CZB-zarr-challenge

Getting started

Installation

1. Clone and install locally

   git clone https://github.com/audreychun/czb-zarr-challenge.git
   cd czb-zarr-challenge
   pip install -e .

2. Install from PyPI

   pip install czb-zarr-challenge

Data Downloads

Download or place your OME-Zarr datasets (e.g., KazanskyStar_converted.zarr or 20241107_infection.zarr) into data/ before running any commands.

• KazanskyStar dataset: link

• 20241107 dataset: link

  • This one can be downloaded using wget -m -np -nH --cut-dirs=2 -R "index.html*" https://public.czbiohub.org/comp.micro/SWE_2025/20241107_infection.zarr/.

iohub Installation

To get the latest iohub (v0.3.0), clone and install manually:

git clone https://github.com/czbiohub-sf/iohub.git
pip install ./iohub

OME-TIFF → OME-Zarr Conversion

Use the provided tiff_to_zarr.py script. This leverages iohub’s TiffConverter to convert a Micromanager OME-TIFF to OME-Zarr. For Task 1, running the following command will convert Dataset 1 (Micromanager OME-TIFF layout) to OME-Zarr in the same /data folder.

python tiff_to_zarr.py

We will work with data/KazanskyStar_converted.zarr for Dataset 1 from now on.

Metadata Extraction

To generate a metadata .txt file with key information (shape, chunk size, dtype, scale, channel names) from a Zarr store:

czb-zarr-challenge metadata   --zarr_path /path/to/YourData.zarr   --output_path ./example_metadata.txt

Examples

czb-zarr-challenge metadata   --zarr_path data/KazanskyStar_converted.zarr   --output_path output/metadata_KazanskyStar.txt

czb-zarr-challenge metadata   --zarr_path data/20241107_infection.zarr   --output_path output/metadata_20241107.txt

Note: if you're having issues with your environment or PATH, you can always invoke the CLI entrypoint directly with Python:

python -m czb_zarr_challenge.cli metadata \
  --zarr_path data/KazanskyStar_converted.zarr \
  --output_path metadata_KazanskyStar.txt

Code for this metadata extraction is found in get_metadata.py. It uses iohub’s open_ome_zarr(...) to open the store and iohub’s node objects to navigate from Plate -> Position -> ImageArray. The script prints the full OME-Zarr hierarchy to stdout, and writes a human-readable .txt containing:

• Layout (Plate / Position / TiledPosition)
• Shape (e.g., (T, C, Z, Y, X))
• Chunk size (e.g., (1, 1, 10, 512, 512))
• Dtype (e.g., float32)
• Scale (e.g., (1.0, 0.5, 0.5))
• Channel names (e.g., ['Plate3D', 'nuclei_DAPI', 'virus_mCherry'])

Inference Profiling

The custom PyTorch DataLoader (in dataloader.py) uses iohub and OME-Zarr to read 5D volumes from Dataset 2. We profile both dataloading and inference times using a pretrained ResNet-18.

czb-zarr-challenge run_inference   --zarr_path data/20241107_infection.zarr

This will print a tqdm progress bar and then report:

Average batch loading time:   X.XXXX s
Average inference time:       Y.YYYY s

• The run_inference subcommand calls profile_inference(...) in inference.py.
• It takes the center Z slice of each 5D batch, resizes to 224×224, normalizes (using ImageNet stats), and runs a forward pass on ResNet-18.

Nuclei Segmentation

To segment nuclei using the DAPI channel and save labels back into the same OME-Zarr store:

czb-zarr-challenge segment   data/20241107_infection.zarr   --nuclei-channel nuclei_DAPI   --label-name nuclei_labels

• zarr_path – path to the OME-Zarr store
• --nuclei-channel (default: nuclei_DAPI)
• --label-name (default: nuclei_labels)

Main logic is found in segmentation.py. For each timepoint, it:

1. Reads a 3D volume from the DAPI channel.
2. Applies a global Otsu threshold → 3D blob detection (LoG) → 3D watershed.
3. Builds a (T, 1, Z, Y, X) label volume (dtype = uint16).
4. Writes it under /labels/<label_name>/0 in the same Zarr group, preserving chunking and compressor.

Parallelized Segmentation

You can also compare single-worker vs. parallel segmentation times:

czb-zarr-challenge segment_p   data/20241107_infection.zarr   --nuclei-channel nuclei_DAPI   --label-name nuclei_labels   --profile

• The --profile flag runs segmentation twice (once on a single worker, once on eight workers) and prints elapsed times for comparison.

Visualization

To visualize cell infection (with optional segmentation contours):

czb-zarr-challenge visualize   data/20241107_infection.zarr   output/

Examples

# plot nuclei overlay only, for timepoints 0, 5, and 10
czb-zarr-challenge visualize   data/20241107_infection.zarr   output/ --timepoints 0 5 10

# plot segmentation outlines
czb-zarr-challenge visualize   data/20241107_infection.zarr   output/

# plot timelapse of cell infection
czb-zarr-challenge visualize   data/20241107_infection.zarr   output/   --z-slice 8   --no-segmentation   --show-infection   --intensity-increase 3.0   --count-increase 60

• zarr_path – input OME-Zarr store
• output_dir – directory to save PNGs and GIFs
• --timepoints (〈int〉…) – specific timepoints to visualize (default = all)
• --z-slice 〈int〉 (default = middle slice)
• --no-segmentation (disable contour overlay)
• --show-infection (recolor nuclei in magenta when virus volume crosses threshold)
• --intensity-increase 〈float〉 (default = 3.0) – fold-change threshold for infection onset
• --count-increase 〈int〉 (default = 60) – absolute voxel count increase threshold for infection onset

• The visualize subcommand calls visualize_infection(...) in visualize.py.
• It can overlay segmentation contours and recolor infected nuclei once criteria are met.
• Infection criteria are based on in-nucleus infected voxel count, which are computed in infection_dynamics.py.

Dependencies & Environment

To reproduce all functionality:

1. Create a fresh virtual environment:

   python3 -m venv venv_repro
   source venv_repro/bin/activate
   pip install --upgrade pip

2. Install dependencies:

   pip install -r requirements.txt

---