This page contains supplementary methods documenting case study 2, demonstrating scalable synapse detection and analysis in the NeuroData framework.
Starting point: aligned image data in ndstore Ending point: Automated spatial synapse density analysis, leveraging store, explore, and analyze NeuroData services.
/figs: Used to assemble the final figure 6 for the manuscript.spatialsyn_figs.ipynb: Jupyter notebook containing reproducible analysis and figure generation. (~10 minutes to run)put_mask*.py: Used to upload source mask data, created using ilastikmerged_mask_v5_layer23only_DSP.csv: Synapse density source data using the layer 2/3 mask (used for paper)merged_mask_v6_layerall_DSP.csv: Alternative mask for generating valid synapse density regionsget_matrix.py: python function used to reproducibly compute densities and metadata across all 100k windowstest_bock_matrix.sh: driver to execute get_matrix.py on MARCCsynapse_viz.ipynb: an interactive notebook to generate alternate figures and explore (under development)This section provides a more extensive overview of the processing pipeline used to analyze synaptic density across the Bock data volume
To generate putative synapse detections, we used a version of VESICLE-RF, a robust, state of the art classifier that balances performance and scalability. It is open-source and fully documented here: docs.neurodata.io/vesicle
We leveraged a version of the manual truth labels created by Kreshuk 2014, et. al for training and validation. We ensured that our validation set was non-overlapping with our training data.
Classifier training is efficient and was completed on a laptop. Classifier deployment was originally run using LONI pipeline, a meta-scheduler we used on a Son of Grid Engine cluster with approximately 400 cores. Processing took about 3 days, and used cuboids of data at scale 1 (8x8x45 nm) in an embarassingly parallel approach. Subsequently, we leveraged back-end ndstore tools to merge across the cuboid boundaries. A RESTful endpoint also exists to complete this step client side. This ultimately resulted in about 11.6 million putative synapses, at an estimated precision of 0.69 and recall of 0.62.
We investigated the result using our scalable visualization tools and observed that our detector did not generalize well to Layer I data, near the pia. To simplify our analysis and reduce bias, we created a mask (using Ilastik) of blank areas, our estimation of Layer I tissue, and other known areas of non-uniformity (e.g., blood vessels and soma). We uploaded this mask to NeuroData for inspection and provenance. We then developed a sampling strategy to compute density estimates. We divided the data into non-overlapping 5x5x5 um blocks (with a surrounding padded region). We counted synapses in the block that (i) had more than 50% of their labeled voxels inside the core cuboid, and (ii) were not touching a masked voxel. To compute density, we divided this by the number of microns in the non-masked region of that block (125 um^3 for regions with no masking). We saved this information to a csv file for each block and then merged all of the 100,000 csv files after processing. We computed this data using MARCC (marcc.jhu.edu), a large enterprise, slurm-based cluster.
Please see our Jupyter notebook for reproducible analysis steps to test for synaptic density uniformity (spatialsyn_figs.ipynb).
All of our data is viewable using NeuroData resources and can be examined, reproduced and extended.
viz.neurodata.io/MP4merged/. Traceability between the csv file and this viz project is straightforward using the provided metadata.openconnecto.me/data/public/ndpaper/