A robust motion correction tool for cardiac extracellular volume mapping
Shufang Liu1,2,3, Lin Zhang3,4, Pauline Ferry3,4, Andrei Codreanu5, Anne Menini2, and Freddy Odille3,4

1Institut für Informatik, Technology University of Munich, Munich, Germany, 2GE Global Research, Munich, Germany, 3Imagerie Adaptative Diagnostique et Interventionnelle, Université de Lorraine, Nancy, France, 4U947, INSERM, Nancy, France, 5Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg

Synopsis

This work present a robust motion correction framework for T1 mapping and ECV mapping. Motion correction within one series and between different series are discussed. Validation is performed on 2 patient data and 6 volunteer data.

Background

Extracellular volume (ECV) can provide a quantitative assessment of the myocardial fibrosis. To obtain the ECV map, T1 weighted series before and after contrast agent injection need to be acquired and pixel-wise curve fitting is used to get the T1 maps. Despite the patients are instructed to hold their breath, motion can still be observed both within one T1 series (termed intra-series motion) and between the pre- and post- contrast series (termed inter-series motion). Nonrigid image registration can be used to correct the inter-image motion [1,2]. Despite these advances, robust registration in both inversion recovery (IR) and saturation recovery (SR) techniques remains challenging due to the drastic contrast changes and due to the large motion between pre- and post- contrast images (typically acquired at 10-15 min interval). In this work we propose a post-processing framework with minimum user input for robust correction of both intra- and inter-series motion, based on non-rigid registration. Improved robustness is obtained by a preprocessing the images to minimize contrast changes and landmark tracking is used in order to deal with large deformations.

Methods

Intra-series correction: First images are preprocessing including a) removing large contrast changes by subtraction of a low frequency version of the images, b) histogram matching. And then the sum-of-squared-differences (SSD) metric is minimized to extract the motion field [3]. Additionally, a predictive motion model is investigated to further improve the registration results. In this method, a region of interest (ROI) containing the heart is detected by thresholding a reduced resolution T1 map from the raw data. Registration in this ROI is likely to fail in frames near the nulling point of the IR or SR curve. Motion inside this ROI is assumed to be correlated with motion outside the ROI. A statistical model is therefore constructed by partial least-squares regression (PLSR), using frames far away from the nulling point (training frames). Then in the other frames (prediction frames), motion inside the ROI is predicted from the outer region using the learnt statistical model (Figure 1).

Inter-series correction: To correct the inter-series motion, several landmarks are picked on the pre-contrast images by the user. The corresponding points on the post-contrast images are detected automatically by block matching methods. Then the global non-rigid motion is estimated by interpolating the motion from the landmarks. Finally the motion is refined using non-rigid registration.

MRI data: Preliminary results from 2 patients and 6 volunteers with both pre- and post- contrast images (GE 3T MR scanners) are processed. For each volunteer, both IR (MOLLI) and SR (SMART1map) single-shot sequences are acquired during repeated breath-holds. Images before motion correction were manually segmented by a cardiologist. DICE overlap measures [2] and Myocardial Boundary Error (MBE) in pixel units were computed to quantify the improvement after motion correction.

Result

Figure 2 shows the DICE score and MBE of each series before and after the intra-series registration step. When the prediction step was turned off, the average DICE score was improved from 0.969±0.014 to 0.974±0.009 (p = 0.0694), and the MBE decreased from 0.917±0.372 to 0.746±0.233 (p = 0.0133) for overall series. Results were similar with the predictive model on, with DICE scores of 0.975±0.008 and MBE of 0.741±0.219. However the differences between results obtained with and without the predictive motion model were not statistically significant. Figure 3 illustrate the registration result between pre- and post- contrast images of a patient with anteroseptal infarction by computing the ECV map. Before registration, the ECV map is corrupted by the inter-series motion, after motion registration, the infarction area can be clearly observed. Figure 3c gives a quantitative measurement of the DICE coefficient for the inter-series registration from all the datasets including SMART1map and MOLLI from both patients and volunteers, the DICE score increased from 0.88±0.09 to 0.96±0.02 (p=0.001).

Conclusion

The proposed post-processing framework can be used to calculate ECV maps in T1 mapping clinical studies based on either IR or SR techniques. More patient datasets will be necessary in order to conclude whether the predictive model can bring a significant improvement to the software. In future work we will investigate an automatic selection of the landmarks.

Acknowledgements

No acknowledgement found.

References

[1] Xue, 2014,MRM; [2] Roujol, 2015, MRM; [3] Odille, 2014, JMRI

Figures

PLSR model for motion correction within one T1 series. The T1 series is divide as well registered and registration failed. A PLSR model is calculated from the well registered images and applied to registration failed

a) DICE score before and after registration. b) Myocardial Boundary Error before and after registration

a) T1 map of pre-contrast images_after intra-series registration; b) T1 map of post-contrast images after intra-series registration; c) statistic measure of DICE coefficient; d) ECV map without motion correction; e) ECV map with after motion correction; f) Delayed Enhancement image



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1862