Machine Learning and Artificial Intelligence

1348227 - Scanner-Independent MyoMapNet for Accelerated Cardiac T1 Mapping Across Vendors and Field Strength

Thursday, January 26, 2023

3:50 PM – 4:00 PM

Location: Silver Pearl 2

Presenting Author(s)

AA

Amine Amyar, PhD

Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States

Primary Author(s)

AA

Amine Amyar, PhD

Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States

Co-Author(s)

Ahmed S. Fahmy, PhD

Instructor in Medicine
Harvard Medical School
Boston, Massachusetts, United States
RG

Rui Guo, PhD

Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Beijing, Massachusetts, United States
KN

Kei Nakata, MD

Clinical fellow
Mie University Hospital
Tsu, Mie, Japan
ES

Eiryu Sai, MD, PhD

Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Tokyo, Massachusetts, United States
JR

Jennifer Rodriguez

Clinical Trials Specialist
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
JC

Julia Cirillo, BSc

Research Assistant
Beth Israel Deaconess Medical Center, Massachusetts, United States
KP

Karishma Pareek, MD

Resident Physician
Boston University School of Medicine and Boston Medical Center, United States
Jiwon Kim, MD

Associate Professor of Medicine
Weill Cornell Medicine
New York, New York, United States
Robert M. Judd, PhD

Professor Emeritus
Duke University
Durham, North Carolina, United States
Frederick L. Ruberg, MD

Professor of Medicine and Radiology
Boston University School of Medicine and Boston Medical Center
Boston, Massachusetts, United States
Jonathan W. Weinsaft, MD

Professor of Medicine
Weill Cornell Medical College
New York, New York, United States
RN

Reza Nezafat, PhD

Professor
Harvard Medical School
Boston, Massachusetts, United States

Disclosure(s):

Jiwon Kim, MD: No financial relationships to disclose

Robert M. Judd, PhD: No relevant disclosure to display

Jonathan W. Weinsaft, MD: No relevant disclosure to display

Reza Nezafat, PhD: No financial relationships to disclose

Background:

In T₁ mapping, a series of T₁ weighted (T₁w) images are collected and numerically fitted to a 2 or 3-parameter model of the signal recovery to estimate voxel-wise T₁ values. To reduce the scan time, one can collect fewer T₁w images, albeit at the cost of precision or/and accuracy. Recently, feasibility of using a neural network instead of conventional 2- or 3-parameter fit modeling has been demonstrated (1-3) to reduce the scan time to 3-4 heartbeats. In MyoMapNet, four T₁w images are collected after a single inversion recovery, which will then be used as inputs to a deep learning model to estimate T₁. However, prior studies used data from a single vendor and field-strength, therefore the generalizability of the models has not been established. The goal of our study is to improve the performance of the previous models by proposing a scanner-independent MyoMapNet architecture.

Methods:

We propose a scanner-independent MyoMapNet (SI-MyoMapNet) architecture by incorporating the relevant scanner information as additional inputs to the model (Fig. 1). SI-MyoMapNet is a deep fully convolutional neural network based on U-Net to generate T₁ map from four T₁w images, corresponding inversion-recovery times (TI), vendor, and field strength. Modified Look-Locker (MOLLI) images from patients undergoing clinical CMR at three different medical centers were extracted: 1) 1249 patients from Beth Israel Deaconess Medical Center (BIDMC) collected using Siemens 3T (MAGNETOM Vida), 2) 99 patients from Weill Cornell Medical Center (Cornell) collected using Siemens 1.5T (MAGNETOM Sola fit), and 3) 75 patients from Boston Medical Center (BMC) collected using Philips 1.5T (Achieva). MOLLI T₁ mapping was performed at each site using vendor-provided imaging protocol. This retrospectively collected dataset were divided into a training/validation (N=853/285) and testing dataset (N=285). The model was trained using mean absolute error cost function for 3000 epochs with an early stopping of 70 to avoid overfitting. The batch size, learning rate, weight decay were 64, 0.01, and 0.001 respectively.

Results:

Fig. 2 shows representative native T₁ maps from 3 different subjects, imaged at Siemens 3T (BIDMC), Siemens 1.5T (Cornell) and Philips 1.5T (BMC) from the testing dataset. The SI-MyoMapNet T₁ maps had similar image quality as MOLLI sequence. Native T₁ values using SI-MyoMapNet strongly correlated with MOLLI in BIDMC (myocardium: r=0.93, P< 0.001; blood: r=0.90, P< 0.001), in Cornell (myocardium: r=0.97, P=0.04; blood: r=0.91, P=0.99), and in BMC (myocardium: r=0.86, P=0.34; blood: r=0.90, P=0.43). In Bland-Altman analysis, SI-MyoMapNet and MOLLI native and post-contrast T₁ were in good agreement in all three sites Fig. 3.

Conclusion:

Inclusion of field-strength and vendor as additional information to the deep learning architecture improves generalizability of MyoMapNet, allowing cardiac T₁ mapping using different vendors or field-strength.