| Popis: |
In order to understand how the brain works, we need to understand how its neural circuits process information. Electron microscopy remains the only imaging technique capable of providing sufficient resolution to reconstruct the dense connectivity between all neurons in a circuit. Automated electron microscopy techniques are approaching the point where usefully large circuits might be successfully imaged, but the development of automated reconstruction techniques lags far behind. No fully-automated reconstruction technique currently produces acceptably accurate reconstructions, and semi-automated approaches currently require an extreme amount of manual effort. This reconstruction bottleneck places severe limits on the size of neural circuits that can be reconstructed. Improved automated reconstruction techniques are therefore highly desired and under active development. The human brain contains ~86 billion neurons and ~80% of these are located in the cerebellum. Of these cerebellar neurons, the vast majority are granule cells. The axons of these granule cells are called parallel fibres and tend to be oriented in approximately the same direction, making 2+1D reconstruction approaches feasible. In this work we focus on the problem of reconstructing these parallel fibres and make four main contributions: (1) a model-based algorithm for reconstructing 2D parallel fibre cross-sections that achieves state of the art 2D reconstruction performance; (2) a fully-automated algorithm for reconstructing 3D parallel fibres that achieves state of the art 3D reconstruction performance; (3) a semi-automated approach for reconstructing 3D parallel fibres that significantly improves reconstruction accuracy compared to our fully-automated approach while requiring ~40 times less labelling effort than a purely manual reconstruction; (4) a "gold standard" ground truth data set for the molecular layer of the mouse cerebellum that will provide a valuable reference for the development and benchmarking of reconstruction algorithms. |
