Neuronal Subcompartment Classification and Merge Error Correction

Abstract

Recent advances in 3d electron microscopy are yielding ever larger reconstructions of brain tissue, encompassing thousands of individual neurons interconnected by millions of synapses. Interpreting reconstructions at this scale demands advances in the automated analysis of neuronal morphologies, for example by identifying morphological and functional subcompartments within neurons. We present a method that for the first time uses full 3d input (voxels) to automatically classify reconstructed neuron fragments as axon, dendrite, or somal subcompartments. Based on 3d convolutional neural networks, this method achieves a mean f1-score of 0.972, exceeding the previous state of the art of 0.955. The resulting predictions can support multiple analysis and proofreading applications. In particular, we leverage finely localized subcompartment predictions for automated detection and correction of merge errors in the volume reconstruction, successfully detecting 90.6% of inter-class merge errors with a false positive rate of only 2.7%.

A connectomic study of a petascale fragment of human cerebral cortex

“…2E). A classification model predicted axon, dendrite, astrocyte, soma, cilium, and axon initial segment classes at skeleton node locations distributed throughout each cell (Li et al, 2020). Occasional agglomeration errors could produce merges between nearby objects, such as a passing axon and dendrite (Fig.…”

“…The model outputs were probabilities for four classes: axon, dendrite, astrocyte, or soma, and models were trained via stochastic gradient descent with batch size 64 for up to 1. We built on the approach of Li et al (Li et al, 2020) to use subcompartment predictions to fix segmentation errors based on the observation that while FFN base segments are largely free of merge errors, occasionally in FFN agglomeration two base segments with inconsistent classes (e.g. axon versus dendrite) are erroneously merged.…”

“…Prior to processing, the agglomeration graph for each object was modified to remove cycles by breaking an arbitrary edge of each cycle. We also simplified on Li et al (Li et al, 2020) by considering only the highest probability predicted class label for each node, rather than working with predicted class probabilities per node.…”

“…Applying the same set of cuts to 20201916 resulted in segmentation 20201916c3. Because we removed somas first to avoid spurious suggested cuts at the soma / branch interface, we did not fix any agglomeration errors involving the soma cluster of base segments, although methods to address this were proposed for songbird data (Li et al, 2020).…”

Petascale neural circuit reconstruction: automated methods

“…Shapson-Coe et al 2021), or to inform automated techniques to clean an initial segmentation (A. Shapson-Coe et al 2021;H. Li et al 2020).…”