mermaid: iMagE Registration via autoMAtIc Differentiation [*]

mermaid is a registration toolbox, which supports various image registration methods. In particular, it focuses on nonparametric registration approaches (including stationary velocity fields and large discplacement diffeomorphic metric mapping models) though simple affine registration is also possible. As it is entirely written in pyTorch it allows for rapid prototyping of new image registration approaches and similiarity measures. To keep track of registration parameters it makes use of json configuration files which entirely describe registration algorithms. mermaid provides optimization-based registration approaches, but the companion-package easyreg (https://github.com/uncbiag/easyreg) adds support for deep-learning registration models by building on top of mermaid’s transformation models.

mermaid was primarily developed by:

• Marc Niethammer
• Zhengyang Shen
• Roland Kwitt

Notes

• Installation
• Parameters
• Mermaid configurations
• TODOs
• How to write your own registration model
• Simple example
• Step by step registration example
• RDMM example on synthetic data
• Example Gallery

Jupyter notebook examples

• Demonstration of mermaid’s simple registration interface

Package reference

• Simple interface
• Registration models
  • Model factory
  • Forward models
  • Model evaluation
  • Registration networks
• Similarity measures
  • Similarity measure factory
  • Similarity helper OMT
• Basic Numerics
  • Finite differences
  • Runge-Kutta integrators
  • Spline interpolation
• Optimizers
  • Multiscale optimizer
  • Custom optimizers
  • Optimizer data loaders
• Visualization
  • Visualize registration results
  • Viewers
• Configurations
  • Config parser
  • Load default settings
  • Parameter module
• Image IO
  • fileio
  • Example generation
  • Image sampling
  • Image manipulations
• Regularization and Smoothing
  • Regularizer factory
  • Smoother factory
  • Deep smoothers
• Utilities and Extensions
  • Data wrapper
  • Utilities
  • Custom pyTorch extensions
  • Spatial transforms
• Deep networks
  • Deep network implementations
  • Noisy convolution

Indices and tables

• Index
• Module Index

[*]	OK, we just liked the acronym.