FAQ

Frequently Asked Questions about MONAI Physio.

General Questions

What is MONAI Physio?

MONAI Physio is a collection of methods, workflows, tutorials, and CLI tools for creating personalized physiological digital twins: starting from a 3D medical image of a subject, extracting anatomic models, and then using AI surrogates to estimate the subject’s physiological processes (initially cardiac and respiratory motion, expanding to electrophysiology, blood flow, and organ perfusion).

What data formats are supported?

Input: NRRD, MHA, NIfTI, DICOM
Output: USD (Universal Scene Description), VTK

Do I need NVIDIA Omniverse?

Omniverse is the recommended way to view the USD scenes: its RTX renderer is what evaluates the material properties assigned to each tissue. See Viewing USD Files. For the intermediate results you can also use:

PyVista, for the intermediate .vtp / .vtu meshes
ParaView, likewise for the VTK files

Installation Questions

Do I need a GPU?

No. A plain pip install monai-physio works without a GPU, and runs every workflow including the AI-surrogate ones - just significantly slower than a GPU-enabled install. At import time a UserWarning is emitted (visible by default in all standard Python runs):

CuPy is not installed - GPU-accelerated mesh operations will fall back to
NumPy and run significantly slower. Every workflow still runs. Re-install
with uv to get CuPy and CUDA-enabled PyTorch in one step (pip alone will
not select the correct CUDA wheel):
  uv pip install 'monai-physio[cuda12]'  # CUDA 12.6
  uv pip install 'monai-physio[cuda13]'  # CUDA 13

CPU-only mode is suitable for evaluation and small datasets. For production workloads an NVIDIA GPU is strongly recommended.

Which CUDA version is required?

Both CUDA 13 and CUDA 12.6 are supported. CUDA 12.6 is the smoothest install, because it is the newest combination with prebuilt torch-scatter wheels on every supported platform - nothing compiles:

uv pip install "monai-physio[cuda12]"   # CUDA 12.6
uv pip install "monai-physio[cuda13]"   # CUDA 13, builds torch-scatter

Each extra installs CuPy and pins PyTorch, torchvision, and torchaudio to the matching wheel index (https://download.pytorch.org/whl/cu126 or cu130). On CUDA 13 there is no prebuilt torch-scatter wheel, so it compiles from source and needs a CUDA toolkit plus a C++ toolchain; see Installation.

To let uv pick the PyTorch build from your driver without naming a CUDA version - at the cost of CuPy, which has no auto-detect option:

uv pip install --torch-backend=auto monai-physio

What Python version is required?

Python 3.11, 3.12 and 3.13 are supported.

The floor is set by nvidia-physicsnemo (a base dependency): it supports >= 3.11, < 3.14, and the AI-surrogate tutorials need it. The rest of the library would run on 3.10, but declaring 3.10 would promise an install that cannot resolve.

Usage Questions

How long does processing take?

Typical processing time for 10-frame cardiac CT (with GPU):

4D to 3D conversion: ~1 minute
Registration: ~5-10 minutes
Segmentation: ~1-2 minutes
USD creation: ~1 minute
Total: ~10-15 minutes

Which segmentation method should I use?

TotalSegmentator: Fast, good quality, general purpose
Simpleware: Best quality for cardiac imaging, requires Simpleware Medical
NV-Segment-CTMR: CT and MRI, 345 classes; weights are licensed for non-commercial academic research only

See Segmentation Modules for comparison.

Which registration method should I use?

Greedy: CPU-capable classical deformable registration; what Tutorials 1 and 3 use by default
ICON: Recommended for cardiac/lung (fast, GPU), and finetunable on your own cohort - see Tutorial 2
ANTs: Best for brain imaging and general purpose
Greedy+ICON (RegisterImagesGreedyICON, a RegisterImagesChain preset): Greedy for the coarse alignment, ICON for the refinement

See Image Registration Modules for comparison.

Troubleshooting

See Troubleshooting for common issues and solutions.