural arbors is a serious bottleneck towards the reconstruction of entire neural circuits at the cellular level. Whole dendritic and axonal trees can be visualized with various combinations of histological and microscopic techniques. Digital tracing transforms the voxel-based content of the image stacks into a set of interconnected vectors representing the coordinates and orientation of each branch. The resulting 3D reconstructions ( are compact, detailed, and useful, but this manual process is too labor intensive (hours to days for dendrites, or many months for the full axon of a single projection cell). Despite several available programs for automated reconstruction, the vast majority of axons and dendrites are still traced manually. The confidence of computational scientists that the problem is very solvable contrasts the frustration of experimental neuroanatomists with the practical inadequacy of the existing solutions. A collegial competition is even being organized to challenge algorithm developers with data provided by potential users. Such an unconventional interaction will likely lead to significant advancements and an explicit assessment of the remaining obstacles. Why is automation so challenging? The answer may reside in the distinction between the older, incomplete description of vision as a bottom-up process (pixels-edges-features-objects-scenes, or retina-LGN-V1-V2-higher areas), and the now accepted view of mixed expectation/sensation (a continuous interaction between imagination and the image). Most algorithms for automated neuronal reconstructions lack the top-down modulation component. We developed several statistical models to simulate neuronal morphology by stochastic resampling experimental data (Donohue&Ascoli: PLoS CB 08 + J Comput NS 05; Samsonovich&Ascoli: Hippocampus 05 + J Neurosci Res 03; Ascoli et al.: Anat Embryol 01). We propose that these simulation frameworks can be leveraged to dramatically improve automated tracing routines with an iterative approach. The computational models generated from a first set of (either manual or automated) reconstructions are used as the “expectation” inference in the tracing of the next generation of neurons, by assigning probabilistic weights to all ambiguous choices. As more neurons are reconstructed, the increasing statistical power makes the model more predictive and thus the successive generations of traced data more accurate.