Reference OME-Zarr for 3D time-lapse light-sheet microscopy with nuclei tracking

This Zenodo publication provides an example dataset to serve as a template for discussing specification requirements in OME-Zarr and related formats for light-sheet microscopy data. It contains a single movie of a mouse intestinal organoid from the publication “Multiscale light-sheet organoid imaging framework”. The dataset includes multiple image restoration steps (raw and deconvolved), maximum intensity projections along z, multiple label images containing instance segmentations of objects (nucleus, cell and lumen), and a lineage tree resulting from nuclear tracking. Moreover, it contains additional metadata, such as per-frame POSIX timestamps, experimental metadata, and normalization statistics. Together, these components represent a live imaging dataset at an advanced processing stage and can serve as an example of a collection of (extension) components as proposed in OME-Zarr RFC8.

Utilized Specifications

Wherever possible, we opted for existing specifications to represent the dataset components and used the latest stable release at the time of publication. These include:

• OME-Zarr v0.5: images, label images, MIPs

• GEFF v1.2: lineage tree and nuclear features

• DCA v0.2.8: experimental metadata, channel metadata, normalization statistics

• ngio generic table v1 (Parquet backend): POSIX timestamps

Dataset Components

Image Data

Timelapse movies are stored as 5D multiscale OME-Zarr v0.5 arrays (t, c, z, y, x) with 5 levels of detail. The dataset contains 2 channels: a membrane marker (mem9) and a nuclear marker (H2B). Due to high anisotropy at the highest resolution level (z: 2.0µm, y/x: 0.26µm), images were downsampled by factors of 2 in x and y only. Each restoration step is stored as a separate OME-Zarr, with both sharing a physical coordinate system (raw.ome.zarr and deconv.ome.zarr).

Label Images

Label images are stored as 4D multiscale arrays (t, z, y, x) with a downsampling schedule matching intensity images. In accordance with the OME-Zarr v0.5 specification, they have been placed under the labels group of deconv.ome.zarr, e.g., deconv.ome.zarr/labels/nucleus. Label IDs for nuclei were generated for each timepoint independently and do NOT represent object identity across timepoints (see Lineage Tree). Label IDs for nuclei and cells cannot be expected to match. Nuclei and cells that are not part of the tracking solution are not included in the label images, leading to sparse segmentation towards later timepoints (~350 in the full dataset). The label image deconv.ome.zarr/labels/lumen contains a semantic segmentation of lumen (label_id = 1) and epithelium (label_id = 2).

Maximum Intensity Projections

Maximum intensity projections along z are stored as 5D multiscale OME-Zarr v0.5 arrays with a singleton z dimension. Non-projected dimensions match their corresponding dimensions in physical extent and spatial sampling of the volumes. Both restoration steps were projected (raw_mip-z.ome.zarr and deconv_mip-z.ome.zarr). Label images were not projected.

Lineage Tree

A lineage tree of tracked nuclei is stored as a Graph Exchange File Format (GEFF) v1.2 group under deconv.ome.zarr/tracks/nucleus.geff. The graph contains the following node properties:

• t_idx: index into the t dimension of the arrays

• z, y, x: centroid position in physical coordinates

• t: time elapsed since the first frame in seconds

• label_id: corresponding object in `labels/nucleus`

• track_id: object identities over time (i.e., tracklets or tracks)

• nucleus_*: features relating to nuclear morphology

• cell_*: features relating to cellular morphology

• dendrogram_*: y-coordinates for dendrogram visualization

• is_root_node: true for root nodes with in-degree = 0

• is_leaf_node: true for leaf nodes with out-degree = 0

• is_split_node: true for nodes with out-degree > 1

• is_merge_node: true for nodes with in-degree > 1

And edge properties:

• displacement: Euclidean distance between source and target node

• edge_type: one of TRACK, SPLIT, MERGE

GEFF axes use a time-unit of frame to match the OME-Zarr metadata of the arrays, which omit the time unit and assume a scale of 1 to allow the representation of non-uniform temporal sampling.

Timestamps for Non-Uniform Temporal Sampling

To accommodate delays at the microscope (e.g., interruptions during buffer exchange) or experimentally preferred variable temporal sampling, we include absolute timestamps for each frame as a POSIX timestamp float. Timestamps are stored as an ngio generic table v1 written using the Parquet backend. The table contains two columns:

• t_idx: index into the t dimension of the arrays

• posix_timestamp: seconds (+ fractional sub-seconds) since 00:00:00 01.01.1970 UTC

DCA Metadata

Additional metadata not covered by other specifications are written according to the Dynamic Cell Atlas (DCA) v0.2.8 specifications. They are written to each multiscale OME-Zarr at the top-level zarr.json under the “attributes:dca” key as a sibling to the “attributes:ome” key. The top-level metadata keys are:

• channels: additional fields adding to ome:omero:channels

• acquisition: microscope type and imaging conditions, includes a copy of the posix_timestamp column under acquisition.timestamps

• biosample: ontological description of the biological sample

• specimen: metadata relating to sample preparation

• study: metadata relating to the lab of origin, experimenter, and related study

• study_component: ontological description of the experiment

• image_data: processing steps conducted

• normalization_statistics: per-channel global and per-channel-per-timepoint intensity quantiles

DCA metadata were validated using the validator tool with the v0.2.8 tagged version of the repository.

Subsampled Datasets

The datasets are all based on one movie (001-full), strategically downsampled in time to improve accessibility and address different use cases. Each dataset is packaged as a separate zip file.

The datasets come in three sizes:

1. full: 667 frames, ~40GB

2. small: ~50 frames, ~2.5GB

3. mini: 5 frames, ~0.25GB

Downsampled datasets are subsampled:

1. as contiguous time slices (small, mini)

2. with reduced temporal resolution (small-lowT, mini-lowT)

3. simulating variable acquisition rate (small-varT, mini-varT)

All datasets contain at least one division and one merger event. Lineage trees of datasets with non-contiguously sampled timepoints were updated to reflect the lineage relationships of the full dataset, but might contain nodes with more than two children (see dataset_lineage_trees.png, mini-lowT).