Normalization Statistics Specification

Overview

This specification defines lightweight per-channel intensity normalization statistics for multichannel light microscopy datasets stored in OME-NGFF v0.5 format.

Motivation

Microscopy datasets often contain multiple channels with different intensity distributions, which requires intensity normalization for downstream analysis and visualization. For AI model training on such datasets, samples are commonly smaller random crops from a larger image. However, the intensity distribution of a random crop often does not match the overall intensity distribution of the full image. By pre-computing and storing summary statistics at both dataset and per-timepoint granularity, consumers can apply different types of data normalization efficiently without reading full images.

Location in Zarr Store

Normalization statistics are stored as part of the dca object directly within the zarr.json of the multiscales image group (i.e. the same zarr.json that contains ome.multiscales). The normalization_statistics field contains channel normalization objects with zero-based channel indices (as a string), and OPTIONAL metadata.

Example

{
  "ome": {
    "multiscales": [ "..." ]
  },
  "dca": {
    "version": "...",
    "url": "...",
    "channels": [ "..." ],
    "normalization_statistics": {
      "0": {
        "dataset_statistics": {
          "p1": 42.0,
          "p5": 62.5,
          "p95": 192.0,
          "p99": 218.5,
          "p95_p5": 129.5,
          "p99_p1": 176.5,
          "mean": 126.45,
          "std": 38.12,
          "median": 124.0,
          "iqr": 51.3
        },
        "timepoint_statistics": {
          "0": {
            "p1": 43.0,
            "p5": 63.1,
            "p95": 193.2,
            "p99": 219.0
          },
          "1": {
            "p1": 41.5,
            "p5": 61.8,
            "p95": 190.7,
            "p99": 216.3
          }
        }
      },
      "1": {
        "dataset_statistics": {
          "p1": 10.2,
          "p5": 18.7,
          "p95": 245.1,
          "p99": 251.8
        }
      },
      "metadata": {
        "grid": 16
      }
    }
  }
}

Channel Normalization Fields

See the JSON Schema (#/$defs/ChannelNormalization) for the authoritative definition.

Each channel normalization object MUST contain:

• dataset_statistics — intensity statistics computed over all spatial dimensions and timepoints for that channel.

Each channel normalization object MAY contain:

• timepoint_statistics — an object mapping zero-based timepoint indices (as strings) to per-timepoint intensity statistics. If present, it MUST contain entries for all timepoints in the dataset.

Intensity Statistics Fields

Required Fields

Both dataset_statistics and each value in timepoint_statistics MUST contain the following fields.

Field	Type	Description
p1	number	1st percentile of pixel intensities.
p5	number	5th percentile of pixel intensities.
p95	number	95th percentile of pixel intensities.
p99	number	99th percentile of pixel intensities.

Optional Fields

Field	Type	Description
p95_p5	number	Robust range: 95th percentile minus 5th percentile.
p99_p1	number	Wide robust range: 99th percentile minus 1st percentile.
mean	number	Arithmetic mean of pixel intensities.
std	number	Standard deviation of pixel intensities.
median	number	Median (50th percentile) of pixel intensities.
iqr	number	Interquartile range (75th percentile minus 25th percentile).

Normalization Metadata

The normalization_statistics object itself MAY contain:

• metadata — an optional object for recording parameters used when computing normalization statistics. Useful for capturing implementation details that may affect reproducibility, for example downsampling factor before computing statistics.