Standalone cellfinder cell detection algorithm

This package implements the cell detection algorithm from Tyson, Rousseau & Niedworok et al. (2021) without any dependency on data type (i.e. it can be used outside of whole-brain microscopy).

cellfinder-core supports the cellfinder software for whole-brain microscopy analysis, and the algorithm can also be implemented in napari using the cellfinder napari plugin.

cellfinder-core supports Python >=3.7, and works across Linux, Windows, and should work on most versions of macOS (although this is not tested).

Assuming you have a Python environment set up (e.g. using conda), you can install cellfinder-core with:

pip install cellfinder-core

Once you have installed napari. You can install napari either through the napari plugin installation tool, or directly from PyPI with:

pip install cellfinder-napari

N.B. To speed up cellfinder, you need CUDA & cuDNN installed. Instructions here.

Before using cellfinder-core, it may be useful to take a look at the preprint which outlines the algorithm.

The API is not yet fully documented. For an idea of what the parameters do, see the documentation for the cellfinder whole-brain microscopy image analysis command-line tool (cell candidate detection, cell candidate classification). It may also be useful to try the cellfinder napari plugin so you can adjust the parameters in a GUI.

from cellfinder_core.main import main as cellfinder_run
import tifffile

signal_array = tifffile.imread("/path/to/signal_image.tif")
background_array = tifffile.imread("/path/to/background_image.tif")

voxel_sizes = [5, 2, 2] # in microns
detected_cells = cellfinder_run(signal_array,background_array,voxel_sizes)

The output is a list of imlib Cell objects. Each Cell has a centroid coordinate, and a type:

print(detected_cells[0])
# Cell: x: 132, y: 308, z: 10, type: 2

Cell type 2 is a "real" cell, and Cell type 1 is a "rejected" object (i.e. not classified as a cell):

from imlib.cells.cells import Cell
print(Cell.CELL)
# 2

print(Cell.NO_CELL)
# 1

If you want to save the detected cells for use in other BrainGlobe software (e.g. the cellfinder napari plugin), you can save in the cellfinder XML standard:

from imlib.IO.cells import save_cells
save_cells(detected_cells, "/path/to/cells.xml")

You can load these back with:

from imlib.IO.cells import get_cells
cells = get_cells("/path/to/cells.xml")

cellfinder-core supports most array-like objects. Using Dask arrays allows for lazy loading of data, allowing large (e.g. TB) datasets to be processed. cellfinder-core comes with a function (based on napari-ndtiffs) to load a series of image files (e.g. a directory of 2D tiff files) as a Dask array. cellfinder-core can then be used in the same way as with a numpy array.

from cellfinder_core.main import main as cellfinder_run
from cellfinder_core.tools.IO import read_with_dask

signal_array = read_with_dask("/path/to/signal_image_directory")
background_array = read_with_dask("/path/to/background_image_directory")

voxel_sizes = [5, 2, 2] # in microns
detected_cells = cellfinder_run(signal_array,background_array,voxel_sizes)

import tifffile
from pathlib import Path

from cellfinder_core.detect import detect
from cellfinder_core.classify import classify

signal_array = tifffile.imread("/path/to/signal_image.tif")
background_array = tifffile.imread("/path/to/background_image.tif")
voxel_sizes = [5, 2, 2] # in microns

home = Path.home()
install_path = home / ".cellfinder" # default

start_plane=0
end_plane=-1
trained_model=None
model_weights=None
model="resnet50_tv"
batch_size=32
n_free_cpus=2
network_voxel_sizes=[5, 1, 1]
soma_diameter=16
ball_xy_size=6
ball_z_size=15
ball_overlap_fraction=0.6
log_sigma_size=0.2
n_sds_above_mean_thresh=10
soma_spread_factor=1.4
max_cluster_size=100000
cube_width=50
cube_height=50
cube_depth=20
network_depth="50"

cell_candidates = detect.main(
    signal_array,
    start_plane,
    end_plane,
    voxel_sizes,
    soma_diameter,
    max_cluster_size,
    ball_xy_size,
    ball_z_size,
    ball_overlap_fraction,
    soma_spread_factor,
    n_free_cpus,
    log_sigma_size,
    n_sds_above_mean_thresh,
)

if len(cell_candidates) > 0: # Don't run if there's nothing to classify
    classified_cells = classify.main(
        cell_candidates,
        signal_array,
        background_array,
        n_free_cpus,
        voxel_sizes,
        network_voxel_sizes,
        batch_size,
        cube_height,
        cube_width,
        cube_depth,
        trained_model,
        model_weights,
        network_depth,
    )

The training data needed are matched pairs (signal & background) of small (usually 50 x 50 x 100um) images centered on the coordinate of candidate cells. These can be generated however you like, but I recommend using the Napari plugin.

cellfinder-core comes with a 50-layer ResNet trained on ~100,000 data points from serial two-photon microscopy images of mouse brains (available here).

Training the network is likely simpler using the command-line interface or the Napari plugin, but it is possible through the Python API.

from pathlib import Path
from cellfinder_core.train.train_yml import run as run_training

# list of training yml files
yaml_files = [Path("/path/to/training_yml.yml)]

# where to save the output
output_directory = Path("/path/to/saved_training_data")

home = Path.home()
install_path = home / ".cellfinder"  # default

run_training(
    output_directory,
    yaml_files,
    install_path=install_path,
    learning_rate=0.0001,
    continue_training=True, # by default use supplied model
    test_fraction=0.1,
    batch_size=32,
    save_progress=True,
    epochs=10,
)

More documentation about cellfinder and other BrainGlobe tools can be found here.

This software is at a very early stage, and was written with our data in mind. Over time we hope to support other data types/formats. If you have any questions or issues, please get in touch by email, on the forum or by raising an issue.

cellfinder takes a stitched, but otherwise raw dataset with at least two channels:

• Background channel (i.e. autofluorescence)
• Signal channel, the one with the cells to be detected:

Raw coronal serial two-photon mouse brain image showing labelled cells

Classical image analysis (e.g. filters, thresholding) is used to find cell-like objects (with false positives):

Candidate cells (including many artefacts)

A deep-learning network (ResNet) is used to classify cell candidates as true cells or artefacts:

Cassified cell candidates. Yellow - cells, Blue - artefacts

If you find this plugin useful, and use it in your research, please cite the preprint outlining the cell detection algorithm:

Tyson, A. L., Rousseau, C. V., Niedworok, C. J., Keshavarzi, S., Tsitoura, C., Cossell, L., Strom, M. and Margrie, T. W. (2021) “A deep learning algorithm for 3D cell detection in whole mouse brain image datasets’ PLOS Computational Biology, 17(5), e1009074 https://doi.org/10.1371/journal.pcbi.1009074

If you use this, or any other tools in the brainglobe suite, please let us know, and we'd be happy to promote your paper/talk etc.