Detection of Myocardial Ischemia in Stress Perfusion CMR with Order-of-magnitude Lower Contrast Dose Enabled by Multi-Stage Deep Learning (RF_TH_273)

Thursday, January 30, 2025

5:50 PM – 6:00 PM East Coast USA Time

Location: Blue Room PreFunction

Presenting Author(s)

KY

Khalid Youssef, PhD

Assistant Research Professor
Indiana University, Department of Radiology and Imaging Sciences

Primary Author(s)

KY

Khalid Youssef, PhD

Assistant Research Professor
Indiana University, Department of Radiology and Imaging Sciences

Co-Author(s)

LZ

Luis F. Zamudio Rivero, MSc

Research Associate
Laboratory for Translational Imaging of Microcirculation, Krannert Cardiovascular Research Center, Indiana University School of Medicine, Indianapolis, IN
BH

Bobak Heydari, MD

Professor, Harvard Medical School
Brigham and Women's Hospital
AH

Andrew G. Howarth, MD, PhD

Associate Professor
Libin Cardiovascular Institute; University of Calgary, Canada
Rohan Dharmakumar, PhD

Executive Director
Indiana University School of Medicine
Behzad Sharif, PhD

Associate Professor of Biomedical Engineering
Purdue University

Background: Detection of inducible myocardial ischemia based on stress first-pass perfusion (FPP) CMR imaging is a cornerstone in the diagnosis and management of ischemic heart disease. Typically, this assessment relies on myocardial blood flow (MBF) quantification using FPP studies, which typically requires a 0.05–0.1 mmol/kg dose of a Gadolinium (Gd)-based contrast agent. Deposition of Gd-based contrast agents in the brain and its potential cumulative effects have been well publicized.¹ Such concerns may have implications for the feasibility of serial FPP studies. We propose a deep learning-based approach that leverages multi-stage training (MST) ^2-4 to enable accurate ischemia detection in stress CMR, even with a »10-fold reduction in Gd contrast dose.

Methods: The proposed methodology (Fig 1) includes two convolutional neural networks ensembles (CNN-E) and an MST. The first CNN-E identifies the region of interest (ROI) encompassing the right and left ventricles (RV and LV). The 2^nd CNN-E is trained using a patch-wise approach for myocardial segmentation in the ROI ⁵. Finally, the MST network quantifies the blood flow in the myocardium. RV segmentation mask is used to detect the insertion point and divide each short-axis slice into 6 sectors. The training dataset included stress/rest perfusion images from 180 volunteer patients with suspected ischemia who underwent dual-bolus FPP exam at 3T. The diagnostic performance for the proposed deep learning technique was evaluated using a cohort of 30 patients who underwent stress FPP with back-to-back FPP imaging using a 90% reduced contrast dose (“dilute bolus” scan) followed by the conventional full dose FPP scan, all during vasodilator stress. The ground truth was generated based on visual assessment by expert cardiologist (standard read of clinically-indicated stress study). The diagnostic accuracy of MBF values derived from the proposed method was compared to those obtained using the established Fermi-constrained deconvolution technique, with receiver operating characteristic (ROC) curves generated at both the per-vessel and per-patient levels.

Results: Representative results (Fig 2) demonstrate robust performance, even in cases with 90% Gd contrast dose reduction and low SNR. Diagnostic performance (ROC curve) analysis of the patient cohort (Fig 3) show that the proposed method significantly outperformed the traditional Fermi deconvolution approach in distinguishing healthy from diseased vessels and patients (p< 0.05 for per-vessel detection).

Conclusion: This study introduces a novel multi-stage deep learning framework that enables accurate myocardial ischemia detection in stress CMR with an order-of-magnitude reduction in Gd contrast dose. Our results indicate that the deep learning framework may achieve a high diagnostic performance for detection of stress-induced ischemia. By reducing the need for Gd contrast dose, this approach holds promise for broader clinical use, enhancing patient safety especially in clinical settings that require serial imaging.