Next-generation file formats (NGFF)

1. Introduction

Bioimaging science is at a crossroads. Currently, the drive to acquire more, larger, preciser spatial measurements is unfortunately at odds with our ability to structure and share those measurements with others. During a global pandemic more than ever, we believe fervently that global, collaborative discovery as opposed to the post-publication, "data-on-request" mode of operation is the path forward. Bioimaging data should be shareable via open and commercial cloud resources without the need to download entire datasets.

At the moment, that is not the norm. The plethora of data formats produced by imaging systems are ill-suited to remote sharing. Individual scientists typically lack the infrastructure they need to host these data themselves. When they acquire images from elsewhere, time-consuming translations and data cleaning are needed to interpret findings. Those same costs are multiplied when gathering data into online repositories where curator time can be the limiting factor before publication is possible. Without a common effort, each lab or resource is left building the tools they need and maintaining that infrastructure often without dedicated funding.

This document defines a specification for bioimaging data to make it possible to enable the conversion of proprietary formats into a common, cloud-ready one. Such next-generation file formats layout data so that individual portions, or "chunks", of large data are reference-able eliminating the need to download entire datasets.

1.1. Why "NGFF"?

A short description of what is needed for an imaging format is "a hierarchy of n-dimensional (dense) arrays with metadata". This combination of features is certainly provided by HDF5 from the HDF Group, which a number of bioimaging formats do use. HDF5 and other larger binary structures, however, are ill-suited for storage in the cloud where accessing individual chunks of data by name rather than seeking through a large file is at the heart of parallelization.

As a result, a number of formats have been developed more recently which provide the basic data structure of an HDF5 file, but do so in a more cloud-friendly way. In the PyData community, the Zarr [zarr] format was developed for easily storing collections of NumPy arrays. In the ImageJ community, N5 [n5] was developed to work around the limitations of HDF5 ("N5" was originally short for "Not-HDF5"). Both of these formats permit storing individual chunks of data either locally in separate files or in cloud-based object stores as separate keys.

A current effort is underway to unify the two similar specifications to provide a single binary specification. The editor’s draft will soon be entering a request for comments (RFC) phase with the goal of having a first version early in 2021. As that process comes to an end, this document will be updated.

1.2. OME-NGFF

The conventions and specifications defined in this document are designed to enable next-generation file formats to represent the same bioimaging data that can be represented in OME-TIFF and beyond. However, the conventions will also be usable by HDF5 and other sufficiently advanced binary containers. Eventually, we hope, the moniker "next-generation" will no longer be applicable, and this will simply be the most efficient, common, and useful representation of bioimaging data, whether during acquisition or sharing in the cloud.

Note: The following text makes use of OME-Zarr [ome-zarr-py], the current prototype implementation, for all examples.

2. On-disk (or in-cloud) layout

An overview of the layout of an OME-Zarr fileset should make understanding the following metadata sections easier. The hierarchy is represented here as it would appear locally but could equally be stored on a web server to be accessed via HTTP or in object storage like S3 or GCS.

2.1. Images

The following layout describes the expected Zarr hierarchy for images with multiple levels of resolutions and optionally associated labels.

.                             # Root folder, potentially in S3,
│                             # with a flat list of images by image ID.
│
├── 123.zarr                  # One image (id=123) converted to Zarr.
│
└── 456.zarr                  # Another image (id=456) converted to Zarr.
    │
    ├── .zgroup               # Each image is a Zarr group, or a folder, of other groups and arrays.
    ├── .zattrs               # Group level attributes are stored in the .zattrs file and include
    │                         #  "multiscales" and "omero" below)
    │
    ├── 0                     # Each multiscale level is stored as a separate Zarr array,
    │   ...                   # which is a folder containing chunk files which compose the array.
    ├── n                     # The name of the array is arbitrary with the ordering defined by
    │   │                     # by the "multiscales" metadata, but is often a sequence starting at 0.
    │   │
    │   ├── .zarray           # All image arrays are 5-dimensional
    │   │                     # with dimension order (t, c, z, y, x).
    │   │
    │   └─ t                  # Chunks are stored with the nested directory layout.
    │      └─ c               # All but the last chunk element are stored as directories.
    │         └─ z            # The terminal chunk is a file. Together the directory and file names
    │            └─ y         # provide the "chunk coordinate" (t, c, z, y, x), where the maximum coordinate
    │               └─ x      # will be `dimension_size / chunk_size`.
    │
    └── labels
        │
        ├── .zgroup           # The labels group is a container which holds a list of labels to make the objects easily discoverable
        │
        ├── .zattrs           # All labels will be listed in `.zattrs` e.g. `{ "labels": [ "original/0" ] }`
        │                     # Each dimension of the label `(t, c, z, y, x)` should be either the same as the
        │                     # corresponding dimension of the image, or `1` if that dimension of the label
        │                     # is irrelevant.
        │
        └── original          # Intermediate folders are permitted but not necessary and currently contain no extra metadata.
            │
            └── 0             # Multiscale, labeled image. The name is unimportant but is registered in the "labels" group above.
                ├── .zgroup   # Zarr Group which is both a multiscaled image as well as a labeled image.
                ├── .zattrs   # Metadata of the related image and as well as display information under the "image-label" key.
                │
                ├── 0         # Each multiscale level is stored as a separate Zarr array, as above, but only integer values
                │   ...       # are supported.
                └── n

2.2. High-content screening

The following specification defines the hierarchy for a high-content screening dataset. Three groups must be defined above the images:

• the group above the images defines the well and MUST implement the well specification. All images contained in a well are fields of view of the same well

• the group above the well defines a row of wells

• the group above the well row defines an entire plate i.e. a two-dimensional collection of wells organized in rows and columns. It MUST implement the plate specification

.                             # Root folder, potentially in S3,
│
└── 5966.zarr                 # One plate (id=5966) converted to Zarr
    ├── .zgroup
    ├── .zattrs               # Implements "plate" specification
    ├── A                     # First row of the plate
    │   ├── .zgroup
    │   │
    │   ├── 1                 # First column of row A
    │   │   ├── .zgroup
    │   │   ├── .zattrs       # Implements "well" specification
    │   │   │
    │   │   ├── 0             # First field of view of well A1
    │   │   │   │
    │   │   │   ├── .zgroup
    │   │   │   ├── .zattrs   # Implements "multiscales", "omero"
    │   │   │   ├── 0
    │   │   │   │   ...       # Resolution levels
    │   │   │   ├── n
    │   │   │   └── labels    # Labels (optional)
    │   │   ├── ...           # Fields of view
    │   │   └── m
    │   ├── ...               # Columns
    │   └── 12
    ├── ...                   # Rows
    └── H

3. Metadata

The various .zattrs files throughout the above array hierarchy may contain metadata keys as specified below for discovering certain types of data, especially images.

3.1. "multiscales" metadata

Metadata about the multiple resolution representations of the image can be found under the "multiscales" key in the group-level metadata.

"multiscales" contains a list of dictionaries where each entry describes a multiscale image.

Each dictionary contained in the list MUST contain the field "datasets", which is a list of dictionaries describing the arrays storing the individual resolution levels. Each dictionary in "datasets" MUST contain the field "path", whose value contains the path to the array for this resolution relative to the current zarr group. The "path"s MUST be ordered from largest (i.e. highest resolution) to smallest.

It SHOULD contain the field "name".

It SHOULD contain the field "version", which indicates the version of the multiscale metadata of this image (current version is 0.2).

It SHOULD contain the field "type", which gives the type of downscaling method used to generate the multiscale image pyramid.

It SHOULD contain the field "metadata", which contains a dictionary with additional information about the downscaling method.

{
    "multiscales": [
        {
            "version": "0.2",
            "name": "example",
            "datasets": [
                {"path": "0"},
                {"path": "1"},
                {"path": "2"}
            ],
            "type": "gaussian",
            "metadata": {                                       # the fields in metadata depend on the downscaling implementation
                "method": "skimage.transform.pyramid_gaussian", # here, the paramters passed to the skimage function are given
                "version": "0.16.1",
                "args": "[true]",
                "kwargs": {"multichannel": true}
            }
        }
    ]
}

If only one multiscale is provided, use it. Otherwise, the user can choose by name, using the first multiscale as a fallback:

datasets = []
for named in multiscales:
    if named["name"] == "3D":
        datasets = [x["path"] for x in named["datasets"]]
        break
if not datasets:
    # Use the first by default. Or perhaps choose based on chunk size.
    datasets = [x["path"] for x in multiscales[0]["datasets"]]

3.2. "omero" metadata

Information specific to the channels of an image and how to render it can be found under the "omero" key in the group-level metadata:

"id": 1,                              # ID in OMERO
"name": "example.tif",                # Name as shown in the UI
"version": "0.2",                     # Current version
"channels": [                         # Array matching the c dimension size
    {
        "active": true,
        "coefficient": 1,
        "color": "0000FF",
        "family": "linear",
        "inverted": false,
        "label": "LaminB1",
        "window": {
            "end": 1500,
            "max": 65535,
            "min": 0,
            "start": 0
        }
    }
],
"rdefs": {
    "defaultT": 0,                    # First timepoint to show the user
    "defaultZ": 118,                  # First Z section to show the user
    "model": "color"                  # "color" or "greyscale"
}

See https://docs.openmicroscopy.org/omero/5.6.1/developers/Web/WebGateway.html#imgdata for more information.

3.3. "labels" metadata

The special group "labels" found under an image Zarr contains the key labels containing the paths to label objects which can be found underneath the group:

{
  "labels": [
    "orphaned/0"
  ]
}

Unlisted groups MAY be labels.

3.4. "image-label" metadata

Groups containing the image-label dictionary represent an image segmentation in which each unique pixel value represents a separate segmented object. image-label groups MUST also contain multiscales metadata and the two "datasets" series MUST have the same number of entries.

The colors key defines a list of JSON objects describing the unique label values. Each entry in the list MUST contain the key "label-value" with the pixel value for that label. Additionally, the "rgba" key MAY be present, the value for which is an RGBA unsigned-int 4-tuple: [uint8, uint8, uint8, uint8] All label-values must be unique. Clients who choose to not throw an error should ignore all except the _last_ entry.

Some implementations may represent overlapping labels by using a specially assigned value, for example the highest integer available in the pixel range.

The properties key defines a list of JSON objects which also describes the unique label values. Each entry in the list MUST contain the key "label-value" with the pixel value for that label. Additionally, an arbitrary number of key-value pairs MAY be present for each label value denoting associated metadata. Not all label values must share the same key-value pairs within the properties list.

The source key is an optional dictionary which contains information on the image the label is associated with. If included it MAY include a key image whose value is the relative path to a Zarr image group. The default value is "../../" since most labels are stored under a subgroup named "labels/" (see above).

"image-label":
  {
    "version": "0.2",
    "colors": [
      {
        "label-value": 1,
        "rgba": [255, 255, 255, 0]
      },
      {
        "label-value": 4,
        "rgba": [0, 255, 255, 128]
      },
      ...
      ],
    "properties": [
      {
        "label-value": 1,
        "area (pixels)": 1200,
        "class": "foo"

      },
      {
        "label-value": 4,
        "area (pixels)": 1650
      },
      ...
      ]
  },
  "source": {
    "image": "../../"
  }
]

3.5. "plate" metadata

For high-content screening datasets, the plate layout can be found under the custom attributes of the plate group under the plate key.

acquisitions

An optional list of JSON objects defining the acquisitions for a given plate. Each acquisition object MUST contain an id key providing an unique identifier within the context of the plate to which fields of view can refer to. It SHOULD contain a name key identifying the name of the acquisition. It SHOULD contain a maximumfieldcount key indicating the maximum number of fields of view for the acquisition. It MAY contain a description key providing a description for the acquisition. It MAY contain a startime and/or endtime key specifying the start and/or end timestamp of the acquisition using an epoch string.

columns

A list of JSON objects defining the columns of the plate. Each column object defines the properties of the column at the index of the object in the list. If not empty, it MUST contain a name key specifying the column name.

field_count

An integer defining the maximum number of fields per view across all wells.

name

A string defining the name of the plate.

rows

A list of JSON objects defining the rows of the plate. Each row object defines the properties of the row at the index of the object in the list. If not empty, it MUST contain a name key specifying the row name.

version

A string defining the version of the specification.

wells

A list of JSON objects defining the wells of the plate. Each well object MUST contain a path key identifying the path to the well subgroup.

For example the following JSON object defines a plate with two acquisition and 6 wells (2 rows and 3 columns), containing up 2 fields of view per acquistion.

"plate": {
    "acquisitions": [
        {
            "id": 1,
            "maximumfieldcount": 2,
            "name": "Meas_01(2012-07-31_10-41-12)",
            "starttime": 1343731272000
        },
        {
            "id": 2,
            "maximumfieldcount": 2,
            "name": "Meas_02(201207-31_11-56-41)",
            "starttime": 1343735801000
        }
    ],
    "columns": [
        {
            "name": "1"
        },
        {
            "name": "2"
        },
        {
            "name": "3"
        }
    ],
    "field_count": 4,
    "name": "test",
    "rows": [
        {
            "name": "A"
        },
        {
            "name": "B"
        }
    ],
    "version": "0.2",
    "wells": [
        {
            "path": "2020-10-10/A/1"
        },
        {
            "path": "2020-10-10/A/2"
        },
        {
            "path": "2020-10-10/A/3"
        },
        {
            "path": "2020-10-10/B/1"
        },
        {
            "path": "2020-10-10/B/2"
        },
        {
            "path": "2020-10-10/B/3"
        }
    ]
  }

3.6. "well" metadata

For high-content screening datasets, the metadata about all fields of views under a given well can be found under the "well" key in the attributes of the well group.

images

A list of JSON objects defining the fields of views for a given well. Each object MUST contain a path key identifying the path to the field of view. If multiple acquisitions were performed in the plate, it SHOULD contain an acquisition key identifying the id of the acquisition which must match one of acquisition JSON objects defined in the plate metadata.

version

A string defining the version of the specification.

For example the following JSON object defines a well with four fields of views. The first two fields of view were part of the first acquisition while the last two fields of view were part of the second acquisition.

"well": {
    "images": [
        {
            "acquisition": 1,
            "path": "0"
        },
        {
            "acquisition": 1,
            "path": "1"
        },
        {
            "acquisition": 2,
            "path": "2"
        },
        {
            "acquisition": 2,
            "path": "3"
        }
    ],
    "version": "0.2"
  }

4. Implementations

Projects which support reading and/or writing OME-NGFF data include:

bigdataviewer-ome-zarr

Fiji-plugin for reading OME-Zarr.

bioformats2raw

A performant, Bio-Formats image file format converter.

omero-ms-zarr

A microservice for OMERO.server that converts images stored in OMERO to OME-Zarr files on the fly, served via a web API.

idr-zarr-tools

A full workflow demonstrating the conversion of IDR images to OME-Zarr images on S3.

OMERO CLI Zarr plugin

An OMERO CLI plugin that converts images stored in OMERO.server into a local Zarr file.

ome-zarr-py

A napari plugin for reading ome-zarr files.

vizarr

A minimal, purely client-side program for viewing Zarr-based images with Viv & ImJoy.

All implementations prevent an equivalent representation of a dataset which can be downloaded or uploaded freely. An interactive version of this diagram is available from the OME2020 Workshop. Mouseover the blackboxes representing the implementations above to get a quick tip on how to use them.

Note: If you would like to see your project listed, please open an issue or PR on the ome/ngff repository.

5. Citing

Next-generation file format (NGFF) specifications for storing bioimaging data in the cloud. J. Moore, et al. Editors. Open Microscopy Environment Consortium, 29 March, 2021. This edition of the specification is https://ngff.openmicroscopy.org/0.2/. The latest edition is available at https://ngff.openmicroscopy.org/latest/. (doi:10.5281/zenodo.4282107)

6. Version History