magdi-data 0.5.0a218

MAGDI data

MAGDI Data

This python package named magdi_data is responsible for data loading to be used for AI training. It also contains the definition of a specific dataset structure.

Setup

• pip install -r requirements.txt to run AI experiments
• pip install -r requirements-tests.txt to run pytest
• Copy .env-example as .env file and set your credentials and other env variables there

dataset_meta.json

The dataset_meta.json describes a dataset in the Occurrence Instance Format. Its properties have to follow a strict format to be readable by humans and machines.

How to create a dataset

Read README_create_occ_inst_dataset.md

Terms

Occurrence

• An occurrence is a real scanned object. It can have one or more instances (replicates or measurements).

Instance

• An instance is a digitalized object.

Contents

property	description	json type	example
id	Identifier of the Dataset. Is to be used as folder name. Normalized string costing of "name", "release" and "uuid". Is automatically overwritten by each validation by pydantic model.	string	Maize
uuid	Generated UUID string for this dataset.
name	Name of the dataset.
short_name	Abbreviation of the "name".
release	Version of release.
descriptive_metadata	Nested object that holds further optional fields to describe the dataset.
entity	Type of real-world object the dataset pertains.	string	maize
string	2025.07.02
description	Short description of the dataset.	string	NMR dataset with all maize samples. Test set dispersed among sets 7,8, 10, 11, tree set7 samples discarded.
reference	Description or URL where the data was published.	string	Leibniz Institute of Plant Genetics and Crop Plant Research (IPK)
license	Name which license is used for this dataset.	string	IPK proprietary
tags	List of modalities to describe this dataset. On Hugging Face an overview of exsisting modalities can be found: https://huggingface.co/datasets	array of string	["MRI","3D","Image"]
task_categories	List of Machine Learning tasks for which the dataset is intended. Hugging Face gives an overview of possible tasks: https://huggingface.co/tasks	array of string	["image-segmentation"]
labels	OPTIONAL (Needed if labeled data exists). Class label mapping made of key value pairs. Keys must be ascending positive integers starting from zero and are formatted as strings. They cannot have gaps. If a key has not a value, the value must be set as empty string e.g. "1": "".	object	{ "0": "background", "1": "embryo" "2":"endosperm", "3":"aleuron"}
image_format	File ending of the image files.	string	.nii.gz
annotation_format	OPTIONAL (Needed if labeled data exists). File ending of the annotation files.	string	.nii.gz
num_instances_total	Total number of instances.	integer	4
num_instances_training	Number of instances of the training split.	integer	2
num_instances_test	OPTIONAL (Needed if test split exists). Number of instances of the test split.	integer	2
num_instances_validation	OPTIONAL (Needed if validation split exists). Number of instances of the validation split.	integer	2
num_occurrences_total	Total number of occurrences.	integer	4
num_occurrences_training	Number of occurrences of the training split.	integer	2
num_occurrences_test	OPTIONAL (Needed if test split exists). Number of occurrences of the test split.	integer	2
num_occurrences_validation	OPTIONAL (Needed if validation split exists). Number of occurrences of the validation split.	integer	2
image_data_type	Defines the type of data compatible to numpy (np.dtype) how the image information is stored.	string	int16
annotation_data_type	OPTIONAL (Needed if labeled data exists). Defines the type of data compatible to numpy (np.dtype) how the annotation information is stored.	string	uint8
value_channels	Number of channels how the image information is stored, e.g. 1 for MRI, 3 for RGB images.	integer	1
range	Possible value range of voxels across all instances for the entity. This field is set manually.
range_max	Maximum value range across all instances. This field is calculated during validation by pydantic model.
range_min	Minimum value range across all instances. This field is calculated during validation by pydantic model.
range_avg	Average value range across all instances. This field is calculated during validation by pydantic model.
dimensions_max	Dimensions of the largest image.	array of integer	[96, 128, 150]
dimensions_min	Dimensions of the smallest image. Must have the same shape as dimensionsMax.	array of integer	[41, 64, 78]
dimensions_avg	Average of the dimensions over all images. Must have the same shape as dimensionsMax.	array of number	[57.49618320610687, 92.01526717557252, 115.6030534351145]
resolution_unit	Unit for resolution_voxel_size. Voxel size: relative to the real world for each image dimension	string	mm
resolution_voxel_size	OPTIONAL Size of a voxel for each instance. This field is set manually.	array of number	[0.1,0.1,0.1]
resolution_voxel_size_max	Maximum voxel size across all instances. This field is calculated during validation by pydantic model.
resolution_voxel_size_min	Minimum voxel size across all instances. This field is calculated during validation by pydantic model.
resolution_voxel_size_avg	Average voxel size across all instances. This field is calculated during validation by pydantic model.
training	Training split. Contains one json object for each instance of the training split. Each json object holds the path to the image file. For annotated data the path to the annotation file is set. "image" key is mandatory. "annotations" key is only mandatory for annotated data.	array of object	[ { "annotations": "occurrence-0000/instance-0000/annotations.nii.gz", "image": "occurrence-0001/instance-0000/image.nii.gz" } ]
test	OPTIONAL (Needed if test split exists). Test split. Contains one json object for each instance of the test split. Each json object holds the path to the image file. For annotated data the path to the annotation file is set. "image" key is mandatory. "annotations" key is only needed for annotated data.	array of object	[ { "annotations": "occurrence-0004/instance-0000/annotations.nii.gz", "image": "occurrence-0005/instance-0000/image.nii.gz" } ]
validation	OPTIONAL (Needed if validation split exists). Validation split. Contains one json object for each instance of the validation split. Each json object holds the path to the image file. For annotated data the path to the annotation file is set. "image" key is mandatory. "annotations" key is only needed for annotated data.	array of object	[ { "annotations": "occurrence-0002/instance-0000/annotations.nii.gz", "image": "occurrence-0003/instance-0000/image.nii.gz" } ]
data_references	Holds key-value pairs to map directory or files of this dataset to a reference. It is not used for AI training. It is necessary to associate digital instances with its origins of creation. E.g. Occurrence is mapped to the ID of a real world object.	object	{ "occurrence-0000/instance-0000/annotations.nii.gz": "example1", "occurrence-0001/instance-0000/annotations.nii.gz": "example3" }

Example:

{
  "id": "mais_karyopse_2025-07-02_806fd055-b90d-46d5-a1ce-dd7e61b286ee",
  "uuid": "806fd055-b90d-46d5-a1ce-dd7e61b286ee",
  "name": "Mais Karyopse",
  "short_name": "MK2025",
  "entity": "maize",
  "descriptive_metadata": {
    "latin_name": "Frumentum",
    "line_name": "Mais ",
    "structure": "seed",
    "dap": "10 DAP",
    "device": "NMR Device Name",
    "coil": "CRP 13C/1H 5mm 400MHz",
    "measurement_channel": "structure"
  },
  "release": "2021.7.5",
  "description": "Short description",
  "reference": "Reference to teh author",
  "license": "license",
  "tags": [
    "MRI",
    "3D",
    "Image"
  ],
  "task_categories": [
    "image-segmentation"
  ],
  "labels": {
    "0": "background",
    "1": "embryo",
    "2": "endosperm",
    "3": "aleuron"
  },
  "image_format": ".nii.gz",
  "annotation_format": ".nii.gz",
  "num_instances_total": 6,
  "num_instances_training": 2,
  "num_instances_test": 2,
  "num_instances_validation": 2,
  "num_occurrences_total": 6,
  "num_occurrences_training": 2,
  "num_occurrences_test": 2,
  "num_occurrences_validation": 2,
  "image_data_type": "int16",
  "annotation_data_type": "uint8",
  "value_channels": 1,
  "range": [
    0,
    16000
  ],
  "range_max": 1002,
  "range_min": 0,
  "range_avg": 167.57250248655913,
  "dimensions_max": [
    96,
    128,
    150
  ],
  "dimensions_min": [
    41,
    64,
    78
  ],
  "dimensions_avg": [
    57.49618320610687,
    92.01526717557252,
    115.6030534351145
  ],
  "resolution_unit": "mm",
  "resolution_voxel_size": [
    0.1,
    0.1,
    0.1
  ],
  "resolution_voxel_size_max": [
    0.1,
    0.1,
    0.1
  ],
  "resolution_voxel_size_min": [
    0.1,
    0.1,
    0.1
  ],
  "resolution_voxel_size_avg": [
    0.1,
    0.1,
    0.1
  ],
  "training": [
    {
      "annotations": "occurrence-0000/instance-0000/annotations.nii.gz",
      "image": "occurrence-0000/instance-0000/image.nii.gz"
    },
    {
      "annotations": "occurrence-0001/instance-0000/annotations.nii.gz",
      "image": "occurrence-0001/instance-0000/image.nii.gz"
    }
  ],
  "validation": [
    {
      "annotations": "occurrence-0002/instance-0000/annotations.nii.gz",
      "image": "occurrence-0002/instance-0000/image.nii.gz"
    },
    {
      "annotations": "occurrence-0003/instance-0000/annotations.nii.gz",
      "image": "occurrence-0003/instance-0000/image.nii.gz"
    }
  ],
  "test": [
    {
      "annotations": "occurrence-0004/instance-0000/annotations.nii.gz",
      "image": "occurrence-0004/instance-0000/image.nii.gz"
    },
    {
      "annotations": "occurrence-0005/instance-0000/annotations.nii.gz",
      "image": "occurrence-0005/instance-0000/image.nii.gz"
    }
  ],
  "data_references": {
    "occurrence-0000/instance-0000/image.nii.gz": "example1",
    "occurrence-0001/instance-0000/image.nii.gz": "example2",
    "occurrence-0002/instance-0000/image.nii.gz": "example3",
    "occurrence-0003/instance-0000/image.nii.gz": "folder_xyz/exampleA",
    "occurrence-0004/instance-0000/image.nii.gz": "folder_xyz/exampleB",
    "occurrence-0005/instance-0000/image.nii.gz": "folder_xyz/exampleC"
  }
}