Motion Compensation

1357146 - Improved self-navigation in XD-GRASP reconstruction of 3D left atrial LGE using exploratory factor analysis

Friday, January 27, 2023

10:20 AM – 11:00 AM

Location: Pacific Jewel

Presenting Author(s)

DB

Dima Saied Bishara

PhD student
Northwestern University
Evanston, Illinois, United States

Primary Author(s)

DB

Dima Saied Bishara

PhD student
Northwestern University
Evanston, Illinois, United States

Co-Author(s)

KyungPyo Hong, PhD

Research Associate
Northwestern University
Chicago, Illinois, United States
Suvai Gunasekaran, PhD

Postdoctoral Researcher
Northwestern University
Chicago, Illinois, United States
Mark Markl, PhD

Professor
Northwestern University
Daniel C. Lee, MD

Associate Professor of Medicine (Cardiology) and Radiology
Northwestern University
Chicago, Illinois, United States
Bradley D. Allen, MD, MSc, FSCMR

Assistant Professor, Cardiovascular and Thoracic Imaging
Northwestern University
Chicago, Illinois, United States
RP

Rod Passman, MD

Professor of Medicine (Cardiology) and Preventive Medicine
Northwestern University
Chicago, Illinois, United States
DK

Dan Kim, MD, MS

Cardiology Fellow
Loyola University Medical Center
Streamwood, Illinois, United States

Disclosure(s):

Daniel C. Lee, MD: No relevant disclosure to display

Bradley D. Allen, MD, MSc, FSCMR: No relevant disclosure to display

Background: Left atrial (LA) fibrosis can be quantified from late gadolinium enhancement (LGE) images. We recently developed a self-navigated 3D LA LGE pulse sequence using XD-GRASP [1] reconstruction [2]. In our 3D LGE using XD-GRASP, a 1D self-navigation k-space line is acquired along the superior inferior (SI) direction once per heartbeat to capture the respiratory motion, and repeated over 350 heartbeats. In conventional XD-GRASP, principal component analysis (PCA), which assumes correlated variables and no unique variance, is performed to extract the respiratory motion. Another approach is to use exploratory factory analysis (EFA) such as orthogonal rotation, which does not require a prior knowledge about the relationship between variables [3]. In this study, we sought to compare the performance between PCA that analyzes variance and EFA which analyzes covariance in 3D LA LGE with XD-GRASP reconstruction.

Methods:

This is a retrospective study using existing stack-of-stars raw k-space of 17 patients (3 females; mean age = 63.88 yrs) with atrial fibrillation. For details on imaging parameters, please see reference [2]. XD-GRASP reconstruction was performed on a GPU (2 x 32GB Tesla V100) computer using MATLAB.

Contrast adjusting and sharpening: As shown in Figure 1, to enhance the edge profiles of 1D SI projection signal, we performed contrast adjusting (‘imadjust’) for each coil image, then we sharpened the image using the unsharp masking method in MATLAB (‘imsharpen’).
PCA and Promax [4]: We then performed an orthogonal rotation (Promax) on the fully stacked coils data, and conventional PCA for comparison. We computed the blur metric (0 [best] to 1 [worst]) [5] on a cropped image encapsulating the heart region only. The variable normality was tested using the Shapiro-Wilk test. Paired t-test was used to compare the mean blur metric values.

Results:

Motion correction: As shown in Figure 1, a combination of image contrast adjustment, sharpening, and EFA produced the best respiratory signal extraction compared with other combinations. As shown in Figure 2, EFA with preprocessing produced better results than PCA (see both zero-filled NUFFT images).

Blur Metric: According to the Shapiro-Wilk test, the blur metric is normally distributed. The mean blur metric was significantly (p< 0.001) better for EFA (0.38 ± 0.04) than PCA (0.41 ± 0.05), where the relative difference was 7%.

Conclusion: This study shows that EFA extracts better self-navigation signal than PCA for XD-GRASP reconstruction. Future study includes comparison of repeatability of LA fibrosis quantification between PCA and EFA self-navigation signal extraction. Another study includes comparing PCA and EFA on other types of XD-GRASP data.