Image Reconstruction (including machine learning)
Naledi L. Adam, MSc
PhD student
King's College London
LONDON, England, United Kingdom
Naledi L. Adam, MSc
PhD student
King's College London
LONDON, England, United Kingdom
Grzegorz T. Kowalik, PhD
Research Associate
King's College London
London, England, United Kingdom
Andrew Tyler, PhD
Research Associate
King's College London, United Kingdom
Ronald Mooiweer, PhD
MRI Scientist & Research Associate
Siemens Healthcare & King's College London, United Kingdom
Alexander P. Neofytou, MSc
PhD Student
King's College London
London, England, United Kingdom
Sarah McElroy, PhD
Clinical scientist
Siemens Healthcare Limited
London, England, United Kingdom
Karl P. Kunze, PhD
Senior Cardiac MR Scientist
Siemens Healthineers
London, England, United Kingdom
Peter Speier, PhD
Principal Key Expert
Magnetic Resonance, Siemens Healthcare GmbH, Erlangen, Germany
Erlangen, Bayern, Germany
Daniel Stäb, PhD
Senior Scientist
Siemens Healthcare Pty Ltd, Melbourne, Australia
Melbourne, Victoria, Australia
Radhouene Neji, PhD
Siemens Research Scientist
King's College London, United Kingdom
Muhummad Sohaib Nazir, PhD
NIHR Clinical Lecturer in Cardiology
King's College London, United Kingdom
Reza Razavi, MD
Professor of Paediatric Cardiovascular Science
King's College London
London, England, United Kingdom
Amedeo Chiribiri, MD PhD FHEA FSCMR
Professor of Cardiovascular Imaging; Consultant Cardiologist
King's College London
London, England, United Kingdom
Sébastien Roujol, PhD
Reader in Medical Imaging
King's College London
London, England, United Kingdom
Simultaneous multi-slice (SMS) bSSFP perfusion imaging enables increased slice coverage with high spatial resolution1 and shows promise to improve the detection of perfusion defects. Since standard SMS reconstructions based on parallel imaging results in noise enhancement, iterative reconstruction (Iter) with spatio-temporal regularization have been employed 1. However, such reconstruction is computationally intensive which remains a challenge for online use. It may also introduce signal alteration due to the spatio-temporal regularization that may affect myocardial perfusion quantification. In this study, we present a fast bias-free reconstruction for SMS CMR perfusion.
Methods: 17 patients (6f, 60±5y/o) with suspected coronary artery disease were recruited for a stress perfusion scan using a SMS-bSSFP sequence (resolution 1.9x1.9mm2, 6 slices, 80 dynamics, MB factor =2)1 with CAIPIRINHA encoding, GC-LOLA correction2 and “lean” implementation for slice separation3. Images were reconstructed with 1) Iter, 2) TGRAPPA, and 3) TGRAPPA followed by fast AI-based image denoising (proposed). Image denoising used a residual learning 2D Res-DUnet (patch-based learning, patch size=64x64, dropout rate=0.5, learning rate =1e-5, and Adam optimizer) to predict a noise map from a noisy image. Training of the network was based on 300 high SNR CINE images (80%/20% training/validation) on which gaussian noise was added. To compare these three reconstructions, quantitative myocardial sharpness, peak to baseline signal ratio in the myocardium (PBmyo) and LV blood pool (PBblood), image quality (0-poor; 3-excellent), perceived SNR (0-poor; 3-excellent), and image diagnostic value (yes/no) were evaluated.
Results:
Example reconstructions are shown in Figure 1. Iter led to higher perceived SNR (3.0±0.0 vs. 2.0±0.0 (proposed) and 1.3±0.6 (TGRAPPA), p< 0.001 for both), higher image quality (2.7±0.4 vs. 1.8±0.4 (proposed) and 1.3±0.6 (TGRAPPA), p< 0.001 for both) but resulted in altered PBmyo/PBblood (3.9±1.7/11.8±7.0 vs. 2.3±0.5/5.8±1.3 (proposed) and 2.3±0.5/5.8±1.3 (TGRAPPA, serving as reference), p< 0.001) and longer computational time (12 min vs. 20 s (proposed) for an entire dataset) (Figure 2/3). There were no significant differences in terms of myocardial sharpness (p=0.77) and rate of diagnostic images between all techniques (94-100%, p=0.36). In comparison to TGRAPPA, the proposed reconstruction led to higher perceived SNR (p< 0.001), higher image quality (p=0.008), and no significant differences in terms of PBmyo /PBblood (p=0.89/0.53).
Conclusion: The proposed approach enables fast bias-free reconstruction of SMS perfusion images with increased SNR and image quality and no image sharpness loss with respect to TGRAPPA. This technique may provide an attractive online SMS reconstruction and could complement iterative reconstructions by providing a second set of images with high signal fidelity which has potential to improve the accuracy of perfusion quantification.