Overview

Project MONAI

Medical Open Network for AI - Toolkit for Healthcare Imaging

This document identifies key concepts of project MONAI at a high level, the goal is to facilitate further technical discussions of requirements, roadmap, feasibility and trade-offs.

Vision

• Develop a community of academic, industrial and clinical researchers collaborating and working on a common foundation of standardized tools.
• Create a state-of-the-art, end-to-end training toolkit for healthcare imaging.
• Provide academic and industrial researchers with the optimized and standardized way to create and evaluate models

Targeted users

• Primarily focused on the healthcare researchers who develop DL models for medical imaging

Goals

• Deliver domain-specific workflow capabilities
• Address the end-end “Pain points” when creating medical imaging deep learning workflows.
• Provide a robust foundation with a performance optimized system software stack that allows researchers to focus on the research and not worry about software development principles.

Guiding principles

Modularity

• Pythonic -- object oriented components
• Compositional -- can combine components to create workflows
• Extensible -- easy to create new components and extend existing components
• Easy to debug -- loosely coupled, easy to follow code (e.g. in eager or graph mode)
• Flexible -- interfaces for easy integration of external modules

User friendly

• Portable -- use components/workflows via Python “import”
• Run well-known baseline workflows in a few commands
• Access to the well-known public datasets in a few lines of code

Standardisation

• Unified/consistent component APIs with documentation specifications
• Unified/consistent data and model formats, compatible with other existing standards

High quality

• Consistent coding style - extensive documentation - tutorials - contributors’ guidelines
• Reproducibility -- e.g. system-specific deterministic training

Future proof

• Task scalability -- both in datasets and computational resources
• Support for advanced data structures -- e.g. graphs/structured text documents

Leverage existing high-quality software packages whenever possible

• E.g. low-level medical image format reader, image preprocessing with external packages
• Rigorous risk analysis of choice of foundational software dependencies

Compatible with external software

• E.g. data visualisation, experiments tracking, management, orchestration

Key capabilities

Basic features	Example		Notes
Ready-to-use workflows	Volumetric image segmentation		“Bring your own dataset”
Baseline/reference network architectures	Provide an option to use “U-Net”
Intuitive command-line interfaces
Multi-gpu training	Configure the workflow to run data parallel training

Customisable Python interfaces	Example		Notes
Training/validation strategies	Schedule a strategy of alternating between generator and discriminator model training
Network architectures	Define new networks w/ the recent “Squeeze-and-Excitation” blocks		“Bring your own model”
Data preprocessors	Define a new reader to read training data from a database system
Adaptive training schedule	Stop training when the loss becomes “NaN”		“Callbacks”
Configuration-driven workflow assembly	Making workflow instances from configuration file		Convenient for managing hyperparameters

Model sharing & transfer learning	Example		Notes
Sharing model parameters, hyperparameter configurations	Standardisation of model archiving format
Model optimisation for deployment
Fine-tuning from pre-trained models	Model compression, TensorRT
Model interpretability	Visualising feature maps of a trained model
Experiment tracking & management			https://polyaxon.com/

Advanced features	Example		Notes
Compatibility with external toolkits	XNAT as data source, ITK as preprocessor
Advanced learning strategies	Semi-supervised, active learning
High performance preprocessors	Smart caching, multi-process
Multi-node distributed training