MONAI

MONAI is Project MONAI: an open-source PyTorch toolkit for medical imaging (transforms, networks, losses, a dictionary dataset). This page is what it is and how you start. It is not a 3D preprocessing notebook. It is not a 3D augmentation notebook. It is not a PYCAD product.

If you meant load / resample / window / crop a volume → preprocessing 3D volumes in MONAI. If you meant RandAffine / rotate / noise → 3D volume augmentation in MONAI. If you meant DICOM → NIfTI first → DICOM to NIfTI. If you meant imaging ML as a job → machine learning for medical imaging.

The slug is monaigpt because this URL used to sell a 2023 Streamlit docs assistant. That demo is a side note below. The evergreen job is the framework.

PYCAD builds custom web DICOM viewers and imaging models. It does not sell MONAI, a MONAI SaaS, or a “MONAIGPT” product.

What MONAI actually is

Project MONAI (Medical Open Network for AI) is a Linux Foundation project. The useful mental model is three boxes that share a name:

Most people who say “I installed MONAI” mean Core. The 2021 notebooks on this site are Core. Docs: docs.monai.io. Code and tutorials: github.com/Project-MONAI.

It is a library, not a diagnosis. A UNet you train on public NIfTI is not a CAD. A transform compose is not a PYCAD SKU.

How to start (Core)

Use a venv. Current MONAI wants a matching PyTorch. Then:

pip install monai nibabel
python -c "import monai; print(monai.__version__)"

The first real program is a dictionary of paths, not two parallel lists you hope stay aligned. Every transform takes keys=, so intensity hits the CT and spacing hits the CT and the mask:

from monai.transforms import Compose, LoadImaged, EnsureChannelFirstd, Spacingd, ScaleIntensityRanged, ToTensord

train_files = [{"image": "case_001.nii.gz", "label": "case_001_seg.nii.gz"}]

xform = Compose([
    LoadImaged(keys=["image", "label"]),
    EnsureChannelFirstd(keys=["image", "label"]),
    Spacingd(keys=["image", "label"], pixdim=(1.5, 1.5, 2.0), mode=("bilinear", "nearest")),
    ScaleIntensityRanged(keys=["image"], a_min=-200, a_max=200, b_min=0.0, b_max=1.0, clip=True),
    ToTensord(keys=["image", "label"]),
])

EnsureChannelFirstd is the current name. Older notebooks (including the 2021 ones on this site) still say AddChanneld. Same channel dim; different year. Do not copy a 2021 import into a 2026 env and call the traceback a MONAI bug.

After that, the official tutorials repo is the next hour: spleen / BTCV segmentation, a 2D classification, sliding-window infer. The two PYCAD notebooks that stay on their own URLs are the 3D preprocess compose and the 3D aug compose — they are not this page restated.

What this URL used to be

In May 2023 the PYCAD Team shipped a Streamlit demo, monaigpt.streamlit.app, that answered MONAI Core docs questions with GPT-3.5. It did not cover Label or Deploy. It was a community demo, not a product. Official docs and the tutorials repo are the start path. If the Streamlit app is up, it is still just a chatbot over Core docs — not a substitute for running the compose above.

What not to do

• Do not train on PNG screenshots of a viewer and call it a MONAI pipeline. The file is NIfTI (or a DICOM series you converted).
• Do not skip spacing. Two sites with 0.7 mm and 2.5 mm slices are not the same grid.
• Do not treat Dice on a public challenge as a clearance. Metrics live on the evaluation URL; this page is the toolkit.
• Do not invent a PYCAD MONAI cloud. Viewer / model work can use Core. That is not a SKU named MONAI.

Start: pip install monai, the dictionary compose, then one official tutorial. Preprocess and aug stay on 3050 and 3063.