fastMONAI 0.9.1

fastMONAI library

Overview

A low-code Python-based open source deep learning library built on top of fastai, MONAI, TorchIO, and Imagedata.

fastMONAI simplifies the use of state-of-the-art deep learning techniques in 3D medical image analysis for solving classification, regression, and segmentation tasks. fastMONAI provides the users with functionalities to step through data loading, preprocessing, training, and result interpretations.

Note: This documentation is also available as interactive notebooks.

Requirements

• Python: 3.10, 3.11, 3.12, or 3.13 (Python 3.12 recommended)
• GPU: CUDA-compatible GPU recommended for training (CPU supported for inference)

Installation

Environment setup (recommended)

We recommend using a conda environment to avoid dependency conflicts:

conda create -n fastmonai python=3.12

conda activate fastmonai

Quick Install (PyPI)

pip install fastMONAI

Development install (GitHub)

If you want to install an editable version of fastMONAI for development:

git clone https://github.com/MMIV-ML/fastMONAI
cd fastMONAI

# Create development environment
conda create -n fastmonai-dev python=3.12
conda activate fastmonai-dev

# Install in development mode
pip install -e '.[dev]'

Getting started

The best way to get started using fastMONAI is to read our paper and dive into our beginner-friendly video tutorial. For a deeper understanding and hands-on experience, our comprehensive instructional notebooks will walk you through model training for various tasks like classification, regression, and segmentation, running inference with a trained model, and patch-based training for large volumes that do not fit in memory. See the docs at https://fastmonai.no for more information.

Notebook	1-Click Notebook
11b_tutorial_classification.ipynb shows how to construct a binary classification model based on MRI data.
11c_tutorial_regression.ipynb shows how to construct a model to predict the age of a subject from MRI scans (“brain age”).
11d_tutorial_binary_segmentation.ipynb shows how to do binary segmentation (extract the left atrium from monomodal cardiac MRI).
11e_tutorial_multiclass_segmentation.ipynb shows how to perform segmentation from multimodal MRI (brain tumor segmentation).
11f_tutorial_inference.ipynb shows how to run inference on new data with an exported learner.	Runs after 11d (inference on an exported model)
12a_tutorial_patch_training.ipynb shows how to train a patch-based segmentation model on large 3D volumes with lazy loading (memory stays constant).
12b_tutorial_patch_cross_validation.ipynb shows how to run k-fold cross-validation and train a final patch-based model.
12c_tutorial_patch_inference.ipynb shows how to deploy a patch-based model with sliding-window inference on new scans.	Runs after 12b (inference on an exported model)

How to contribute

We welcome contributions! See CONTRIBUTING.md

Citing fastMONAI

If you are using fastMONAI in your research, please use the following citation:

@article{KALIYUGARASAN2023100583,
title = {fastMONAI: A low-code deep learning library for medical image analysis},
journal = {Software Impacts},
pages = {100583},
year = {2023},
issn = {2665-9638},
doi = {https://doi.org/10.1016/j.simpa.2023.100583},
url = {https://www.sciencedirect.com/science/article/pii/S2665963823001203},
author = {Satheshkumar Kaliyugarasan and Alexander S. Lundervold},
keywords = {Deep learning, Medical imaging, Radiology},
abstract = {We introduce fastMONAI, an open-source Python-based deep learning library for 3D medical imaging. Drawing upon the strengths of fastai, MONAI, and TorchIO, fastMONAI simplifies the use of advanced techniques for tasks like classification, regression, and segmentation. The library's design addresses domain-specific demands while promoting best practices, facilitating efficient model development. It offers newcomers an easier entry into the field while keeping the option to make advanced, lower-level customizations if needed. This paper describes the library's design, impact, limitations, and plans for future work.}
}