Cytokit is a collection of tools for quantifying and analyzing properties of individual cells in large fluorescent microscopy datasets with a focus on those generated from multiplexed staining protocols. This includes a GPU-accelerated image processing pipeline (via TensorFlow), CLI tools for batch processing of experimental replicates (often requiring conditional configuration, as things tend go wrong when capturing hundreds of thousands of microscope images over a period of hours or days), and visualization UIs (either Cytokit Explorer or CellProfiler Analyst).
Cytokit runs in a Python 3 environment but also comes (via Docker) with CellProfiler (Python 2) and Ilastik installations.
For more information, see: Cytokit: A single-cell analysis toolkit for high dimensional fluorescent microscopy imaging
Installing and configuring Cytokit currently involves little more than installing nvidia-docker and building or downloading the Cytokit container image, but this inherently limits support to Linux operating systems for GPU-acceleration. Additional limitations include:
- There is currently no CPU-only docker image
- Generating and running pipelines requires working knowledge of JupyterLab and a little tolerance for yaml/json files as well as command lines
- Only tiff files are supported as a raw input image format
- Deconvolution requires manual configuration of microscope attributes like filter wavelengths, immersion media, and numerical aperture (though support to infer much of this based on the imaging platform may be added in the future)
- 3 dimensional images are supported but cell segmentation and related outputs are currently 2 dimensional
- General system requirements include at least 24G RAM and 8G of GPU memory (per GPU)
Once nvidia-docker is installed, the container can be launched and used as follows:
nvidia-docker pull eczech/cytokit:latest
# Set LOCAL_IMAGE_DATA_DIR variable to a host directory for data sharing
# and persistent storage between container runs
export LOCAL_IMAGE_DATA_DIR=/tmp
# Run the container with an attached volume to contain raw images and results
nvidia-docker run --rm -ti -p 8888:8888 -p 8787:8787 -p 8050:8050 \
-v $LOCAL_IMAGE_DATA_DIR:/lab/data \
eczech/cytokitThis will launch JupyterLab on port 8888. After navigating to localhost:8888 and entering the access token
printed on the command line following nvidia-docker run, you can then run an example notebook
like cellular_marker_profiling_example, which can be found
at /lab/repos/cytokit/python/notebooks/examples in the JupyterLab file navigator.
To use a release-specific container, the instructions above can be modified as such where the below
example shows how to launch the 0.1.1 container:
nvidia-docker pull eczech/cytokit:0.1.1
export LOCAL_IMAGE_DATA_DIR=/tmp
nvidia-docker run --rm -ti -p 8888:8888 -p 8787:8787 -p 8050:8050 \
-v $LOCAL_IMAGE_DATA_DIR:/lab/data \
eczech/cytokit:0.1.1One of the goals of Cytokit is to make it as easy as possible to reproduce complicated workflows on big image datasets and to that end, the majority of the logic that drives how Cytokit functions is determined by json/yaml configurations.
Starting from template configurations like this sample Test Experiment and more realistically, this CODEX BALBc1 configuration, pipelines are meant to work as bash scripts executing small variants on these parameterizations for evaluation against one another. Here is a bash script demonstrating how this often works:
EXPERIMENT_DATA_DIR=/lab/data/201801-codex-lung
for REPLICATE in "201801-codex-lung-01" "201801-codex-lung-02"; do
DATA_DIR=$EXPERIMENT_DATA_DIR/$REPLICATE
# This command will generate 3 processing variants to run:
# v01 - Cell object determined as fixed radius from nuclei
# v02 - Cell object determined by membrane stain
# v03 - 5x5 grid subset with deconvolution applied and before/after channels extracted
cytokit config editor --base-config-path=template_config.yaml --output-dir=$DATA_DIR/output \
set processor.cytometry.segmentation_params.nucleus_dilation 10 \
save_variant v01/config reset \
set processor.cytometry.membrane_channel_name CD45 \
save_variant v02/config reset \
set acquisition.region_height 5 \
set acquisition.region_width 5 \
set processor.args.run_deconvolution True \
add operator '{extract: {name:deconvolution, channels:[raw_DAPI,proc_DAPI]}}' \
save_variant v03/config exit
# Run everything for each variant of this experiment
for VARIANT in v01 v02 v03; do
OUTPUT_DIR=$DATA_DIR/output/$VARIANT
CONFIG_DIR=$OUTPUT_DIR/config
cytokit processor run_all --config-path=$CONFIG_DIR --data-dir=$OUTPUT_DIR --output-dir=$OUTPUT_DIR
cytokit operator run_all --config-path=$CONFIG_DIR --data-dir=$OUTPUT_DIR
cytokit analysis run_all --config-path=$CONFIG_DIR --data-dir=$OUTPUT_DIR
done
doneThe above, when executed, would produce several things:
- 5D tiles with processed image data (which can be reused without having to restart from raw data)
- 5D tile extracts corresponding to user-defined slices (e.g. raw vs processed DAPI images above) as well as montages of these tiles (e.g. stitchings of 16 2048x2048 images on 4x4 grid into single 8192x8192 images)
- CSV/FCS files with single-cell data
- Final yaml configuration files representing how each variant was defined
For example, an ad-hoc extraction like this (which could also be defined in the configuration files):
cytokit operator extract --name='primary_markers' --z='best' \
--channels=['proc_dapi','proc_cd3','proc_cd4','proc_cd8','cyto_cell_boundary','cyto_nucleus_boundary']Would produce 5D hyperstack images that could be loaded into ImageJ and blended together:
Human T Cells stained for DAPI (gray), CD3 (blue), CD4 (red), CD8 (green) and with nucleus outline (light green), cell outline (light red)
After processing an experiment, the Explorer UI application can be run within the same docker container for fast visualization of the relationship between spatial features of cells and fluorescent signal intensities:
See the Cytokit Explorer docs for more details.
In addition to Cytokit Explorer, exports can also be generated using CellProfiler (CP) directly. This makes it possible to ammend a configuration with a line like this to generate both CP spreadhseets and a SQLite DB compatible with CellProfiler Analyst (see pub/config/codex-spleen/experiment.yaml):
analysis:
- cellprofiler_quantification:
- export_csv: true
- export_db: true
- export_db_objects_separately: trueThese screenshots from CellProfiler Analyst 2.2.1 show a reconstruction of plots used in the CODEX publication based on data generated by dynamic construction and execution of a CP 3.1.8 pipeline (see pub/analysis/codex-spleen/pipeline_execution.sh):
CellProfiler is not easy to use programmatically as it is used here. There is no official Python API and direct access to the internals has to be informed largely based on tests and other source code, but for any interested power-users, here are some parts of this project that may be useful resources:
- Installation: The Dockerfile shows how to bootstrap a minimal Python 2.7 environment compatible with CellProfiler 3.1.8
- Configuration: The cpcli.py script demonstrates how to build a CP pipeline programmatically (in this case segmented objects are provided to the pipeline that only does quantification and export)
- Analysis: When exported data from CP in a docker container, the paths in csv files or inserted into
a database will all be relative to a container. One simple solution to this problem is to simply create a
local
/lab/datafolder with copies of the information from the container that you would like to analyze.
A little more information on this can be found at pub/analysis/codex-spleen/README.md.
While the purpose of this pipeline is to perform image preprocessing and segmentation, the semantics of that segmentation often change. Depending on the experimental context, the provided cell nucleus segmentation may not be adequate and if a different segmentation methodology is required then any custom logic can be added to the pipeline as in the mc38-spheroid example. Specifically, a custom segmentation implementation is used here to identify spheroids rather than cells.
Errors in processor logs that can safely be ignored:
- tornado.iostream.StreamClosedError: Stream is closed: These often follow the completion of successful pipeline runs. These can hopefully be eliminated in the future with a dask upgrade but for now they can simply be ignored.
As a small piece of standalone functionality, instructions can be found here for how to run deconvolution on CODEX samples: Standalone Deconvolution Instructions
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent