Open-source project
Project-MONAI/MONAILabel avatar
Project-MONAI/MONAILabel

MONAI Label trains the model, the viewer draws the mask

MONAI Label is an intelligent open source image labeling and learning tool.

912 stars280 forksPythonApache-2.0

At a glance

What is it?
MONAI Label is a Python server that trains and serves annotation models on medical images. Its viewers are separate projects, its release tags and dates disagree, and its Python floor comes from one dependency marker.
Who is it for?
Adopt MONAI Label when the annotation work is clinical imaging, when 3DSlicer, OHIF, QuPath, Digital Slide Archive or CVAT is already in your loop, and when a pinned dependency set is preferable to a moving one. It is the wrong shape for a general image labelling project and for anyone who wants a viewer inside the same install.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 17 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • Project-MONAI/MONAILabel912 ★ · +24
Star history
Repository2026-09-122026-11-15
Project-MONAI/MONAILabel888912

24 Stars in the last 64 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

A server for the models, a separate app for the drawing

MONAI Label is a Python package that runs a service, not an application you draw in. The project describes its own shape as serverless in the sense that custom labeling apps are exposed as a service through the MONAI Label Server, and it states plainly that server and client may sit on the same machine or on different ones.

The split matters when you budget an installation. The repository root separates `monailabel/` for the server side, `sample-apps/` for bundled tasks and `plugins/` for viewer integrations. The feature list names 3DSlicer and OHIF for radiology, QuPath, Digital Slide Archive and CVAT for pathology, and CVAT again for endoscopy. Each of those is a separate application with its own release cycle.

What you install from this repository trains and serves models. The component that turns a model's output into a filled contour lives in someone else's project. Anyone scoping an annotation effort has to count both sides.

The Python floor arrives as a SAM-2 marker

The interpreter requirement is expressed as a marker expression rather than a sentence. The SAM-2 dependency is pinned to one commit of another project and carries its own version gate:

code
# SAM-2 @ git+https://github.com/facebookresearch/sam2.git@c2ec8e14a185632b0a5d8b161928ceb50197eddc ; python_version >= '3.10'

That gate means SAM 2 resolves only on 3.10 and newer. Nothing else in the file imposes a Python version, so removing that one line removes the declared floor with it. The commented alternative directly underneath, `#sam2>=0.4.1; python_version >= '3.10'`, shows the package index route was considered and left off by default, and it keeps the same 3.10 marker.

Everything around it is exact or bounded rather than open. `uvicorn==0.29.0`, `pydantic==2.7.0`, `fastapi==0.110.2` and `pydantic-settings==2.2.1` will not move under you, while `setuptools>=61` and `setuptools-scm<8.0.0` leave room at both ends. One more conditional drops `numpymaxflow==0.0.7` on 3.13, and a source comment gives the reason for a second exclusion: `# scipy and scikit-learn latest packages are missing on python 3.8`. Those two marker lines are the only version-conditional code in a file holding more than fifty pinned entries.

Release 0.8.5 is dated a year before 0.8.4

Tag order and calendar order disagree. The three most recent releases are 0.8.5 dated 2023-10-18, 0.8.4 dated 2024-10-17 and 0.8.3 dated 2024-07-23. The highest version number is the oldest of the three by about a year, and nothing later than 2024-10-17 appears in the list, even though the last push to `main` is dated 2026-09-30.

Two readings fit and the project settles neither. Either the lower tags were cut from a maintenance branch after 0.8.5 shipped, or 0.8.5 was tagged from an older tree and development moved on to the branch. What you cannot do is treat the descending tag list as an upgrade path and assume 0.8.5 is what everyone is running.

Version discovery has its own mechanism elsewhere in the tree. `pyproject.toml` and `setup.py` both sit at the root with `versioneer.py` beside them, and the container build stamps the artifact with a number taken from the clock:

code
RUN BUILD_OHIF=false python setup.py bdist_wheel --build-number $(date +'%Y%m%d%H%M')

So a container build produces a different version string from the same source than the package index does. That matters before you try to compare two installed copies of the thing.

setup.py walks two whole directories into every install

`setup.py` walks the tree twice before it configures anything. The data file list starts with the logging config and then absorbs every file under two directories:

code
data_files = [("logconfig", ["monailabel/logging.json"])]
data_files.extend(recursive_files("sample-apps", "monailabel"))
data_files.extend(recursive_files("plugins", "monailabel"))

The helper prints a count for each pass and a total, so a build log is the place to see how many files a distribution carries. The consequence is that sample applications and viewer plugins are package data rather than optional extras. A plain install pulls the whole `sample-apps` tree, and the OHIF viewer source sits inside `plugins/` where the walk will find it.

The container build confirms the cost by spending its first stage on that viewer:

code
FROM ${NODE_IMAGE} AS ohifbuild
COPY plugins/ohifv3 /opt/ohifv3
RUN cd /opt/ohifv3 && ./build.sh /opt/ohifv3/release

A Node image is pulled in purely to build the bundle, and the output is copied back into the Python package path. Setting `BUILD_OHIF=false` skips it. Leaving the variable unset and running a local build reaches for the same Node toolchain whether or not the machine has one.

The final container stage is ubuntu:22.04, not PyTorch

The first line of the Dockerfile is commented out and the second is live, which quietly changes what the image is:

code
#ARG FINAL_IMAGE=pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime
ARG FINAL_IMAGE=ubuntu:22.04
ARG BUILD_IMAGE=python:3.10

The comment above them names the way to switch, pointing at a clean run of the test script followed by a build:

code
./runtests.sh --clean && DOCKER_BUILDKIT=1 docker build -t projectmonai/monailabel:latest .

The default final stage is plain Ubuntu with no CUDA runtime inside it, and the container installs its own Python stack afterwards rather than inheriting one from a base image. That leaves the two stages on different interpreters by default: the wheel is built on `python:3.10` and then copied into an `ubuntu:22.04` filesystem whose system Python is older. Nothing in the file resolves it, and the option the comment offers is to build against a PyTorch image instead.

On a workstation with a GPU driver already configured you may notice neither detail. The first person to run the image on a machine without one will.

DeepGrow exists only in the radiology row

The supported matrix is the most concrete thing in the project and the easiest to misread. Radiology lists `Segmentation`, `DeepGrow`, `DeepEdit` and SAM 2 as a 2D and 3D variant. Pathology lists `DeepEdit`, `NuClick`, `Segmentation`, `Classification` and SAM 2 in 2D. Video lists `DeepEdit`, `Tooltracking`, `InBody/OutBody` and SAM 2 in 2D.

These are not interchangeable labels. `Segmentation` on its own is a plain task. `DeepGrow` appears in the radiology row alone, so a pathology or endoscopy reader looking for click-driven growth will not find it named at all. `NuClick` appears in one row, `Tooltracking` in another, and `InBody/OutBody` belongs to video only.

Viewer pairings are just as uneven. Radiology pairs 3DSlicer, MITK and OHIF with NIfTI, NRRD and DICOM data types for CT and MRI. Pathology pairs Digital Slide Archive, QuPath and CVAT with TIFF and SVS. Video lists CVAT alone, with JPG and 3-channel frames, and it is the only row that claims an automated active learning workflow.

The caveats sit next to the table: those rows are the combinations the maintainers have tested, and the text says that does not mean your dataset or file type will fail.

Five getting started steps stop at the first heading

This is the part to plan around. The repository index names five steps and stops at the first one:

code
- [Step 1. Installation](#step-1-installation)
- [Step 2. MONAI Label Sample Applications](#step-2-monai-label-sample-applications)
- [Step 3. MONAI Label Supported Viewers](#step-3-monai-label-supported-viewers)
- [Step 4. Data Preparation](#step-4-data-preparation)
- [Step 5. Start MONAI Label Server and Start Annotating!](#step-5-start-monai-label-server-and-start-annotating)

Anchors and titles, no commands. Where the operations live is stated twice and both statements agree: the full walkthrough is hosted at `https://monai.readthedocs.io/projects/label/en/latest/`, and the worked examples are notebook tutorials in the Project-MONAI tutorials repository on the `monailabel` path.

What you can assemble from this repository alone is the dependency set, with `monai` at 1.4.0 or newer carrying a long extras list, plus `fastapi`, `uvicorn`, `pydantic`, `girder-client` and `shapely` for the DICOM side, and the container recipe above. How the server is pointed at a folder of images, which environment variables carry those paths and what the sample apps expect are documentation questions rather than repository questions.

Editorial conclusion

Adopt MONAI Label when the annotation work is clinical imaging, when 3DSlicer, OHIF, QuPath, Digital Slide Archive or CVAT is already in your loop, and when a pinned dependency set is preferable to a moving one. It is the wrong shape for a general image labelling project and for anyone who wants a viewer inside the same install. Before committing, check three things against your own setup: whether your Python is 3.10 or newer, which tagged release you actually want given that 0.8.5 is dated a year before 0.8.4, and whether the container's ubuntu:22.04 final stage matches the CUDA runtime you need. The last push to main is dated 2026-09-30 while the newest release is dated 2024-10-17, and planning around the tag rather than the branch is the mistake that list invites.

Frequently asked questions

What is MONAI Label?

MONAI Label is an open source image labeling and learning tool for building annotated datasets and training AI annotation models for clinical evaluation. It runs as a server, and custom labeling apps are exposed as a service through the MONAI Label Server while the viewer that draws the annotations is a separate application.

How do I use MONAI Label?

The server and the client can sit on the same machine or on different machines. Radiological work pairs the server with 3DSlicer, MITK or OHIF over NIfTI, NRRD and DICOM files, pathology pairs it with Digital Slide Archive, QuPath or CVAT over TIFF and SVS, and endoscopy video uses CVAT.

How do I use MONAI Label in 3D Slicer?

The supported matrix lists 3DSlicer first among the radiology viewers, alongside MITK and OHIF, with CT and MRI as the target modalities. The plugin directory holds the 3DSlicer and OHIF integrations and a separate CVAT one, and CVAT is the client listed for pathology and endoscopy instead.

Official sources

  1. License: Apache-2.0
  2. Project-MONAI/MONAILabel on GitHub
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/project-monai-monailabel.svg)](https://hysenlabs.com/projects/project-monai-monailabel)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/project-monai-monailabel.svg)](https://hysenlabs.com/projects/project-monai-monailabel)

Was this useful?