Analysis of different image-registration algorithms for Fourier decomposition MRI in functional lung imaging

Abstract

Background

Motion correction is mandatory for the functional Fourier decomposition magnetic resonance imaging (FD-MRI) of the lungs. Therefore, it is important to evaluate the quality of various image-registration algorithms for pulmonary FD-MRI and to determine their impact on FD-MRI outcome.

Purpose

To evaluate different image-registration algorithms for FD-MRI in functional lung imaging.

Material and Methods

Fifteen healthy volunteers were examined in a 1.5-T whole-body MR scanner (Magnetom Avanto, Siemens AG) with a non-contrast enhanced 2D TrueFISP pulse sequence in coronal view and free-breathing (acquisition time 45 s, 250 images). Three image-registration algorithms were used to compensate the spatial variation of the lungs (fMRLung 3.0, ANTs, and Elastix). Quality control for image registration was performed by edge detection (ED), quotient image criterion (QI), and dice similarity coefficient (DSC). Ventilation, perfusion, and a ventilation/perfusion quotient (V/Q) were calculated using the three registered datasets.

Results

Average computing times for the three image-registration algorithms were 1.0 ± 1.6 min, 38.0 ± 13.5 min, and 354 ± 78 min for fMRLung, ANTs, and Elastix, respectively. No significant difference in the quality of motion correction provided by different image-registration algorithms occurred. Significant differences were observed between fMRLung- and Elastix-based perfusion values of the left lung as well as fMRLung- and ANTs-based V/Q quotient of the right and the entire lung (P < 0.05). Other ventilation and perfusion values were not significantly different.

Conclusion

The mandatory motion correction for functional FD-MRI of the lung can be achieved through different image-registration algorithms with consistent quality. However, a significantly difference in computing time between the image-registration algorithms still requires an optimization.

Keywords

Image-registration algorithm functional magnetic resonance imaging Fourier magnetic resonance imaging lung ventilation lung perfusion

Introduction

Various pulmonary diseases can lead to impairment of the ventilation/perfusion (V/Q) ratio of the lung. Pulmonary perfusion and ventilation are therefore important parameters of lung function (1). Methods for the measurement of pulmonary perfusion and ventilation used in clinical routine are highly dependent on patient compliance. To overcome this limitation, a new method called “Fourier decomposition magnetic resonance imaging” (FD-MRI), which facilitates quick determination of lung perfusion and ventilation without the application of intravenous contrast agents, has previously been proposed in the literature (2,3).

FD-MRI uses a pixel-wise frequency analysis of dynamically acquired MRI datasets to separate the signal changes arising from the respiratory and cardiac cycles (2,4). The pixel-wise analysis of a dataset presupposes that all associated anatomical structures remain spatially immutable. However, this condition is not met because of volume changes and mobility of the lungs during breathing. Therefore, retrospective image registration of the acquired data is mandatory.

The aim of the present study was to evaluate the quality of image registration for FD-MRI using different image-registration algorithms (“fMRLung,” Advanced Normalization Tools [“ANTs”], and “Elastix”) and determine their effects on the quantified perfusion and ventilation parameters.

Material and Methods

Study population

The present study was approved by the local ethics committee, and written informed consent was obtained from all volunteers. Fifteen healthy volunteers (seven women, eight men; mean age = 33.0 ± 3.6 years) without any history of pulmonary or cardiac disease participated in the study.

Image-acquisition protocol

All measurements were performed using a 1.5-T MR scanner (Siemens Magnetom Avanto, Siemens Healthineers, Erlangen, Germany) in the supine position. A six-element surface coil and a 24-element spine coil integrated into the patient table were employed.

A transverse, T2-weighted half-Fourier acquired single shot turbo spin echo (HASTE) sequence (18 sections, section thickness = 8 mm, field of view [FOV] = 360 × 270 mm, TR/TE = 1020/91 ms, matrix = 256 × 256, parallel imaging factor = 2, and acquisition time = 46 s) was performed on all volunteers for anatomical imaging.

For FD-MRI, a single, coronal two-dimensional (2D) true fast imaging with steady-state free precession (TRUFI) image was acquired first, followed by a coronal 2D TRUFI sequence (1 section, section thickness = 10 mm, FOV = 400 × 400 mm, TR/TE = 2.06/0.89 ms, matrix = 128 × 128, parallel imaging factor = 2, images = 250, total acquisition time = 45 s).

Image registration

To maintain the equilibrium state in the TRUFI sequence, the first 10 images of the dataset were excluded from further evaluation (2,3).

An image in the breathing agent position was chosen as a reference image for all analyzed image-registration algorithms. The three image-registration algorithms analyzed in the present study are fMRLung 3.0, ANTs, and Elastix. The settings for each image-registration algorithm were determined in pre-tests to give the best results.

fMRLung 3.0 (Siemens Corporate Research, Princeton, NJ, USA) is a prototype image-registration software based on an image-registration algorithm by Chefd’hotel (5,6). The settings used in the present study were: number of iterations per pyramid level = 16, 16, 8, 8; strength of regularization = 6.0. Employment of this program for image registration in pulmonary FD-MRI has been frequently described in the literature (2,3,7 –9).

ANTs (Penn Image Computing and Science Lab [PICSL]) is an open-source software based on the InsightToolKit library (ITK) (10). In the present study, symmetric-diffeomorphic registration was performed by ANTs with following settings: image dimensions = 2; update field variance in voxel space (weight) = 1; metric radius = 9; number of iterations per level = 50 × 25 × 25; smoothing = Gaussian $[3, 1]$ . Several studies dealing with pulmonary FD-MRI have utilized ANTs for image registration (11 –16).

Elastix (a toolbox for rigid and nonrigid registration of images, University Medical Center Utrecht and contributors, The Netherlands) is an open-source ITK-based software (17,18). The settings for Elastix in the present study were: image metrics, normalized mutual information; maximum number of iterations per level = 2000, 3 levels; transformation model, B-spline; optimizer, adaptive-stochastic gradient descent. Elastix, unlike the other two image-registration algorithms, is not widely used for image-registration in FD-MRI.

To analyze the required calculation time for image registration with the three algorithms, the post-processing time for the image registration was analyzed for all 15 datasets.

Calculation of lung perfusion and ventilation

After image registration, the perfusion and ventilation parameter images of all three registered datasets were calculated using the FD method to determine the impact of image registration on pulmonary FD-MRI outcome (2). For further analysis, lungs were segmented manually.

The fractional ventilation was calculated using the following formula:

V_{F D} = \frac{2 {F D}_{V}}{{F D}_{0} + {F D}_{V}}

where FD_V is the ventilation parameter image after Fourier transformation and FD₀ is the zero-frequency component (19).

The fractional perfusion was obtained as follows:

f = \frac{S^{image, pf} (T_{\exp})}{S^{image, ef} (T_{\exp})} * \frac{1}{2 T_{\exp}}

where

S^{image, p f}

is the signal intensity of a voxel partially filled with blood,

S^{image, e f}

is the signal intensity of a voxel completely filled with blood (voxel filling factor (19,20)), T_exp is the time between two heartbeats derived from the heart rate (Texp = 0.90 ± 0.11 s).

To facilitate a comparison of the different lung volumes of the individuals, they were normalized to mL/min/100 mL. Finally, the V/Q ratio (3) was calculated for each participant.

Statistical analysis

For the evaluation of the image registration quality, an in-house developed MATLAB R2015b script (Mathworks, Natick, NA, USA) was used.

First, the diaphragm movement was estimated using Canny edge detection algorithm (ED) (21), which measures changes in brightness of two adjacent pixels.

Second, motion-induced signal intensity changes across multiple image voxels were measured by quotient image (QI) that is defined as a ratio of a reference image and a test object (22,23). Thus, the greater structural similarity between the images, the closer QI is to 1 (24). The QI analysis was performed over the entire dataset.

Third, the dice similarity coefficient (DSC) was determined to quantitatively estimate the overlap between the areas/volumes after global registration (15,25 –27). The DSC is a commonly used metric to evaluate the image registration accuracy and has already been successfully applied to the FD-MRI data (15,26,27).

In order to assess the level of motion artifacts, all the above-mentioned quality metrics were applied to the unregistered images.

Furthermore, the perfusion and ventilation values for both sides of the lung separately and the entire lung were compared for all three registered datasets. The normal distribution of all data collected was tested using the Kolmogorov–Smirnov test. Subsequently, all data were compared using Student’s t-test. P < 0.05 was considered to indicate a statistically significant difference.

Results

The average computing time differed significantly (P < 0.05) between the image-registration algorithms investigated in the present study. Overall, fMRLung showed the highest computational efficiency, whereas Elastix showed the lowest efficiency (Table 1).

Table 1.

The average computing times for all the three image-registration algorithms.*

	Computing time (min)
fMRLung	1.0 ± 1.6
ANTs	38.0 ± 13.5
Elastix	354 ± 78

Values are given as mean ± SD.

*The optimal quality of image registration by the investigated image-registration algorithms led to significant different computing times (P < 0.05). fMRLung showed the highest computational efficiency, whereas Elastix showed the lowest efficiency.

All three image-registration algorithms resulted in a significant reduction of the diaphragmatic motion compared to the uncorrected image as measured by ED (P < 0.05) (Fig. 1). Furthermore, all the considered image-registration algorithms showed similar image-registration qualities as all values showed no significant differences (P > 0.05) (Table 2). Nevertheless, residual motion of the diaphragm is still visible in the ANTs-based dataset (Fig. 1). All three image-registration algorithms significantly reduced motion artifacts in comparison to the unregistered images (P < 0.05) (Fig. 1).

Fig. 1.

Line diagram of the diaphragmatic movement of a healthy 31-year-old individual with and without motion correction. On the left, the positioning of the measurement cursor is shown for all four measurements. On the right, the diaphragmatic movement is shown without motion correction and with the three image-registration algorithms, respectively. All three image-registration algorithms result in a significant reduction of diaphragmatic motion compared to the uncorrected image (P < 0.05). There was no significant difference in the quality of motion correction among the three image-registration algorithms. However, residual motion of the diaphragm is still slightly visible in the ANTs-based dataset.

Table 2.

QI and DSC results for all three image-registration algorithms and for the dataset without motion correction.*

	QI	DSC
fMRLung	0.98 ± 0.03	0.92 ± 0.04
ANTs	0.97 ± 0.03	0.91 ± 0.04
Elastix	0.96 ± 0.03	0.93 ± 0.04

Values are given as mean ± SD.

*All three image-registration algorithms provide a significant improvement of image quality in comparison to the unregistered dataset (P < 0.05). However, there is no significant difference in the quality of the motion correction between the three image-registration algorithms.

DSC, dice similarity coefficient; QI, quotient image.

The heart (f_H) and respiratory rate (f_R) of the individuals in the present study were 1.13 ± 0.12 Hz (range = 0.86–1.36 Hz) and 0.26 ± 0.06 Hz (range = 0.16–0.35 Hz), respectively.

Calculated pulmonary ventilation and perfusion parameters are shown separately for each image-registration algorithm in Table 3. Fig. 2 shows an example of the ventilation, perfusion, and V/Q ratio parameter maps of a single participant.

Table 3.

Ventilation and perfusion values as well as the V/Q ratio for the left, right, and entire lung for all three image-registration algorithms and averaged over all participants.*

	fMRLung	ANTs	Elastix
Ventilation (mL/min/100 mL)
Right	115.0 ± 9.4 (100.5–132.3)	108.5 ± 12.3 (92.6–129.3)	114.6 ± 9.3 (96.5–131.8)
Left	117.3 ± 11.3 (94.6–135.8)	109.9 ± 12.4 (84.7–130.0)	117.3 ± 9.9 (99.1–133.4)
Total	116.1 ± 8.6 (103.9–132.6)	109.2 ± 10.8 (96.5–129.7)	115.9 ± 8.4 (102.0–132.6)
Perfusion (mL/min/100 mL)
Right	121.1 ± 25.4 (86.2–166.8)	127.1 ± 27.7 (80.1–174.5)	138.5 ± 46.4 (91.8–276.3)
Left	164.4 ± 38.9 (107.1–235.7)	177.0 ± 41.2 (119.8–246.5)	202.4 ± 57.7 (126.7–325.6)
Total	142.7 ± 31.0 (98.0–198.7)	152.0 ± 32.8 (105.1–209.6)	170.5 ± 50.6 (111.8–301.0)
V/Q
Right	0.98 ± 0.1 (0.79–1.21)	0.86 ± 0.2 (0.47–1.16)	0.89 ± 0.2 (0.48–1.22)
Left	0.76 ± 0.2 (0.4–1.09)	0.63 ± 0.2 (0.31–0.88)	0.63 ± 0.2 (0.37–0.95)
Total	0.74 ± 0.2 (0.39–0.95)	0.87 ± 0.2 (0.59–1.11)	0.76 ± 0.2 (0.44–1.06)

Values are given as mean ± SD (range).

*A significant difference was found only for the comparison of the fMRLung- and Elastix-based perfusion values of the left lung (P = 0.04) and the comparison of the ANTs- and fMRLung-based V/Q ratio of the right and the entire lung (P = 0.04, respectively).

V/Q, ventilation/perfusion.

Fig. 2.

Example of ventilation (in mL/min/100 mL), perfusion (in mL/min/100 mL), and V/Q maps (×10⁻³) of a healthy 31-year-old volunteer separated according to the three image-registration algorithms. V/Q, ventilation/perfusion.

Only fMRLung- and Elastix-based perfusion parameters in the left lung showed a significant difference (P < 0.05) (Table 3, Fig. 3). All other perfusion values resulting from the three registered datasets were not significantly different (P > 0.05). There was no significant difference in the calculated ventilation parameters based on the three registered datasets in the present study (P > 0.05).

Fig. 3.

Box plots showing the ventilation, perfusion, and V/Q values of all image-registration algorithms. A significant difference was found only for the comparison of the fMRLung- and Elastix-based perfusion values of the left lung (P = 0.04) and the comparison of the fMRLung- and ANTs-based V/Q ratio of the right and the entire lung (P = 0.04). All other functional lung values were not significantly different (P > 0.05). V/Q, ventilation/perfusion.

The calculated V/Q ratio showed a significant difference for fMRLung- and ANTs-based calculations in the right lung and those averaged over the entire lung (P < 0.05). All other calculated V/Q ratios did not differ significantly (P > 0.05).

Discussion

FD-MRI is a novel acquisition approach for measuring pulmonary perfusion and ventilation during free-breathing without the use of contrast agents. Image registration is mandatory for pulmonary FD-MRI.

Various image registration algorithms for pulmonary FD-MRI have been described in the existing literature (2,3,9,11,12,14,19,28). In the current study, the performance of three different image-registration algorithms for FD-MRI was analyzed.

Among all investigated image-registration algorithms, fMRLung showed the lowest computing time. However, fMRLung is a research black-box software without access to the source code, which only allows a limited selection of parameters and is not freely available. In contrast, the other two image-registration algorithms, ANTs and Elastix, are open source. Both packages contain a wide range of registration algorithms.

Only few FD-MRI studies give insight into the settings, which were used for image registration (16). In the present study, the settings for all three image-registration algorithms were optimized to achieve the best quality of image registration. In the case of ANTs and Elastix, the improvement of the accuracy came at the cost of substantially higher computing time.

The similarity metrics used in the present study (ED, QI and DSC) to quantitatively measure the effectiveness of image registration showed no significant differences between the investigated algorithms (P > 0.05). However, residual motion of the diaphragm was visible in the ANTs-based datasets (Fig. 1). This slightly motion was not detected by the similarity criteria used. A further analysis of the performance of existing methodological approaches for ANTs and Elastix in FD-MRI is necessary to enable both registration algorithms to achieve similar performance to fMRLung, especially in regard to the computing time.

A systematic evaluation of the post-processing tools for the segmentation (15) and retrospective motion correction of the lungs is an important step towards a standardized post-processing pipeline for FD-MRI, allowing comparison of the results obtained in different sites.

Next to ED, QI, and DSC, there are other objective criteria for the assessment of the image-registration quality, such as root-mean-squared-error (RMSE) and DVARS (29,30). In the present study, however, DSC was used as a criterion, as it has previously been used for the registration of the FD-MRI data (27). The two other metrics, ED and QI, have also been applied as quantitative measures of the motion-correlated signal changes (24,30,31).

The impact of the choice of image-registration algorithm on the calculated quantitative ventilation and perfusion parameters was investigated in the current study as well. Only in three cases was there a significant difference between the calculated parameters (fMRLung- and ANTs-based V/Q ratio of the right and the entire lung, P = 0.04; fMRLung- and Elastix-based perfusion of the left lung, P = 0.04).

All other parameters were not significantly different and thus independent of the image-registration algorithm used. In addition, the value ranges of the estimated parameters showed an overlap regardless of whether the differences between them were significant.

The present study has some limitations. The main limitation is that only images in the coronal plane were acquired. Since the effectiveness of the image registration might depend on the imaging plane, acquisition in multiple planes (e.g. axial and coronal) should be considered in future studies. Further, manual segmentation that might be biased and time-consuming was performed in the study. Automatic lung segmentation could be a useful alternative. However, Guo et al. (15) recently showed that there is no significant difference in the performance of manual segmentation and automatic segmentation.

In conclusion, the present study demonstrated that it is possible to perform FD-MRI mandatory image registration with different image-registration algorithms and obtain consistent quality. A further systematic optimization of the registration parameters in order to reduce the computing time would be desirable.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Alexandra Ljimani

References

Jensen

Teeter

England

, et al. Sources of long-term variability in measurements of lung function: implications for interpretation and clinical trial design. Chest 2007; 132:396–402.

Bauman

Puderbach

Deimling

, et al. Non-contrast-enhanced perfusion and ventilation assessment of the human lung by means of fourier decomposition in proton MRI. Magn Reson Med 2009; 62:656–664.

Kjørstad

Corteville

DMR

Henzler

, et al. Non-invasive quantitative pulmonary V/Q imaging using Fourier decomposition MRI at 1.5T. Z Med Phys 2015; 25:326–332.

Lujan

Larsen

Balter

, et al. A method for incorporating organ motion due to breathing into 3D dose calculations. Med Phys 1999; 26:715–720.

Hermosillo

Chefd’Hotel

Faugeras

A Variational Approach to Multi-Modal Image Matching. RR-4117. Sophia Antipolis: INRIA, 2001. Available at: https://hal.inria.fr/inria-00072513/document (accessed 23 July 2017).

Chefd’hotel

Hermosillo

Faugeras

Flows of diffeomorphisms for multimodal image registration. In: Proceedings IEEE International Symposium on Biomedical Imaging. Piscataway, NJ: IEEE, 2002:753–756.

Zapke

Topf

H-G

Zenker

, et al. Magnetic resonance lung function – a breakthrough for lung imaging and functional assessment? A phantom study and clinical trial. Respir Res 2006; 7:106.

Wujcicki

Corteville

Materka

, et al. Perfusion and ventilation filters for Fourier-decomposition MR lung imaging. Z Für Med Phys 2015; 25:66–76.

Bondesson

Schneider

Gaass

, et al. Nonuniform Fourier-decomposition MRI for ventilation- and perfusion-weighted imaging of the lung. Magn Reson Med 2019; 82:1312–1321.

10.

Avants

Tustison

Song

, et al. A reproducible evaluation of ANTs similarity metric performance in brain image registration. NeuroImage 2011; 54:2033–2044.

11.

Behrendt

Voskrebenzev

Klimeš

, et al. Validation of Automated Perfusion-Weighted Phase-Resolved Functional Lung (PREFUL)-MRI in Patients With Pulmonary Diseases. J Magn Reson Imaging 2020; 52:103–114.

12.

Voskrebenzev

Gutberlet

Becker

, et al. Reproducibility of fractional ventilation derived by Fourier decomposition after adjusting for tidal volume with and without an MRI compatible spirometer. Magn Reson Med 2016; 76:1542–1550.

13.

Schönfeld

Cebotari

Voskrebenzev

, et al. Performance of perfusion-weighted Fourier decomposition MRI for detection of chronic pulmonary emboli. J Magn Reson Imaging 2015; 42:72–79.

14.

Voskrebenzev

Greer

Gutberlet

, et al. Detection of chronic lung allograft dysfunction using ventilation-weighted Fourier decomposition MRI. Am J Transplant 2018; 18:2050–2060.

15.

Guo

Capaldi

DPI

McCormack

, et al. A framework for Fourier-decomposition free-breathing pulmonary 1 H MRI ventilation measurements. Magn Reson Med 2019; 81:2135–2146.

16.

Voskrebenzev

Gutberlet

Kaireit

, et al. Low-pass imaging of dynamic acquisitions (LIDA) with a group-oriented registration (GOREG) for proton MR imaging of lung ventilation. Magn Reson Med 2017; 78:1496–1505.

17.

Klein

Staring

Murphy

, et al. elastix: A Toolbox for Intensity-Based Medical Image Registration. IEEE Trans Med Imaging 2010; 29:196–205.

18.

Shamonin

Bron

Lelieveldt

BPF

, et al. Fast parallel image registration on CPU and GPU for diagnostic classification of Alzheimer’s disease. Front Neuroinform 2014; 7:50.

19.

Kjørstad

Corteville

DMR

Henzler

, et al. Quantitative lung ventilation using Fourier decomposition MRI; comparison and initial study. Magn Reson Mater Phys Biol Med 2014; 27:467–476.

20.

Fischer

Pracht

Arnold

JFT

, et al. Assessment of pulmonary perfusion in a single shot using SEEPAGE. J Magn Reson Imaging 2008; 27:63–70.

21.

Canny

A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell 1986; 8:679–698.

22.

Shashua

Riklin-Raviv

The quotient image: class-based re-rendering and recognition with varying illuminations. IEEE Trans Pattern Anal Mach Intell 2001; 23:129–139.

23.

Riklin-Raviv

Shashua

The quotient image: Class based recognition and synthesis under varying illumination conditions. In: Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149). Vol. 2. Piscataway, NJ: IEEE, 1999:566–571.

24.

Wang

Generalized quotient image. IEEE CVPR 2004; 2:498–505.

25.

Dice

LR.

Measures of the Amount of Ecologic Association Between Species. Ecology 1945; 26:297–302.

26.

Zou

Warfield

Bharatha

, et al. Statistical Validation of Image Segmentation Quality Based on a Spatial Overlap Index. Acad Radiol 2004; 11:178–189.

27.

Guo

Yuan

Rajchl

, et al. Globally optimal co-segmentation of three-dimensional pulmonary ¹H and hyperpolarized 3He MRI with spatial consistence prior. Med Image Anal 2015; 23:43–55.

28.

Bauman

Puderbach

Heimann

, et al. Validation of Fourier decomposition MRI with dynamic contrast-enhanced MRI using visual and automated scoring of pulmonary perfusion in young cystic fibrosis patients. Eur J Radiol 2013; 82:2371–2377.

29.

Power

Mitra

Laumann

, et al. Methods to detect, characterize, and remove motion artifact in resting state fMRI. NeuroImage 2014; 84:320–341.

30.

Klein

Andersson

Ardekani

, et al. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage 2009; 46:786–802.

31.

Wang

Bovik

Sheikh

, et al. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 2004; 13:600–612.