Bayesian Particle Instance Segmentation for Electron Microscopy Image Quantification

This repository contains the official implementation of bpartis — a Bayesian deep neural network for nanoparticle instance segmentation.

Try an in-browser demo on your electron microscopy images here.

If you would like to use a pretrained bpartis model, we strongly recommend using imagedataextractor to do so.

import cv2
from imagedataextractor.segment import ParticleSegmenter

image = cv2.imread('<path/to/image>')  # PIL can also be used
segmenter = ParticleSegmenter()

segmentation, uncertainty, _ = segmenter.segment(image)

More detailed information can be found in the imagedataextractor segmentation documentation.

If you are interested in training bpartis to reproduce our results, follow these steps:

1. Clone the repository

git clone https://github.com/by256/bpartis.git

2. Install requirements

python3 -m pip install -r requirements.txt

3. Download the EMPS dataset from here.

4. Train the BPartIS model on the EMPS dataset.

python bpartis/train.py --data-dir=<path/to/emps/dir/> --device=cuda --epochs=300 --save-dir=bpartis/saved_models/

If you use bpartis in your work, please cite the following work:

B. Yildirim, J. M. Cole, "Bayesian Particle Instance Segmentation for Electron Microscopy Image Quantification", J. Chem. Inf. Model. (2021) https://doi.org/10.1021/acs.jcim.0c01455

@article{doi:10.1021/acs.jcim.0c01455,
	author = {Yildirim, Batuhan and Cole, Jacqueline M.},
	title = {Bayesian Particle Instance Segmentation for Electron Microscopy Image Quantification},
	journal = {Journal of Chemical Information and Modeling},
	volume = {61},
	number = {3},
	pages = {1136-1149},
	year = {2021},
	doi = {10.1021/acs.jcim.0c01455},
	note ={PMID: 33682402},
	URL = {https://doi.org/10.1021/acs.jcim.0c01455}
}

This project was financially supported by the Science and Technology Facilities Council (STFC) and the Royal Academy of Engineering (RCSRF1819\7\10).