mdreg

Full Changelog: https://github.com/QIB-Sheffield/mdreg/compare/v0.1.0...v0.3.0

Description

Python implementation of model-based image coregistration for quantitative medical imaging applications.

The distribution comes with a number of common signal models and uses ITK-Elastix for deformable image registration.

Installation

Run pip install mdreg.

Example data

Example data in DICOM format are provided for testing the setup.

How to use

Input data must be image arrays in numpy format, with dimensions (x,y,z,t) or (x,y,t). To perform MDR on an image array im with default settings do:

from mdreg import MDReg

mdr = MDReg()
mdr.set_array(im)
mdr.fit()

When fitting is complete the motion-corrected data are in mdr.coreg in the same dimensions as the original im. The calculated deformation fields in format (x,y,d,t) or (x,y,z,d,t) can be found as mdr.deformation. The dimension d holds x, y components of the deformation field, and a third z component if the input array is 3D.

The default settings will apply a linear signal model and coregistration as defined in the elastix parameter file Bsplines.txt.

Customization

MDR can be configured to apply different signal models and elastix coregistration settings. A number of example models and alternative elastix parameter files are included in the distribution as templates.

The following example fits a mono-exponential decay and applies an elastix parameter file par_file optimized for a previous DTI-MRI study:

from mdreg import MDReg
from mdreg.models import exponential_decay

mdr = MDReg()
mdr.set_array(im)
mdr.signal_model = exponential_decay
mdr.read_elastix(par_file)
mdr.fit()

The signal model often depends on fixed constants and signal parameters such as sequence parameters in MRI, or patient-specific constants. These should all be grouped in a list and set before running the signal model.

Equally elastix parameters can be fine tuned, either by importing a dedicated elastix file, or by modifying the settings.

Then a number of parameters are available to optimize MDR such as the precision (stopping criterion) and maximum number of iterations.

Some examples:

from mdreg import MDReg
from mdreg.models import exponential_decay

t = [0.0, 1.25, 2.50, 3.75]     # time points for exponential in sec

mdr = MDReg()
mdr.set_array(im)
mdr.signal_parameters = t
mdr.signal_model = exponential_decay
mdr.set_elastix(MaximumNumberOfIterations = 256)   # change defaults
mdr.precision = 0.5         # default = 1
mdr.max_iterations = 3      # default = 5
mdr.fit()

mdreg comes with a number of options to export results and diagnostics:

mdr.export_unregistered = True      # export parameters and fit without registration
mdr.export_path = filepath          # default is a results folder in the current working directory
mdr.export()                        # export results after calling fit.

This export creates movies of original images, motion corrected images, modelfits, and maps of the fitted parameters.

Model fitting without motion correction

MDReg also can be used to perform model fitting without correcting the motion. The following script fits a linearised exponential model to each pixel and exports data of model and fit:

from mdreg import MDReg
from mdreg.models import exponential_decay

mdr = MDReg()
mdr.set_array(im)
mdr.signal_model = linear_exponential_decay
mdr.fit_signal()
mdr.export_data()
mdr.export_fit()

Defining new MDR models

A model must be defined as a separate module or class with two required functions main() and pars().

pars() must return a list of strings specifying the names of the model parameters. main(im, const) performs the pixel based model fitting and has two required arguments. im is a numpy ndarray with dimensions (x,y,z,t), (x,y,t) or (x,t). const is a list of any constant model parameters.

The function must return the fit to the model as an numpy ndarray with the same dimensions as im, and an ndarray pars with dimensions (x,y,z,p), (x,y,p) or (x,p). Here p enumerates the model parameters.

Context

mdreg was first developed for use in quantitative renal MRI in the iBEAt study, and validated against group-wise model-free registration (Tagkalakis F, et al. Model-based motion correction outperforms a model-free method in quantitative renal MRI. Abstract-1383, ISMRM 2021).

Acknowledgement

The iBEAt study is part of the BEAt-DKD project. The BEAt-DKD project has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement No 115974. This Joint Undertaking receives support from the European Union's Horizon 2020 research and innovation programme and EFPIA with JDRF. For a full list of BEAt-DKD partners, see www.beat-dkd.eu.

Authors

Kanishka Sharma, Joao Almeida e Sousa, Steven Sourbron