Groupwise Non Rigid Registration For Temporal Myocardial Arterial Spin Labeling Images
Veronica Aramendía-Vidaurreta1, Pedro Macías-Gordaliza2,3, Marta Vidorreta4, Rebeca Echeverria-Chasco1, Gorka Bastarrika1, Arrate Muñoz-Barrutia2,3, and María Fernández-Seara1

1Radiology, Clínica Universidad de Navarra, Pamplona, Spain, 2Universidad Carlos III de Madrid, Madrid, Spain, 3Instituto de Investigación Sanitaria Gregorio Marañón, Madrid, Spain, 4Siemens Healthineers, Madrid, Spain


Arterial Spin Labeling (ASL) enables quantitative measurement of myocardial blood flow (MBF) by averaging over multiple ASL pairs providing a voxelwise map in units of milliliters of blood per gram of tissue per minute (ml/g/min). However, its estimation accuracy in free breathing acquisitions depends critically on the quality of the image registration algorithm. In this work, a groupwise non-rigid registration method with a similarity measure based on Principal Component Analysis (PCA) was applied to ASL images of the heart acquired during free breathing. The method was compared against a pair-wise registration algorithm provided by the advanced normalization tools software (ANTs). Results demonstrate the feasibility of using PCA-groupwise for temporal ASL image registration.


In myocardial ASL, FAIR labeling is commonly used to acquire a temporal set of images by alternating selective (Label) and non-selective (Control) inversions, to label arterial blood in the coronary arteries. Prior to quantification, these images are pairwise subtracted and averaged to increase SNR. Typically, label and control images (which have different contrast, Fig1) are independently registered to a manually chosen reference image pair, incorporating bias into the procedure[1][2]. Recently, groupwise registration approaches, where all images are simultaneously aligned to a mean space, have been successfully employed in quantitative MRI[3]. The purpose of this project was to evaluate a groupwise registration technique in ASL images. This method was based on minimizing a Principal Component Analysis (PCA) metric, taking advantage of the fact that label-control intensity changes depend on a low-dimensional acquisition model[3].


Scanning protocol: A cardiac MRI study was performed in 11 subjects on a 3T Skyra. The scanning session included: localizers to obtain a mid-ventricular short-axis plane of the myocardium, a baseline image and 60 free-breathing FAIR-ASL images with bSSFP readout. (FOV:300x300mm; Matrix:96x96; Slice thickness:10mm; Flip Angle:70º; Grappa-2i; TI=1s; TR=4RR). Four saturation pulses were applied before the inversion pulse. Tagging and imaging were centered at mid-diastole using ECG-triggering.

PCA registration: Temporal ASL images were represented as data points in a matrix (Fig2), from where data correlation(K) was calculated[1]. When correctly coregistered, K should follow a particular intensity model, represented by temporally alternating low(label) and high(control) intensity values for ASL. Deviation from the expected model is due to motion. PCA is calculated on K and its cost function minimized to fit the expected model. This method was included within the Elastix framework to perform registration[4].

Data processing: Coregistered images were pair-wise subtracted and averaged. A circular ROI was manually drawn in the myocardium based on the mean ASL image. Regional MBF was estimated(Fig1).

Comparison of registration methods: No images were discarded to simplify comparison between registration algorithms (ANTs details:[2]). First, temporal SNR (tSNR), mean MBF values and MBF maps were computed. Then, parameter differences between registrations were explored through: (1) calculation of image intensity differences by subtracting only label images. If properly coregistered, difference signal will mainly come from the presence of noise. Therefore, absolute values were considered and divided by the baseline image for calibration across volunteers. Same procedure was repeated for control images; (2) registered images were smoothed before MBF quantification with a Gaussian filter of increasing standard deviation specified by sigma (12 values, from 0.5 to 6) in order to assess possible effects of the interpolation steps during registration.

Results and Discussion

Figure 1 shows a label-control image pair together with the MBF maps from both registration methods. PCA maps are visually more homogeneous.

Figure 3 shows the ASL mean boxplot per volunteer, mean tSNR and mean MBF values. Mean±SD across volunteers are: tSNR (PCA: 1.29±2.47; ANTs: 3.11±2.74) and MBF (PCA: 1.54±1.19; ANTs: 3.00±2.19). tSNR tends to be higher for ANTs suggesting a stronger smoothing effect of the algorithm. Typically in deformable registrations, interpolation smoothing is applied through a kernel to obtain the new intensity in the fixed space. Although both algorithms use linear interpolation, differences in kernel size and operating function could be the cause of the smoothing present in ANTs coregistered images, This artifactual increase of signal could explain why MBF values of three volunteers (2,6,7) exceed physiological literature range for ANTs but not for PCA [6].

Figure 4 shows the spurious difference signal obtained from subtracting and averaging control-label series independently. High motion is present in volunteer3 (not properly coregistered by ANTs). Label images mean difference signal is similar between algorithms (ANTs:0.0021 and PCA:0.0022). A paired two tails t-test (p-value=0.86) indicates similar noise levels in label images with both algorithms. Control images present higher values of spurious signal (ANTs:0.0089 and PCA:0.0046) showing the negative effects of misregistration and partial volume due to the high intensity of myocardial blood within the ventricles. A paired two tails t-test between registration methods was not significant (p-value=0.29).

Figure 5 shows increasing mean estimated MBF values from registered images after smoothing, confirming that interpolation smoothing during registration considerably affects MBF values. For two volunteers (4 and 8), higher resolution scans would have been needed to avoid partial volume artifacts due to their thin myocardium.


This work demonstrates the feasibility of using PCA-groupwise for temporal ASL image registration. Interpolation smoothness during registration may affect MBF quantification, especially in low resolution images of myocardium, due to the presence of voxels with high intensity signal from the ventricular blood pool. Although no statistically significant differences between registration methods were found, MBF values computed after PCA coregistration are closer to physiological values for some subjects.


This work has been supported by Asociación de Amigos de la Universidad de Navarra and the projects TEC2015-73064-EXP and TEC2016-78052-R from the Spanish Ministry of Economy, Industry and Competitiveness.


[1] A. Javed, T. R. Jao, and K. S. Nayak, “Motion correction facilitates the automation of cardiac ASL perfusion imaging,” J. Cardiovasc. Magn. Reson., vol. 17, no. Suppl 1, pp. 1–2, 2015.

[2] V. A. Vidaurreta, A. G. Osés, G. Bastarrika, and M. A. F. Seara, “OPTIMAL REPETITION TIME FOR MYOCARDIAL ARTERIAL SPIN LABELING IN HUMANS,” in ESMRMB, 2017.

[3] W. Huizinga et al., “PCA-based groupwise image registration for quantitative MRI,” Med. Image Anal., vol. 29, pp. 65–78, 2016.

[4] S. Klein, M. Staring, K. Murphy, M. A. Viergever, and J. P. W. Pluim, “Elastix: A toolbox for intensity-based medical image registration,” IEEE Trans. Med. Imaging, vol. 29, no. 1, pp. 196–205, 2010.

[5] Z. Zun, E. C. Wong, and K. S. Nayak, “Assessment of myocardial blood flow (MBF) in humans using arterial spin labeling (ASL): feasibility and noise analysis.,” Magn. Reson. Med., vol. 62, no. 4, pp. 975–983, 2009.

[6] P. Chareonthaitawee et al., “Heterogeneity of resting and hyperemic myocardial blood flow in healthy humans,” Cardiovasc. Res., vol. 50, pp. 151–161, 2001.


Figure 1. (Top) A label and control image, followed by the MBF equation used for quantification [5], where C, T and B represent the mean myocardial signal in the ROI of the control, tagged and baseline images; TI: inversion time (1s); T1: T1 of arterial blood (1.644s 3T);

(Bottom) MBP map for PCA and ANTs registrations for a representative volunteer (Vol 9). MBF maps have been overlapped to the mean ASL image. Values represent ml/g/min units (values < 0.0 appear dark blue, values > 3.0 do not appear).

Figure 2. PCA groupwise metric used for simultaneously non rigid registration of ASL images. The 60 temporal ASL images, which follow an alternating low-label and high-control intensity model, were transformed into a new matrix (M) whose columns represent the images (G) and rows each of the voxels within the image. Correlation matrix (K) and dissimilarity metric (D) were used as defined by Huizinga et al [1].

Figure 3: Summary of mean ASL, temporal SNR (tSNR) and MBF results. All parameters have been calculated taking into account the manual myocardial ROI. (Top) boxplot of the mean ASL values obtained for all registered ASL pairs; (Middle) tSNR calculated as the temporal mean ASL divided by the temporal standard deviation. (Bottom) MBF values for 11 volunteers obtained for ANTs and PCA registration methods. Red lines represent the maximum and minimum intervals of physiological literature range[5].

Figure 4: Image intensity differences obtained by subtracting only registered controls (top) and only registered labels (bottom) for both registration methods. Absolute difference values were divided by the baseline image intensity for calibration across volunteers.

Figure 5: Estimated mean MBF values represented as a function of the Gaussian filter applied to the registered images with increasing standard deviation (specified by sigma) for both registration methods. Red lines represent the MBF physiological interval for healthy subjects at rest [4]. Sigma value of 0 represents the estimated MBF value right after registration. Neither of the algorithms applied a previous smoothing step before registration.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)