A Novel Semi-automated Proofreading and Mesh Error Detection Pipeline for Neuron Extension

Abstract

The immense scale and complexity of neuronal electron microscopy (EM) datasets pose significant challenges in data processing, validation, and interpretation, necessitating the development of efficient, automated, and scalable error-detection methodologies. This paper proposes a novel approach that employs mesh processing techniques to identify potential error locations near neuronal tips. Error detection at tips is a particularly important challenge since these errors usually indicate that many synapses are falsely split from their parent neuron, injuring the integrity of the connectomic reconstruction. Additionally, we draw implications and results from an implementation of this error detection in a semi-automated proofreading pipeline. Manual proofreading is a laborious, costly, and currently necessary method for identifying the errors in the machine learning based segmentation of neural tissue. This approach streamlines the process of proofreading by systematically highlighting areas likely to contain inaccuracies and guiding proofreaders towards potential continuations, accelerating the rate at which errors are corrected.

Keywords: EM connectomics; Neural Morphology; Proofreading.

Figure 1:

Errors in 15 expert-labelled neurons. A: A representative example of a poorly-reconstructed neuron mesh. Pink points indicate the tips of dendrites, error or not. Yellow points indicate axonal tips. B: Number of synapses and 3D volume missing from each neuron. This is the number of synapses and amount of volume located in all of the true extensions, as labeled by an expert proofreader. C: The number of ‘encapsulated’ errors within each of the expert-labeled neurons as detailed in section 2.1

Figure 2:

Performance of the error detector against the evaluation dataset. A: A representative example of a success case for facet detection. In the xz plane, the dropped slices that cause the split are visible. B: A representative example of a success case for defect detection. A darkly stained body interrupts the process. C: Total number of suggestions versus the total number of error locations for each neuron. D: Precision, recall, and F1 scores of the facet and defect errors together, with double counts removed.

Figure 3:

Comparison of the time each expert proofreader took on semi-automated vs manual tasks. Every proofreader who completed both task types performed faster when given automated tasks, and the automated tasks also had long tails in the positive direction, indicating that many ‘easy’ tasks were able to be completed almost immediately.