An MR Motion Correction toolbox for registration and evaluation

Thomas Küstner^1,2, Verena Neumann², Martin Schwartz^1,2, Christian Würslin^1,3, Petros Martirosian¹, Sergios Gatidis¹, Nina F. Schwenzer¹, Fritz Schick¹, Bin Yang², and Holger Schmidt¹

¹Department of Radiology, University Hospital Tübingen, Tübingen, Germany, ²Institute of Signal Processing and System Theory, University of Stuttgart, Stuttgart, Germany, ³University of Stanford, Palo Alto, CA, United States

Synopsis

Motion estimation is an important task in MRI. For retrospective motion correction, there is often an image-based registration involved. Hence, the extraction of a reliable and accurate motion model for the underlying application is mainly dependent on the chosen image registration procedure. There are several different image registration methods available, but visualization and evaluation of the derived displacement fields and transformed images often remains an open topic. In the spirit of a reproducible research and for streamlining and simplifying the process, we provide GUIs and evaluation methods to perform and analyze image registration techniques which will be made publicly available.

Purpose

Motion visualization and estimation is an important task in MRI. All methods aim to correct or omit motion-induced artifacts. The techniques can be distinguished into prospective or retrospective methods. In the case of a retrospective motion correction (MC) or motion estimation, an image registration is often involved. Hence the chosen image registration method is one of the key steps to derive an accurate motion field. There is a variety of applications, like respiratory MC¹, cardiac MC² or abdominal DWI³, for which a motion model can be applied to. Motion models are often derived image-based by registration of images acquired at different motion states. It is therefore essential to first derive an accurate motion model and then to analyze and evaluate the resulting displacements and transformed images. There are different image registrations available, like parametric B-Spline models^4,5, non-parametric diffusion-based methods⁶ or optical flow based methods^7,8 offering a variety of parameters to tune the algorithms for their specific tasks. These algorithms are often just command-line based which make them not very easy and handy to use. Furthermore some algorithms require certain data formats which demand a data conversion. After image registration, visualization can be performed by some other third-party programs, e.g. 3DSlicer⁹, which maybe require another data conversion. Despite their open-source architecture, it can get time-consuming and difficult to implement new algorithms or to perform a thorough analysis with the given tools. Since many research groups work primarily with Matlab due to its simple access to first develop new algorithms, a more streamlined process within this scope would be desired.
As indicated by Rohlfing et al.¹⁰, a mere evaluation based on intensity similarity metrics is often not sufficient to quantify and evaluate the underlying motion. Looking at anatomic landmark points or regions helps to identify the potential feasibility of the performed image registration algorithms.
In the spirit of a reproducible research and in order to simplify and streamline the motion correction and evaluation procedure, we propose a Matlab toolbox with easy extendable graphical user interfaces (GUI) to perform and analyze image registrations.

Methods

A Matlab toolbox consisting of three GUIs was developed which allows performing image registrations and analyzing the obtained displacement fields, see Fig.1. First the registration GUI (RegGUI, Fig.2) loads the 2D or 3D image data from DICOM, NIfTI, MAT, GIPL or MHD format and allows the selection of the registration algorithms: elastix¹¹, hierarchical adaptive local affine registration⁵, Local All-Pass (LAP)⁸ and Demons⁶ with their respective parametrizations. Furthermore, a plugin for the Matlab tool imagine¹² is provided to streamline the process of image loading. After image registration, the images and displacement fields can be viewed and analyzed in an evaluation GUI (EvalGUI, Fig.3). Zooming, scaling, rotating and switching between the forward and backward displacement field, as well as between the original and transformed image is supported to comfort the quality assessment. First quantitative insights on the deformation can be gathered and visualized from the divergence, determinant of Jacobian or the absolute displacement of the vector fields. The similarity between the images can be examined via intensity-based metrics. In the case of respiratory motion correction, the images can be analyzed by means of overlap measures of the deformed lung masks which are generated by a modified version of lung segmentation according to Kohlmann¹³. Furthermore, a feature-based evaluation is provided by means of anatomic landmark points (LandmarkGUI, Fig.4). Readers can label anatomic points, lines and region of interests and evaluate the respective Dice overlap measures between the original and transformed images. All obtained results can be exported for further evaluations.
The Matlab code is easily extendable for other algorithms and will be made available online at Matlab Central FileExchange and Github under Apache 2.0 license:
https://github.com/thomaskuestner/MoCoGUI

Results

An exemplary free-breathing coronal dataset was acquired with a spoiled gradient echo sequence¹⁴ and retrospectively gated into four respiratory motion states¹⁵. The gated images are co-registered by means of a non-rigid multilevel and multidimensional cubic B-Spline model of elastix, the optical flow-based LAP algorithm and the diffusion-based Demons algorithm. Fig. 5 shows the end-expiratory images of the different registration algorithms with their corresponding motion fields and their absolute displacements.

Conclusion

The proposed Matlab toolbox provides easy access to perform image registrations and to evaluate, visualize and analyse the obtained deformation fields and transformed images. This work shall help to broaden and simplify the access for motion studies as well as helping to create a reproducible research environment.

Acknowledgements

No acknowledgement found.

References

[1] McClelland et al., Med Image Anal 2013;17. [2] Scott et al., Radiology 2009;250. [3] Martirosian et al., Proc ISMRM 2015. [4] Rueckert et al., TMI 1999;18(8). [5] Buerger et al., Med Image Anal 2011;15(4). [6] Thirion et al., Med Image Anal 1998;2(3). [7] Odille et al., MRM 2008;60(1). [8] Gilliam et al., IEEE Proc ICASSP 2015. [9] Fedorov et al., MRI 2012;30(9). [10] Rohlfing et al., TMI 2012;31(2). [11] Klein et al., TMI 2010;29(1). [12] Würslin, Matlab FEX 2013. [13] Kohlmann et al., Int J Comput Assist Radiol Surg 2015;10(4). [14] Küstner et al., Proc ISMRM Workshop Motion Correction 2014. [15] Küstner et al., IEEE Proc ICASSP 2015.

Figures

Figure 1: System overview of registration and evaluation toolbox.

Figure 2: RegGUI: registration GUI for performing image registration.

Figure 3: EvalGUI: visualization and analysis of motion models derived from image registration.

Figure 4: LandmarkGUI: feature-based evaluation by means of anatomic landmark points.

Figure 5: Comparison between different registration algorithms: a) elastix, b) LAP and c) Demons. End-inspiratory images with overlayed motion field to end-expiratory state (first line), absolute displacement (second line) and divergence of transformation (third line).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)

1847