CMR-Analysis (including machine learning)

1357112 - Data-Driven Native Mapping for Amplifying Native-T1-based Image Contrast to LGE Levels

Wednesday, January 25, 2023

Location: E-POSTER

Presenting Author(s)

KY

Khalid Youssef, PhD

Senior Scientist
Indiana University School of Medicine
Plainfield, Indiana, United States

Primary Author(s)

KY

Khalid Youssef, PhD

Senior Scientist
Indiana University School of Medicine
Plainfield, Indiana, United States

Co-Author(s)

Xinheng Zhang, MSc

Ph.D. Candidate
Indiana University School of Medicine
Indianapolis, Indiana, United States
GY

Ghazal Yoosefian

Student
Indiana University School of Medicine, United States
Yinyin Chen, MD

Attending Radiologist
Zhongshan Hospital, Fudan University and Shanghai Institute of Medical Imaging, Shanghai, China
LOS ANGELES, California, China (People's Republic)
SC

Shing Fai Chan, PhD

Research Assistant Professor
Krannert Cardiovascular Research Center, Indiana University, School of Medicine/IU Health Cardiovascular Institute
Indianapolis, Indiana, United States
HY

Hsin-Jung Yang, PhD

Assistant Professor
Cedars-Sinai Medical Center
Los Angeles, California, United States
KV

Keyur Vora, MD, MSc, FACC

Assistant Professor of Medicine, Division of Cardiology
Krannert Cardiovascular Research Center, Indiana University School of Medicine
Carmel, Indiana, United States
AH

Andrew G. Howarth, MD, PhD

Clinical Co-director
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
AK

Andreas Kumar, MD

doctor
Northern Ontario School of Medicine, Sudbury,Ontario, Canada
Sudbury, Ontario, Canada
Behzad Sharif, PhD

Associate Professor of Medicine
Indiana University School of Medicine
Indianapolis, Indiana, United States
Rohan Dharmakumar, PhD

Professor of Medicine, Radiology & Imaging Sciences, Anatomy, Cell Biology & Physiology
Krannert Cardiovascular Research Center, Indiana University School of Medicine
Indianapolis, Indiana, United States

Disclosure(s):

Behzad Sharif, PhD: No financial relationships to disclose

Background: Late gadolinium enhancement (LGE) cardiac MRI (CMR) is currently the standard for the detection and characterization of chronic myocardial infarction (MI). Despite its capabilities in identifying infarct presence and characteristics, LGE CMR has a key limitation – it requires the use of gadolinium-based contrast agents, which is contraindicated in patients with late-stage chronic kidney disease (CKD). Given the high comorbidity of CKD in patients with heart disease, LGE CMR is estimated to be contraindicated in millions of patients. Late-stage CKD patients who cannot undergo LGE CMR are in great need of chronic MI assessment since the risk of cardiac mortality is substantially elevated in this population. Recent studies have shown that MI territories often appear hyperintense on native T1 maps. However, the presence of fat or iron within the MI zone can compromise image contrast and impede the visual detection of infarcted regions. To overcome the existing limitations, we developed and tested a data driven mapping model based on deep neural networks as an alternative to generic fitting algorithms.

Methods: We conducted a pilot study with a dataset of highly aligned native-T1-weighted and LGE images using a large animal model of chronic MI with histological evidence of varied tissue composition within the MI territories. We trained a deep neural network model to map pixels from native-T1-weighted image series to corresponding pixels in LGE images to produce a visually enhanced native-T1-mapping alternative that we refer to as data-driven native mapping (DNM). We tested the performance of DNM against native T1 mapping, using LGE as the ground truth based on semi-automated quantification performed by three CMR readers.

Results: Our results showed that DNM led to significant improvement in CNR over native T1 mapping between chronic MI and remote territories: 15.01 ± 2.88 (DNM) vs 5.64 ± 1.58 (native T1 maps), p< 0.05. CNR of LGE was 15.51 ± 2.43, which was significantly different from native T1 maps but not different from DNM (p=0.4). Moreover, the infarct areas as determined using LGE were strongly correlated with DNM compared to native T1 maps: (R²: 0.71 for native-T1 maps vs. LGE and 0.85 for DNM vs. LGE, p < 10^-5).

Conclusion:

Pixel-level information from native-T1-weighted images carry information that can be utilized by data-driven mapping models to maximize image contrast between chronic MI and remote territories and significantly increase CNR levels for enhanced visualization of chronic MI territories. While the proposed data-driven native modeling approach shows much promise as a viable non-contrast alterative to LGE for characterization of chronic MI territories, additional studies employing clinical data are necessary for clinical implementation.