tl;dr: Working with large medical or scientific images for machine learning? -> Use MLArray.
MLArray is a purpose-built file format for N-dimensional medical and scientific array data in machine learning workflows. It replaces the usual patchwork of source formats and late-stage conversions to NumPy/Zarr/Blosc2 by layering standardized metadata on top of a Blosc2-backed storage layout, so the same files work reliably across training, analysis, and visualization tools (including Napari and MITK).
Installation
You can install mlarray via pip:
pip install mlarray
To enable the mlarray_convert CLI command (conversion with MedVol 1.0 or newer), install MLArray with the extra dependencies:
pip install mlarray[all]
Documentaion
See the documentation for the API reference, the metadata schema, usage examples or CLI usage.
Usage
Below are common usage patterns for loading, saving, and working with metadata.
Default usage
import numpy as np
from mlarray import MLArray
array = np.random.random((128, 256, 256))
image = MLArray(array) # Create MLArray image
image.save("sample.mla")
image = MLArray("sample.mla") # Loads image
Memory-mapped usage
from mlarray import MLArray
import numpy as np
# read-only, partial access (default)
image = MLArray.open("sample.mla", mmap_mode='r')
crop = image[10:20, 50:60] # Read crop
# read/write, partial access
image = MLArray.open("sample.mla", mmap_mode='r+')
image[10:20, 50:60] *= 5 # Modify crop in memory and disk
# read/write, partial access, create/overwrite
array = np.random.random((128, 256, 256))
image = MLArray.create("sample.mla", shape=array.shape, dtype=array.dtype, mmap_mode='w+')
image[...] = array # Modify image in memory and disk
Fast reading
MLArray.open reads lazily: image[...] decompresses only the chunks a patch touches. The first open in a process also turns on blosc2's read-ahead (4 threads), which fetches those chunks in parallel and makes reads of data not in the page cache several times faster, especially on network storage. Change it with mlarray.set_read_ahead(n) (0 turns it off) or the BLOSC_PREFETCH environment variable. Files are written with nnU-Net's layout and compression settings; see ML Optimization for details.
Metadata
Metadata is explicit and domain-neutral: what each axis is, where the array sits in physical space, what its values mean, optional domain profiles, and your own metadata. Geometry is never implied: spacing, origin and direction always refer to a declared space.
import numpy as np
from mlarray import MLArray
array = np.random.random((64, 128, 128))
image = MLArray(
array,
axes=["z", "y", "x"],
spacing=(1.5, 1.0, 1.0), # per array axis, in array order
origin=(10.0, 10.0, 30.0), # centre of index (0, 0, 0)
direction=[[0, 0, 1], [0, 1, 0], [1, 0, 0]], # column c: unit vector of array axis c
space="patient", # DICOM patient frame (LPS)
values={"kind": "measurement", "unit": "hounsfield_unit"},
user={"patient_id": "123"}, # any metadata of your own
)
print(image.spacing) # (1.5, 1.0, 1.0)
print(image.letter_codes()) # ('S', 'P', 'L')
image.set_geometry(spacing=(1.5, 5.3, 1.0))
image.user["study_id"] = "study-001"
image.save("with-metadata.mla")
# Open memory-mapped
image = MLArray.open("with-metadata.mla", mmap_mode='r+')
image.user["study_id"] = "new-study" # Modify metadata
image.close() # Close and save metadata, only necessary to save modified metadata
Segmentations, ML profiles and derived images:
import numpy as np
from mlarray import MLArray
base = MLArray("with-metadata.mla")
seg = MLArray(
np.zeros(base.shape, dtype=np.uint8),
copy=base, # geometry, values, profiles and user metadata of base (never its source)
values={"kind": "label", "labels": {"0": {"name": "background"}, "1": {"name": "liver"}}},
profiles={"ml": {"split": "train"}},
)
seg.save("seg.mla")
See Usage and Schema for the full metadata model.
Patch size variants
Default patch size (192):
from mlarray import MLArray
image = MLArray("sample.mla") # Existing file
image.save("default-patch.mla") # Keeps existing layout metadata
loaded = MLArray("sample.mla")
image = MLArray(loaded.to_numpy(), patch_size='default')
image.save("default-patch-relayout.mla") # Uses constructor patch_size='default' (192)
Custom isotropic patch size (512):
from mlarray import MLArray
loaded = MLArray("sample.mla")
image = MLArray(loaded.to_numpy(), patch_size=512)
image.save("patch-512.mla")
Custom non-isotropic patch size:
from mlarray import MLArray
loaded = MLArray("sample.mla")
image = MLArray(loaded.to_numpy(), patch_size=(128, 192, 256))
image.save("patch-non-iso.mla")
Manual chunk/block size:
from mlarray import MLArray
loaded = MLArray("sample.mla")
image = MLArray(
loaded.to_numpy(),
patch_size=None,
chunk_size=(1, 128, 128),
block_size=(1, 32, 32),
)
image.save("manual-chunk-block.mla")
Let Blosc2 itself configure chunk/block size:
from mlarray import MLArray
loaded = MLArray("sample.mla")
image = MLArray(loaded.to_numpy(), patch_size=None)
# If patch_size, chunk_size and block_size are all None, Blosc2 will auto-configure chunk and block size
image.save("blosc2-auto.mla")
CLI
mlarray_header
Print the metadata of a .mla file as JSON (--source and --user add the original metadata of
converted files and the user metadata).
mlarray_header sample.mla
mlarray_convert
Convert between MLArray and NIfTI, NRRD, MetaImage, MGH/MGZ and DICOM with
MedVol 1.0 or newer (pip install "mlarray[all]"). Nothing is reoriented, rescaled or
dropped silently: the array keeps its stored order and dtype, the geometry is taken from the file's own fields,
the original header is kept in source so the file can be written back faithfully, and every choice is reported.
mlarray_convert sample.nii.gz sample.mla
mlarray_convert sample.mla roundtrip.nii.gz
mlarray_convert dicom_dir/ ct.mla --dicom-tilt exact
mlarray_convert ct.mla out_dir/ --target-format dicom # re-emits the original DICOM series
Ambiguities are resolved by explicit MedVol policies, e.g. --nifti-form sform when a NIfTI's qform and sform
disagree, or --world-frame ras to store the patient space in RAS instead of DICOM's LPS.
Contributing
Contributions are welcome! Please open a pull request with clear changes and add tests when appropriate.
Acknowledgments
This repository is developed and maintained by the Applied Computer Vision Lab (ACVL) of Helmholtz Imaging and the Division of Medical Image Computing at DKFZ.
Metadata
Release files for mlarray 1.0.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| mlarray-1.0.0.tar.gz | 670.5 kB | With provenance | Details for mlarray-1.0.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| mlarray-1.0.0-py3-none-any.whl | Python 3 | none | any | Details for mlarray-1.0.0-py3-none-any.whl |
Total release size: 725.7 kB
Release files / mlarray-1.0.0.tar.gz
| Download URL | mlarray-1.0.0.tar.gz |
|---|---|
| Size | 670.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
419fbf711bdb2e33c6c079d26dbeed28eace1fac34cb00ffc09062ea529b9f90
|
|
BLAKE2b-256 checksum How to use checksums |
2253d10f9d18adddf8cbf962e04d2f0557c96ebecab300472bd00f5cc50fc907
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 6, 2026.
Transparency logRelease files / mlarray-1.0.0-py3-none-any.whl
| Download URL | mlarray-1.0.0-py3-none-any.whl |
|---|---|
| Size | 55.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
17ce59f1cf185f6cc125eddc883f1fb28e014161a2bbfb191b422dbdd0adca45
|
|
BLAKE2b-256 checksum How to use checksums |
5af82ae96849ecc9b032b46fa8266a91d46bd50bc06a868e6286752852b7777c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 6, 2026.
Transparency logSecurity for mlarray 1.0.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for mlarray-1.0.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 6, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.