ErrorNet: Error Detection on Meshes of Neuronal Reconstruction

Abstract

This thesis is focused on error detection on reconstructed neurons from EM cortex data. We present ErrorNet, a neural network that operates on 3D point clouds coming from meshes of neural segments and aims to identify merge and split errors. We prove that the network can identify errors accurately, at least when restricted to a certain subset of cells. Our classifier operates on sparse mesh data and uses minimal preprocessing, unlike previous error detection methods. Our starting point is Pointnet, a pioneer in 3D classification using unordered point clouds. One of our improvements is the incorporation of multiple views in the Pointnet pipeline, which drastically improves the network error detection performance. We also show promising results when the locations are shifted a few $\mu m$, indicating that the network has the potential of being used in practice.