Local Shape Descriptors for Neuron Segmentation

Abstract

We present a simple, yet effective, auxiliary learning task for the problem of neuron segmentation in electron microscopy volumes. The auxiliary task consists of the prediction of Local Shape Descriptors (LSDs), which we combine with conventional voxel-wise direct neighbor affinities for neuron boundary detection. The shape descriptors are designed to capture local statistics about the neuron to be segmented, such as diameter, elongation, and direction. On a large study comparing several existing methods across various specimen, imaging techniques, and resolutions, we find that auxiliary learning of LSDs consistently increases segmentation accuracy of affinity-based methods over a range of metrics. Furthermore, the addition of LSDs promotes affinitybased segmentation methods to be on par with the current state of the art for neuron segmentation (Flood-Filling Networks, FFN), while being two orders of magnitudes more efficient—a critical requirement for the processing of future petabyte-sized datasets. Implementations of the new auxiliary learning task, network architectures, training, prediction, and evaluation code, as well as the datasets used in this study are publicly available as a benchmark for future method contributions.

Petascale neural circuit reconstruction: automated methods

“…These errors show that it can be challenging for a convolutional net to (1) erase the membranes of intracellular organelles while (2) marking plasma membranes, a tension that is intrinsic to the problem of neuronal boundary detection. Possible solutions may include (1) learning an object-centered representation (Lee et al 2021;Sheridan et al 2021) to better understand the scene, and (2) conservatively detecting all biological membranes to oversegment intracellular structures and later postprocessing them, possibly with the help of biological priors (Krasowski et al 2018).…”

Improved blind demixing methods for recovering dense neuronal morphology from barcode imaging data

“…Second, given an estimated barcode library, we would like improved methods for recovering the morphology corresponding to each barcode from the imagestack. One promising direction would be to use more sophisticated approaches for 3D neuronal recovery, adapting architectures and loss functions that have proven useful in the electron microscopy image processing literature [47,48]. We hope to explore these directions in future work.…”

Ultrastructural readout of in vivo synaptic activity for functional connectomics

“…For visualization and analysis, we used convolutional neural networks to automatically segment neurons (Sheridan et al, 2021), active zones (Heinrich et al, 2018), mitochondria and vesicles. For the latter, we used a 3D U-Net (Ronneberger et al, 2015) to predict, for each voxel, whether it was part of a vesicle, and if so, whether it was photoconverted or not.…”

“…We then trained a final 3D U-Net based on a previously published architecture (Funke et al, 2019). The network learned a 10D embedding describing local object shape to aid in the prediction of a 3D affinity graph (Sheridan et al, 2021). We taught the network to learn to predict zero affinities in areas of the ground-truth which were labelled as glia and artifacts to prevent false merges between neurons.…”