Bootstrapper

• Installation
• Getting Started
• Examples
• Citation
• Funding

A toolkit for bootstrapping and refining 3D instance segmentations and models from sparse 2D labels.

• 🖥️ CLI: Command-line interface for training, prediction, post-processing
• ⚙️ Configurable Models: Flexible and hackable models using config files
• 🧱 Blockwise Processing: Efficient handling of large volumes

Tested on Ubuntu 22.04, Rocky Linux 8.10, and macOS 15.1.1 (Apple Silicon).

Installation

Note: Rust is necessary if you wish to use mwatershed for segmentation. Install from rustup.rs

To install bootstrapper, we first recommend creating a new conda environment:

conda create -n bs -c conda-forge python=3.12 graph-tool boost
conda activate bs

Then, install bootstrapper with:

pip install git+https://github.com/ucsdmanorlab/bootstrapper.git

Getting Started

Bootstrapper has the following main components:

• Volumes: 3D Zarr image arrays, with or without labels or masks
• Models: PyTorch models for training and prediction
• Commands: Configurable CLI commands for training, prediction, post-processing, and evaluation

The following are the primary Bootstrapper commands, typically run in the given order:

• bs prepare : Prepare volumes and config files for the following steps

• bs train : Train a model

• bs predict : Run inference on a volume

• bs segment : Segment affinities

• bs evaluate : Evaluate segmentations against ground truth or model predictions

• bs filter : Refine segmentations to create pseudo-ground truth

A round is a cycle of the above commands run on a set of volumes.

• Use bs prepare to create config files for one or multiple rounds.
• Refined segmentations from one round become training labels for the next round.

Bootstrapper also has:

• bs run: Runs the appropriate command for the given config file.
• bs view : A wrapper for neuroglancer -d
• bs utils offers functions for data manipulation and preprocessing.

Examples

• CREMI

Citation

For questions about the preprint or this repository, please contact vvenu@utexas.edu

If you find Bootstrapper useful in your research, please consider citing our preprint:

@article {Thiyagarajan2024.06.14.599135,
	author = {Thiyagarajan, Vijay Venu and Sheridan, Arlo and Harris, Kristen M. and Manor, Uri},
	title = {A deep learning-based strategy for producing dense 3D segmentations from sparsely annotated 2D images},
	year = {2024},
	doi = {10.1101/2024.06.14.599135},
	URL = {https://www.biorxiv.org/content/early/2024/06/15/2024.06.14.599135},
}

Funding

Chan-Zuckerberg Imaging Scientist Award DOI https://doi.org/10.37921/694870itnyzk from the Chan Zuckerberg Initiative DAF, an advised fund of Silicon Valley Community Foundation (funder DOI 10.13039/100014989).

NSF NeuroNex Technology Hub Award (1707356), NSF NeuroNex2 Award (2014862)