Towards electrocardiogram-free diffusion tensor cardiac magnetic resonance using a machine learning approach (RF_TH_219)

Thursday, January 30, 2025

3:00 PM – 3:10 PM East Coast USA Time

Location: Blue Room PreFunction

Presenting Author(s)

DH

David Han, BSc

Medical Student
Case Western Reserve University School of Medicine, Cardiovascular Innovation Research Center

Primary Author(s)

DH

David Han, BSc

Medical Student
Case Western Reserve University School of Medicine, Cardiovascular Innovation Research Center

Co-Author(s)

MC

Moses P. Cook, BSc

Doctoral Candidate
Sunnybrook Research Institute - University of Toronto, Canada

Co-Author(s)

MC

Moses P. Cook, BSc

Doctoral Candidate
Sunnybrook Research Institute - University of Toronto, Canada
DK

Danielle Kara, PhD

Staff Scientist
Cleveland Clinic
DM

Dingheng Mai

PhD Candidate
Cleveland Clinic
YL

Yuchi Liu, PhD

Research Scientist
Siemens Medical Solutions USA, Inc.
PS

Pasquale Santangeli, MD, PhD

Director
Ventricular Tachycardia (VT) program in the Cardiac Electrophysiology and Pacing Section of the Tomsich Family Department of Cardiovascular Medicine, Miller Family Heart, Vascular & Thoracic Institute, Cleveland Clinic
OW

Oussama Wazni, MD, MBA

Physician
Cardiac Electrophysiology and Pacing at Sydell and Arnold Miller Family Heart, Vascular & Thoracic Institute, Cleveland Clinic
Deborah Kwon, MD, FSCMR

Director of Cardiac MRI
Cleveland Clinic
Christopher Nguyen, PhD, FSCMR, FACC

Director, Cardiovascular Innovation Research Center
Cleveland Clinic

Background: Diffusion tensor cardiac magnetic resonance (DT-CMR) can quantify tissue microstructure, such as cardiac tissue anisotropy and the heart’s 3D double helix structure. However, DT-CMR is logistically complicated to set up, which has limited efforts towards clinical translation. DT-CMR requires electrocardiogram (ECG) triggering, which is unreliable in patients with premature left ventricular (LV) contraction or other arrhythmias that may cause ECG gating irregularities. It also requires breath holds, which are difficult for heart failure (HF) patients. Inconsistent breath holds and ECG data may cause image artifacts and spurious results. Thus, there is a need for a free-breathing, non-triggered (ungated) DT-CMR technique to improve comfort and expand access to tissue microstructure characterization to patients with HF or arrhythmias.

Methods: Reference DT-CMR¹ images were scanned using the following parameters: TR/TE = 6000/67 ms; b-value = 500 s/mm²; diffusion directions = 12; diffusion-weighted image (DWI) averages = 8, resolution = 2.7x2.7x8 mm³; slices = 12. An ungated DT-CMR sequence was adapted from the reference standard, except for TR = 2700 while scanning continuously for 30 minutes. LV segmentation of the ungated images was performed using a U-Net model, which was trained using PyTorch on ungated DT-CMR images from human subjects (Figure 1).
Four healthy volunteers (age 23 ± 1.225, 3 males) were prospectively scanned under an IRB approved protocol at 3T (MAGNETOM Cima.X, Siemens Healthineers, Forchheim, Germany). Each volunteer underwent the breath-held, ECG-gated (gated) DT-CMR protocol followed by the ungated DT-CMR protocol. The ungated DT-CMR images were segmented by the U-Net² model to characterize cardiac motion. The mode of the area for each LV segmentation was used to derive the b0 and DWIs accepted for tensor computation. DWIs were then registered to the b0 images. These images were used to compute the mean diffusivity (MD), fractional anisotropy (FA), and helix angle (HA). Average values were compared to those from the reference gated DT-CMR protocol.

Results: After training, the U-Net model achieved a Dice score of 0.85 on a test set of images that were kept separate from the training set. Images for which the model created masks generally had good quality DT parameter maps (Figure 2). A Wilcoxon signed-rank test demonstrated no significant differences between MD (p=0.62) nor FA (p=0.25). Bland-Altman plots showed a positive bias of 0.06 x 10^-3mm²/s for MD and a positive bias of 0.03 for FA (Figure 3).

Conclusion:

Herein, the feasibility of an ungated DT-CMR protocol is presented. Future optimization of the deep-learning model and thresholding algorithm will help improve scan time and image quality for clinical applications.