MONAI

MONAI (Medical Open Network for AI) is an open-source, PyTorch-based framework for deep learning in healthcare imaging, developed by Project MONAI. The project says it was started by NVIDIA and King's College London and is now a community of 30+ institutions. The current release listed is 1.6.1, and the code is licensed under Apache 2.0.

What It Is

MONAI Core is a training and research library for medical imaging AI. It provides flexible pre-processing for multi-dimensional medical imaging data, plus domain-specific networks, losses and evaluation metrics. It is built on PyTorch, so standard PyTorch components such as DataLoader work with it, and it supports multi-GPU, multi-node data parallelism. It is installed with pip, conda or a Docker image.

Domain tooling and models

The project site describes DICOM and NIfTI I/O, spatial transforms for 3D medical volumes, and segmentation metrics such as Dice, Hausdorff and Surface Distance. It lists reference implementations including MAISI, UNETR, SwinUNETR, VISTA-3D and Auto3DSeg. MONAI Bundles package weights, training configs, metadata and inference code, and the same format is used by the Model Zoo for sharing community models.

The wider MONAI ecosystem

Core is one of three projects that share a bundle format. MONAI Label is an active-learning annotation server that plugs into 3D Slicer, OHIF and MITK. MONAI Deploy covers the path to clinical pipelines, with DICOM and FHIR I/O and containerized MAP packaging. Tutorials, Colab notebooks such as the MedNIST demo, and API documentation on Read the Docs support onboarding.