VESICLE Dataset for Synaptic Connectomics

• VESICLE Dataset is a comprehensive resource providing volumetric EM data and expert-curated synapse annotations for mammalian connectomics.
• It includes approximately 50,000 manually labeled synapses and reference implementations (VESICLE-RF and VESICLE-CNN) with F1 scores up to 0.85.
• The dataset integrates seamlessly via RAMON standards and OCP APIs, enabling scalable, reproducible brain circuit reconstruction and analysis.

The VESICLE dataset is a comprehensive resource for the volumetric evaluation of synaptic interfaces in mammalian brain tissue using large‐scale electron microscopy (EM) and computer vision approaches. It comprises an anisotropic EM volume of mouse somatosensory cortex, expert‐curated synapse annotations, reference algorithms for synapse detection, standardized benchmarks, and is distributed through the Open Connectome Project platform. This framework addresses the challenges of synapse identification in non-post‐stained, anisotropic data and serves as a foundation for large‐scale connectomics and brain circuit reconstruction (Roncal et al., 2014).

1. Imaging Properties and Volume Characteristics

The raw data consist of serial section EM images of mouse somatosensory cortex acquired at approximately 3 × 3 × 30 nm resolution, with an in‐plane pixel size originally near 5 nm. For practical processing, images are downsampled by a factor of two along the X and Y axes, resulting in a final voxel sampling of $s_{x} = 6$ nm, $s_{y} = 6$ nm, and $s_{z} = 30$ nm. The full inscribed cuboid covers approximately $60,000\,\mu m^3$ of tissue, comprising about $N = 1 \times 10^{10}$ voxels and occupying $220$ GB on disk in its original form or $56$ GB post‐downsampling.

The relationship between the volume $V$ and the number of voxels $N$ is expressed by

$V = N \cdot s_x \cdot s_y \cdot s_z.$

This imaging protocol preserves native anisotropy without post‐staining or isotropic slicing, a deviation from many preceding synaptic datasets.

2. Annotation Protocols and Ground Truth Generation

Approximately $50,000$ synaptic contacts within the EM volume are manually annotated by expert neurobiologists to provide ground truth labels for supervised algorithm development and benchmarking. The annotation process combines recognized EM criteria, including the presence of dark post‐synaptic densities, membrane apposition, and vesicle clustering to validate putative boutons. Annotators systematically browse the three‐dimensional image stacks, defining dense label‐map volumes for each synapse and marking the centroid coordinates. Labels are archived in both dense three‐dimensional bitmask formats and as lists of $3$D centroid coordinates.

All annotations and metadata adhere to the RAMON (Reusable Annotation Markup for Open coNnectomes) data model, facilitating standardization and interoperability within connectomics pipelines.

3. Dataset Organization and Accessibility

The VESICLE dataset and all associated annotations are hosted on the Open Connectome Project (OCP) infrastructure and are accessible via RESTful HDF5/TIFF interfaces. The data organization is as follows:

Directory Path	Content Description	Format(s)
/vesicle/em/	Raw EM volumes	HDF5
/vesicle/labels/	Synapse label maps	HDF5, TIFF (compressed)
/vesicle/coords/	Synapse centroid coordinates	CSV
/vesicle/config/	Model definitions and Random Forest parameters	Caffe, config files

Two separate $1024 \times 1024 \times 100\,\mu m^3$ subvolumes are reserved for benchmarking: AC4 (training) and AC3 (testing), ensuring non‐overlapping evaluation for reproducibility. Access is open source under a permissive research license at http://openconnecto.me/vesicle, with an encouragement to cite Gray Roncal et al. (BMVC 2015) and adhere to non‐commercial and attribution‐only terms.

4. Reference Implementations and Performance Benchmarks

Two principal reference implementations are provided:

• VESICLE-RF: A Random Forest classifier utilizing $200,000$ balanced synapse/non‐synapse voxel samples and ten handcrafted features including multi‐scale intensity, local binary patterns, gradient magnitudes, structure tensor responses, and vessel proximity. Training requires approximately $10$ minutes and $8$ GB RAM; full‐volume inference (on the $60,000\,\mu m^3$ cube) takes about $10$ minutes on a multicore CPU (MATLAB implementation).
• VESICLE-CNN: A deep learning classifier with three convolutional and two fully connected layers, designed in the style of N3‐architectures, which performs pixelwise classification using $65 \times 65$ input patches. Training on one GPU requires roughly $56$ hours; inference on the test cuboid is approximately $39$ hours.

When evaluated on the AC3 held‐out test volume, both VESICLE‐RF and VESICLE‐CNN outperform the prior method of Becker et al. 2013 in synaptic precision–recall, with the following performance metrics: - VESICLE‐RF: maximum $F_1$ score near $0.80$. - VESICLE‐CNN: maximum $F_1$ score near $0.85$. At high recall levels (e.g., recall $\approx 0.9$), precision remains approximately $0.7$.

VESICLE‐RF demonstrates scalability, having been executed on the entire $220$ GB downsampled dataset in a distributed manner (OCP/LONI Pipeline), detecting approximately $50,335$ synapses in about $24$ hours wall‐time on a CPU cluster. Application to a post‐stained $5$ TB FIB‐SEM dataset yielded $11.6$ million synapses (precision $\approx 0.69$, recall $\approx 0.60$).

5. Integration, Interoperability, and Use Cases

The dataset is designed for seamless integration into standard connectomics workflows. Since all data conform to the RAMON schema and OCP REST API, VESICLE outputs can be incorporated directly into segmentation pipelines based on techniques such as flood‐filling networks or supervoxel methods, as well as visualized or manually proofread using platforms such as CATMAID or Neuroglancer.

The primary application is semi‐ or fully automated brain circuit reconstruction. Detected synaptic contacts can be mapped onto neuronal fragments, supporting rapid estimation of the brain's connectivity graph. This approach is intended to accelerate the field’s shift towards the analysis of petabyte‐scale, high‐throughput image datasets.

6. Prospective Extensions and Utility for Connectomics

Planned extensions include region‐growing algorithms for comprehensive synapse morphology extraction, the development of very‐deep CNN architectures and $3$D‐aware input features, and guided active learning protocols that exploit vesicle‐based priors from the Random Forest pipeline. The framework encourages experimentation, pipeline development, and benchmarking within the broader connectomics community.

In summary, the VESICLE dataset represents a rigorously benchmarked, large‐scale, expert‐labeled anisotropic EM volume with associated open‐source algorithms and infrastructure engineered to establish a scalable platform for mammalian synapse detection and volumetric connectomics (Roncal et al., 2014).