Toward Streaming Synapse Detection with Compositional ConvNets

Shibani Santurkar, David M. Budden, Alexander Matveev, Heather Berlin, Hayk Saribekyan, Yaron Meirovitch, Nir ShavitarXiv 2017 · OpenReview

Abstract

Connectomics is an emerging field in neuroscience that aims to reconstruct the 3-dimensional morphology of neurons from electron microscopy (EM) images. Recent studies have successfully demonstrated the use of convolutional neural networks (ConvNets) for segmenting cell membranes to individuate neurons. However, there has been comparatively little success in high-throughput identification of the intercellular synaptic connections required for deriving connectivity graphs. In this study, we take a compositional approach to segmenting synapses, modeling them explicitly as an intercellular cleft co-located with an asymmetric vesicle density along a cell membrane. Instead of requiring a deep network to learn all natural combinations of this compositionality, we train lighter networks to model the simpler marginal distributions of membranes, clefts and vesicles from just 100 electron microscopy samples. Our approach to synapse detection is both more accurate than previous state-of-the-art (7% higher recall and 5% higher F1-score) and yields a 20-fold speed-up compared to the previous fastest implementations. We demonstrate by reconstructing the first complete, directed connectome from the largest available anisotropic microscopy dataset (245 GB) of mouse somatosensory cortex (S1) in just 9.7 hours on a single shared-memory CPU system.

The Challenge of Scale and Complexity

The field of connectomics aims to map the "wiring diagram" of the brain by reconstructing every neuron and synapse from high-resolution electron microscopy (EM) images. While researchers have become proficient at segmenting the shapes of neurons (the "nodes" of the graph), identifying the synapses (the "edges") remains a massive bottleneck.

This challenge is two-fold. First, the scale is immense; a tiny 1 mm³ cube of brain tissue generates petabytes of data. Existing high-accuracy deep learning models are prohibitively slow, often taking hundreds of hours to process a single gigabyte. Second, synapses are small, sparse, and visually diverse. Previous attempts to increase speed using "hand-crafted" features often lacked the accuracy needed for reliable scientific discovery. The goal of this work is a "microscope-pace" pipeline—a system fast enough to process data as quickly as modern microscopes can stream it.

Methods: A Compositional Approach

The core innovation of this study is a compositional approach to detection. Instead of training one massive, "black-box" deep network to identify a synapse directly, the researchers break the synapse down into its biological components: neuron membranes, intercellular clefts, and synaptic vesicles.

The compositional pipeline: marginal ConvNets identify membranes, clefts, and vesicles, which are then integrated into a final synapse detection.

The system uses a bank of three lightweight ConvNets (specifically a "MaxoutNet" architecture) to generate probability maps for these three features in parallel. Because these "marginal" features are simpler and more abundant than the synapses themselves, the networks require very little training data—only 100 samples. Finally, a rules-based heuristic module integrates these maps. It applies biological constraints—for example, a synapse must be located on a membrane and must be flanked by a high density of vesicles—to filter out false positives and infer the direction of the connection.

Key Results and Throughput

The system achieved an F1-score of 0.847, outperforming the previous state-of-the-art deep learning models and Random Forest methods. More importantly, the speed improvements were transformative: the system is 11,000x faster than the most accurate previous ConvNet and 20x faster than the previous fastest Random Forest implementation.

3D reconstruction of the 245 GB mouse somatosensory cortex (S1) dataset, completed in under 10 hours.

To demonstrate real-world utility, the researchers reconstructed the first complete, directed connectome from a 245 GB dataset of the mouse somatosensory cortex (S1) in just 9.7 hours on a single machine. This proves that with smart engineering—such as using SIMD instructions and efficient job scheduling—standard multi-core CPUs can outperform expensive GPU clusters for this specific task.

Conclusion

By explicitly modeling the "compositionality" of synapses, this research reduces the search space for AI and allows for the use of shallower, more efficient networks. While the system is highly accurate, future work will address remaining "split errors" in neuron segmentation and explore automated learning of integration rules to further improve robustness across different tissue types.

Summary updated

Abstract

Connectomics is an emerging field in neuroscience that aims to reconstruct the 3-dimensional morphology of neurons from electron microscopy (EM) images. Recent studies have successfully demonstrated the use of convolutional neural networks (ConvNets) for segmenting cell membranes to individuate neurons. However, there has been comparatively little success in high-throughput identification of the intercellular synaptic connections required for deriving connectivity graphs. In this study, we take a compositional approach to segmenting synapses, modeling them explicitly as an intercellular cleft co-located with an asymmetric vesicle density along a cell membrane. Instead of requiring a deep network to learn all natural combinations of this compositionality, we train lighter networks to model the simpler marginal distributions of membranes, clefts and vesicles from just 100 electron microscopy samples. These feature maps are then combined with simple rules-based heuristics derived from prior biological knowledge. Our approach to synapse detection is both more accurate than previous state-of-the-art (7% higher recall and 5% higher F1-score) and yields a 20-fold speed-up compared to the previous fastest implementations. We demonstrate by reconstructing the first complete, directed connectome from the largest available anisotropic microscopy dataset (245 GB) of mouse somatosensory cortex (S1) in just 9.7 hours on a single shared-memory CPU system. We believe that this work marks an important step toward the goal of a microscope-pace streaming connectomics pipeline.
BibTeX
@misc{santurkar2017toward,
  author = {Shibani Santurkar and David M. Budden and Alexander Matveev and Heather Berlin and Hayk Saribekyan and Yaron Meirovitch and Nir Shavit},
  title = {{Toward Streaming Synapse Detection with Compositional ConvNets}},
  year = {2017},
  eprint = {1702.07386},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1702.07386},
}
Figure