/Demeter_metamorphosis

Open library for LDDMM and Metamorphosis in PyTorch

Primary LanguagePython

Demeter 0.2

(DiffEoMorphic METamorphic Regristration)

Warning : This depot is still in progress.

You are seeing the evolution of a new python librairy, intended to solve diffeomoprhic registration problems using metamorphosis, LDDMM and derived methods. For more details on the methods please read the joined PDF : Metamorphic_image_registration_afrancois_pgori_jglaunes.

Why Demeter ? Because in ancient Rome, Demeter was the goddess of agriculture. Isn't it the perfect protection for or vectors fields ?

Requirement

In this project depend on plural libraries, some quite basics like :

  • Numpy
  • math
  • matplotlib
  • time

and some more complex. make sure that you have :

Jupyter Notebooks and examples.

You can see results and usage exemples in the following jupyter notebooks

  • figs_geodesic_stability_lddmm.ipynb : Figure from the article aiming to show the stability of semi-Lagrangian schemes over Eulerian ones for LDDMM.
  • fig_lddmm_vs_metamorphoses.ipynb : Figure from the article aiming to compare our implementation of LDDMM and Metamorphosis.
  • toyExample_weightedMetamorphosis.ipynb : Examples and results of Weighted Metamorphosis method on a synthetic toy example image.
  • brain_weightedMetamorphosis.ipynb : Examples and results of Weighted Metamorphosis method on real data.

Note : vedo visualisation currently does not work in anaconda environment such as jupyter or ipython. If you want to display a vedo visualisation that we implemented, run them in a python terminal. See :

  • examples/metamorphosis_3D.py
  • examples/visualizeMetamorphosis_3D.py

Files

  • metamorphosis.py : is containing the two classes for LDDMM/metamorphic regristration

    • class Geodesic_integrator : (abstract) is the class 'tool kit' for integrating over a geodesic
    • class Optimise_geodesicShooting : (abstract) is the class 'tool kit' for optimizing using geodesic shooting methods. It works in combinaison with a child class of Geodesic_integrator.
    • class Metamophosis_path: (derived from Geodesic_integrator) is made for integration over LDDMM or Metamorphosis geodesics.
    • class Optimise_Metamorphosis: (derived from Optimise_geodesicShooting) Optimize a Metamorphosis_path object.
    • class Residual_norm_function: (abstract) Generic 'tool kit' class for defining a custom norm on $z_0$
    • class Residual_norm_identity: (derived from Residual_norm_function) definition of $|\sqrt{M_t}z_t|_{L_2}^2 = \lbrace z_t M_t ,z_t \rbrace$
    • class Residual_norm_borderBoost: (derived from Residual_norm_function) Experimental
    • class Weighted_meta_path: (derived from Geodesic_integrator) is made for integration with Weighted Metamorphosis method. It is defined with a function derived from an child object of Residual_norm_function.
    • class Weighted_optim: (derived from Optimise_geodesicShooting) Optimize a Weighted_meta_path object.

  • my_optim.py : is containing the gradient descent class.

  • my_toolbox.py : is containing generic utilitary fonctions

  • my_torchbox.py : is containing utilitary function compatible with PyTorch such as:

    • wrapper for kornia and PyTorch utilitary functions.
    • function for plotting.

  • fft_conv.py : function from Github : fkodom

  • reproducing_kernels.py : Kornia inerited functions to use filter convolution with fft_conv.

  • constants.py : Files with different constants reused at different locations

  • image_3d_visualisation : Visualisation function for 3D volumetric images using matplotlib or vedo.

  • fill_saves_overview.py : Gestion of saved geodesics_shooting in the csv file.

Contact

You can find my contact on my website : http://helios.mi.parisdescartes.fr/~afrancoi/