Image Reconstruction (including machine learning)
Siyeop Yoon, PhD
Post-doc Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Siyeop Yoon, PhD
Post-doc Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Eiryu Sai, MD, PhD
Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Tokyo, Massachusetts, United States
Kei Nakata, MD
Clinical fellow
Mie University Hospital
Tsu, Mie, Japan
Salah Assana, MSc
Research Assistant
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Manuel A. Morales, PhD
Post-doc Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Amine Amyar, PhD
Postdoctoral Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Jennifer Rodriguez
Clinical Trials Specialist
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
Julia Cirillo, BSc
Research Assistant
Beth Israel Deaconess Medical Center, Massachusetts, United States
Beth Goddu, RT
MRI Tech
Beth Israel Deaconess Medical Center, United States
Patrick Pierce, RT
MRI Tech
Beth Israel Deaconess Medical Center, Massachusetts, United States
Warren J. Manning, MD
Professor of Medicine and Radiology
Harvard Medical School
Boston, Massachusetts, United States
Reza Nezafat, PhD
Professor
Harvard Medical School
Boston, Massachusetts, United States
Real-time cine imaging (RT-cine) with physiological exercise (Ex-CMR) can provide diagnostic and prognostic information in patients with ischemic or non-ischemic disease. However, the spatial and temporal resolution of RT-cine is inferior to segmented cine acquisition. This study aimed to develop a 16-fold accelerated RT-cine sequence with a high spatial and temporal resolution (1.8×1.8 mm2 and 28ms).
Methods:
We propose a highly accelerated RT-cine by combining compressed sensing (CS) with a novel Resolution Enhancement Generative Adversarial Inline Neural Network (REGAIN) (Fig 1A). RT-Cine images are prescribed with ½ reduced spatial resolution in the phase-encode direction. The k-space is non-uniformly under-sampled and reconstructed using vendor-provided CS, followed by zero-padding to generate a blurry image. Subsequently, REGAIN enhances the sharpness to restore spatial resolution. REGAIN is based on the generative adversarial network (1) and modified for a flexible input size and a reduced number of parameters (Fig 1B, C). REGAIN is integrated with the scanner using FIRE (2) for inline reconstruction. The generator of REGAIN sharpens the low-resolution image input to achieve higher resolution, and the discriminator classifies a high-resolution image and the generator’s output for adversarial training. For the network training, we collected k-space of segmented cine images from 343 patients who underwent clinical CMR and synthesized low-resolution images by discarding the outer ky lines, analogous to prescribing a low-resolution image, using vendor reconstruction (Fig 1D).
The proposed RT-cine was evaluated in 85 patients with cardiac disease (52 male; 55±15 years) and 45 healthy participants (13 male; 28±10 years) at rest. A subset of subjects (12 patients and 20 healthy) underwent Ex-CMR immediately after exercise using a supine cycle ergometer (Lode, Groningen, The Netherlands) (Fig 2). Images were collected in the short-axis using parameters as follows: a Siemens 3T Vida, TR/TE = 2.8/1.2ms, flip angle 26°, slice thickness 8mm, matrix size 144-192×192, a spatial resolution of 3.6×1.8mm2 (reconstructed via REGAIN to 1.8×1.8 mm2), a temporal resolution of 28-34ms, and 6.8-8.0-fold CS. All images were graded on a 3-point Likert scale for diagnostic quality (1=non-diagnostic to 3=diagnostic) and a 5-point Likert scale for artifact scoring (1=poor to 5=excellent). Image qualities were compared using Wilcoxon signed-rank tests.
Results:
All images were successfully acquired and reconstructed inline (reconstruction time of 52 ms per frame). Fig 3 shows RT-cine images demonstrating excellent image quality with a high spatial-temporal resolution. All images were of diagnostic quality. REGAIN improved artifact score compared to zero-padding (4.2±0.8 vs. 2.7±0.6 and 4.1±1.0 vs. 3.0±0.9, for the rest and Ex-CMR, respectively; all P< .05).
Conclusion: A 16-fold accelerated RT-cine enables high spatial and temporal resolutions with excellent image quality and minimal artifact.