The training library
built for medical imaging
Capabilities
Everything between raw scans and a trained model
Not a general ML framework retrofitted for healthcare: every layer, from data loading to evaluation, is designed around 3D medical volumes.
Medical-Specific Transforms
LoadImaged, Spacingd, RandCropByPosNegLabeld: dictionary transforms that keep image and label in sync
CacheDataset and smart caching for up to 10x faster data loading
Deterministic pipelines for reproducible training runs
DICOM, NIfTI, and PNG/JPEG I/O built in (2D, 3D, and 4D)
State-of-the-Art Architectures
MAISI: latent-diffusion generation of synthetic 3D CT with controllable anatomy
VISTA-3D: universal 3D segmentation, promptable across 100+ anatomical classes
UNETR and SwinUNETR transformer backbones, plus classic UNet variants
40+ pre-trained Model Zoo bundles ready for fine-tuning
Research Workflows
sliding_window_inference for whole-volume prediction on limited GPU memory
Dice, Hausdorff, and Surface Distance metrics with batched GPU evaluation
Multi-GPU (DDP) and mixed-precision training out of the box
TensorBoard visualization and experiment tracking hooks
Quick Start
One install, ten composable modules
Import only what you need (transforms, networks, losses) or use the end-to-end workflows. Every module links to its API reference.
$ pip install monai
from monai.transforms import (
Compose, LoadImage,
EnsureChannelFirst, ScaleIntensity,
)
transforms = Compose([
LoadImage(image_only=True),
EnsureChannelFirst(),
ScaleIntensity(),
])
image = transforms("ct_chest.nii.gz")Want the full walkthrough? Browse the tutorials repository
Core Modules
Transforms
Spatial, intensity, and dictionary-based medical transforms
Networks
UNETR, SwinUNETR, SegResNet, and more reference nets
Losses
DiceLoss, DiceCELoss, FocalLoss for imbalanced anatomy
Metrics
Dice, Hausdorff, Surface Distance, and more
Data
CacheDataset, PersistentDataset, efficient loaders
Inferers
Sliding-window and patch-based inference
Optimizers
Optimizers and LR schedules for medical tasks
Applications
End-to-end apps: Auto3DSeg, MAISI, VISTA
Visualization
2D/3D plotting and TensorBoard utilities
Utils
Determinism, type conversion, helper functions
Automation
Auto3DSeg: segmentation that tunes itself
Point Auto3DSeg at a labeled dataset and it analyzes intensity, size, and spacing, configures candidate algorithms, trains them on your GPUs, and ensembles the winners. The same pipeline is behind multiple MICCAI challenge wins.
How it works
Dataset Analysis
Automatic profiling of intensity ranges, volume sizes, and voxel spacing to pick the right preprocessing.
Algorithm Generation
Generates ready-to-train algorithm folders (SegResNet, SwinUNETR, DiNTS) configured from the data assessment.
GPU-Accelerated Training
Multi-GPU training, validation, and inference without extra configuration.
Model Ensemble
Combines the best-performing candidates into a single, more accurate ensemble.
Proven in competition
MICCAI 2023 Challenges
- Multiple 1st place wins in BraTS 2023 (brain tumors)
- 1st place in KiTS 2023 (kidney segmentation)
- 1st place in SEG.A. 2023 (aorta segmentation)
- 1st place in MVSEG 2023 (mitral valve)
Impact
The default research stack for medical-imaging AI
The numbers are verifiable: each links to its source where one exists.
Peer-reviewed papers
Published research that builds on MONAI, across radiology, pathology, and beyond.
Challenge wins
MONAI-powered entries have topped medical-imaging competitions including BraTS and KiTS.
Model Zoo bundles
Pre-trained, reproducible bundles: weights, configs, and inference code packaged together.
Reference
How to cite MONAI Core
If MONAI supports your research, please cite the framework paper:
@article{cardoso2022monai,
title={MONAI: An open-source framework for deep learning in healthcare},
author={M Jorge Cardoso, Wenqi Li, Richard Brown, Nic Ma, Eric Kerfoot, Yiheng Wang, Benjamin Murrey, Andriy Myronenko, Can Zhao, Dong Yang, Vishwesh Nath, Yufan He, Ziyue Xu, Ali Hatamizadeh, Andriy Myronenko, Wentao Zhu, Yun Liu, Mingxin Zheng, Yucheng Tang, Isaac Yang, Michael Zephyr, Behrooz Hashemian, Sachidanand Alle, Mohammad Zalbagi Darestani, Charlie Budd, Marc Modat, Tom Vercauteren, Guotai Wang, Yiwen Li, Yipeng Hu, Yunguan Fu, Benjamin Gorman, Hans Johnson, Brad Genereaux, Barbaros S Erdal, Vikash Gupta, Andres Diaz-Pinto, Andre Dourson, Lena Maier-Hein, Paul F Jaeger, Michael Baumgartner, Jayashree Kalpathy-Cramer, Mona Flores, Justin Kirby, Lee A D Cooper, Holger R Roth, Daguang Xu, David Bericat, Ralf Floca, S Kevin Zhou, Haris Shuaib, Keyvan Farahani, Klaus H Maier-Hein, Stephen Aylward, Prerna Dogra, Sebastien Ourselin, Andrew Feng},
journal={arXiv:2211.02701},
year={2022}
} Connect
Build with the community
Questions, contributions, and roadmap discussions all happen in the open: on GitHub, Slack, and the working-group calls.
Documentation
API reference and concept guides, from your first transform to distributed training.
GitHub Repository
Source code, issues, and pull requests, with contributions reviewed and merged in the open.
Slack Community
Real-time help and collaboration with maintainers, researchers, and clinicians.
Tutorials
Hands-on notebooks covering segmentation, classification, registration, and generative models.