Learning cellular morphology with neural networks

Philipp J. Schubert, Sven Dorkenwald, Michał Januszewski, Viren Jain, Joergen Kornfeld

Abstract

Reconstruction and annotation of volume electron microscopy data sets of brain tissue is challenging but can reveal invaluable information about neuronal circuits. Significant progress has recently been made in automated neuron reconstruction as well as automated detection of synapses. However, methods for automating the morphological analysis of nanometer-resolution reconstructions are less established, despite the diversity of possible applications. Here, we introduce cellular morphology neural networks (CMNs), based on multi-view projections sampled from automatically reconstructed cellular fragments of arbitrary size and shape. Using unsupervised training, we infer morphology embeddings (Neuron2vec) of neuron reconstructions and train CMNs to identify glia cells in a supervised classification paradigm, which are then used to resolve neuron reconstruction errors. Finally, we demonstrate that CMNs can be used to identify subcellular compartments and the cell types of neuron reconstructions. Volume electron microscopy data of brain tissue can tell us much about neural circuits, but increasingly large data sets demand automation of analysis. Here, the authors introduce cellular morphology neural networks and successfully automate a range of morphological analysis tasks.

Introduction

Mapping the connectome—the dense web of wiring in the brain—requires Volume Electron Microscopy (VEM), which produces massive 3D datasets. While automated tracing (segmentation) has improved significantly, understanding the morphology of these traces remains a major bottleneck. Researchers face a "3D memory wall" where 3D Convolutional Neural Networks (CNNs) are too computationally expensive to process high-resolution features across large volumes. Furthermore, manual annotation of cell types and subcellular compartments is prohibitively slow, and automated tools frequently produce "false mergers" between neurons and glia, necessitating tedious manual proofreading.

Methods

The core innovation presented by Schubert et al. is the Cellular Morphology Neural Network (CMN) architecture. To sidestep the computational costs of 3D voxels, the system samples locations along a reconstructed cell fragment and generates 2D depth-map projections from multiple perspectives.

Overview of the CMN architecture, multi-view projection sampling, and tested applications.

The CMN framework utilizes two primary training paradigms:

  1. Neuron2vec (Unsupervised): A triplet-loss training method allows the network to learn a 10-dimensional latent space where similar-looking cellular fragments are clustered together without human labels.
  2. Supervised Classification: CMNs are trained on manually annotated data for specific tasks, such as identifying cell types (e.g., Medium Spiny Neurons) or subcellular compartments (axon vs. dendrite).

Notably, the CMNs incorporate multi-channel data, including information on mitochondria and synaptic vesicles, which serve as critical morphological signposts for identifying functional compartments.

Results

The CMN approach demonstrated high accuracy across several critical tasks in connectomics, processing roughly 25–58 multi-views per second with GPU acceleration.

Unsupervised Feature Discovery

The Neuron2vec embeddings successfully discovered biologically meaningful categories. As shown in the latent space visualization, the network naturally clustered features like spiny dendrites and smooth axons based purely on shape similarity, proving that deep learning can discover meaningful biological categories autonomously.

Visualization of the Neuron2vec latent space (Z0, Z1) showing the clustering of biologically distinct neuronal features.

Automated Error Correction and Classification

  • Glia Detection: The model achieved an F1-score of 0.979 in identifying glial cells. This allowed for the automated "splitting" of false mergers by treating reconstructions as graphs and removing nodes classified as glia.
  • Subcellular Mapping: CMNs identified axons, dendrites, and somas with an F1-score of approximately 0.98, outperforming traditional methods that rely on hand-crafted skeleton features.
  • Cell Type Identification: The framework successfully classified four major neuron types in the songbird basal ganglia (Area X) with high precision, even when processing only small, local fragments of the cell.

Discussion

The findings suggest that 2D projection-based models can outperform 3D architectures in morphological analysis because they can ingest higher-resolution data for the same computational cost. By capturing the fine texture of cell surfaces and internal organelles, CMNs provide a scalable solution for large-scale VEM datasets. Future work may involve integrating these morphological priors directly into the initial segmentation process and exploring the potential for "universal" morphological embeddings across different species and brain regions.

Summary updated

Abstract

Reconstruction and annotation of volume electron microscopy data sets of brain tissue is challenging but can reveal invaluable information about neuronal circuits. Significant progress has recently been made in automated neuron reconstruction as well as automated detection of synapses. However, methods for automating the morphological analysis of nanometer-resolution reconstructions are less established, despite the diversity of possible applications. Here, we introduce cellular morphology neural networks (CMNs), based on multi-view projections sampled from automatically reconstructed cellular fragments of arbitrary size and shape. Using unsupervised training, we infer morphology embeddings (Neuron2vec) of neuron reconstructions and train CMNs to identify glia cells in a supervised classification paradigm, which are then used to resolve neuron reconstruction errors. Finally, we demonstrate that CMNs can be used to identify subcellular compartments and the cell types of neuron reconstructions. Volume electron microscopy data of brain tissue can tell us much about neural circuits, but increasingly large data sets demand automation of analysis. Here, the authors introduce cellular morphology neural networks and successfully automate a range of morphological analysis tasks.
BibTeX
@inproceedings{schubert2019learning,
  author = {Philipp J. Schubert and Sven Dorkenwald and Michał Januszewski and Viren Jain and Joergen Kornfeld},
  title = {{Learning cellular morphology with neural networks}},
  booktitle = {Nature Communications},
  year = {2019},
  url = {https://openreview.net/forum?id=06JwAJXpC1},
}
Figure