B
Figure 1.
Dark blood single shot PS MDE in a patient with ischemic cardiomyopathy and extensive myocardial infarction. There is a reduction in artifact in (B) the AIR™ Recon DL compared to (A) the standard reconstruction, enabling a more clear view of the ICD. Images courtesy of Erasmus MC.
A
Figure 2.
Patient with cardiac sarcoidosis and an MR-Compatible ICD in an LGE acquisition with a wide-bandwidth MDE sequence. There is a reduction in artifact enabling visualization of the ICD more clearly and with sharper resolution. (B) AIR™ Recon DL set to high compared to (A) the standard reconstruction. Images courtesy of Erasmus MC.
B
Figure 2.
Patient with cardiac sarcoidosis and an MR-Compatible ICD in an LGE acquisition with a wide-bandwidth MDE sequence. There is a reduction in artifact enabling visualization of the ICD more clearly and with sharper resolution. (B) AIR™ Recon DL set to high compared to (A) the standard reconstruction. Images courtesy of Erasmus MC.
‡ Technology in development that represents ongoing research and development efforts. These technologies are not products and may never become products. Not for sale. Not cleared or approved by the US FDA or any other global regulator for commercial availability.
1. van der Velde N, Hassing HC, Bakker BJ, Wielopolski PA, Lebel RM, Janich MA, Kardys I, Budde RPJ, Hirsch A. Improvement of late gadolinium enhancement image quality using a deep learning-based reconstruction algorithm and its influence on myocardial scar quantification. Eur Radiol. 2021 Jun;31(6):3846-3855. doi: 10.1007/s00330-020-07461-w. Epub 2020 Nov 21. PMID: 33219845; PMCID: PMC8128730.
A
Figure 1.
Dark blood single shot PS MDE in a patient with ischemic cardiomyopathy and extensive myocardial infarction. There is a reduction in artifact in (B) the AIR™ Recon DL compared to (A) the standard reconstruction, enabling a more clear view of the ICD. Images courtesy of Erasmus MC.
A
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
B
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
C
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
D
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
E
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
F
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
G
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
H
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
I
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
J
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
K
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
L
Figure 4.
Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
A
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
B
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
C
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
D
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
F
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
G
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
E
Figure 3.
Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
GE_logo.svg
SIGNA_Pulse_Title.svg
arrow_forward.svg
Subscribe Now
Manage Subscription
FOLLOW US
facebook.svg
linkedin.svg
twitter.svg
contact usprivacy policyterms & conditions
© 2022 GENERAL ELECTRIC Company
next arrow.svgprevious arrow.svg
dc_Alexander Hirsch-BW_c.jpg
Alexander Hirsch, MD
Erasmus Medical Center Rotterdam, Netherlands
Marc Sirol_c.jpg
Marc Sirol, MD, PhD
American Hospital of Paris Paris, France
vasanawala-shreyas_c.jpg
Shreyas Vasanwala, MD, PhD
Stanford Children’s Health San Francisco, CA
IN PRACTICE

The power of deep learning in cardiac MR imaging

Based on webinars by Alexander Hirsch, MD, Erasmus Medical Center, Rotterdam, Netherlands, Marc Sirol, MD, PhD, American Hospital of Paris, Paris, France, and Shreyas Vasanawala, MD, PhD, Director of MRI, Stanford Children’s Health, San Francisco, CA
dc_Alexander Hirsch-BW_c.jpg
Alexander Hirsch, MD
Erasmus Medical Center Rotterdam, Netherlands
Marc Sirol_c.jpg
Marc Sirol, MD, PhD
American Hospital of Paris Paris, France
vasanawala-shreyas_c.jpg
Shreyas Vasanwala, MD, PhD
Stanford Children’s Health San Francisco, CA
Imaging of the heart is one of the more challenging MR studies, as the heart’s movement can introduce artifacts to the image.
To better characterize myocardial tissue, Alexander Hirsch, MD, a cardiologist at Erasmus Medical Center in Rotterdam, Netherlands, uses GE Healthcare’s AIR™ Recon DL, a pioneering deep-learning based reconstruction algorithm that improves SNR and image sharpness, enabling shorter scan times. It improves image quality at the foundational level by making use of the raw data to remove image noise and ringing. It’s also possible to use AIR™ Recon DL to simultaneously reduce scan times and acquire high-quality images.
"AIR™ Recon DL increases the sharpness and resolution. And what’s quite nice about it is you have a tunable noise reduction so you can set the noise reduction to a low, medium or high level," says Dr. Hirsch. "It increases sharpness and reduces noise and Gibbs ringing artifact to improve your image quality."
In a recently published peer-reviewed paper on myocardial tissue characterization with deep learning reconstruction1, Dr. Hirsch examined how AIR™ Recon DL can be used in a late gadolinium enhancement sequence to improve image sharpness and quality. He notes the algorithm was originally developed for 2D anatomical sequences and is compatible with many standard sequences and options.
Dr. Hirsch studied late enhancement images from 60 patients, including patients with scarring from ischemic heart disease and others with normal ischemic heart disease and fibrosis. He first looked at the sharpness of the image, particularly at the septum, to assess how blood entered the myocardium. Images enhanced with AIR™ Recon DL showed a significant improvement in sharpness.
"There is a clear increase in image sharpness using the deeplearning reconstruction, beginning at the low noise reduction level," he explains.
In addition, three investigators independently scored image quality as poor, fair, good, very good or excellent. "AIR™ Recon DL images demonstrated a reduction in noise level and the image quality was clearly improved," compared to the other images, he says.
In a patient with an ischemic cardiomyopathy and extensive myocardial infarction, AIR™ Recon DL clearly shows the location of the myocardial infarction, even with noise reduction. "We used the high noise reduction, with more homogeneous signal from the blood and a nice deviation between the blood and the myocardial infarction," by using a dark blood PS MDE sequence, he says (Figure 1).
IP_Cardiac Figure 1 Image A.jpg
A
IP_Cardiac Figure 1 Image B.jpg
B
Figure 1. Dark blood single shot PS MDE in a patient with ischemic cardiomyopathy and extensive myocardial infarction. There is a reduction in artifact in (B) the AIR™ Recon DL compared to (A) the standard reconstruction, enabling a more clear view of the ICD. Images courtesy of Erasmus MC.
In another patient with myocardial infarction, cardiac sarcoidosis and an MR-Conditional implantable cardioverter defibrillator (ICD), a single-shot free breathing MDE sequence with AIR™ Recon DL at high noise reduction delivered a clear view of the heart. In the presence of MR-Conditional implants, a wideband MDE adiabatic pulse (available in all MDE sequences across the GE portfolio) can be applied to improve tissue nulling, making it more robust against inversion inaccuracies caused by implants. Additionally, a reduction in artifact with AIR™ Recon DL enabled clear visualization of late enhancement of the whole septum inferior and the right ventricle, even in the presence of an implanted ICD (Figure 2).
IP_Cardiac Figure 2 Image A.jpg
A
IP_Cardiac Figure 2 Image B.jpg
B
Figure 2.
Patient with cardiac sarcoidosis and an MR-Compatible ICD in an LGE acquisition with a wide-bandwidth MDE sequence. There is a reduction in artifact enabling visualization of the ICD more clearly and with sharper resolution. (B) AIR™ Recon DL set to high compared to (A) the standard reconstruction. Images courtesy of Erasmus MC.
Dr. Hirsch says AIR™ Recon DL is useful for not only late gadolinium enhancement, but also for other anatomical images, such as T1 or T2 images. For example, he used a single-shot free breathing T2 FSE with noise reduction in a patient with a constrictive pericarditis. "We used only the free-breathing sequence, and you could see where the pericardium is thickened, in the inferior wall and the lateral wall with the right ventricle. We could visualize the four chambers to also see the thickened pericardium at the basal and mid-ventricle levels.
"Late gadolinium enhancement and parametric mapping are establishing imaging techniques in cardiovascular progression imaging. However, new developments, including deep-learning reconstruction techniques, lead to improvements in the image quality," he says.
For example, deep-learning reconstruction techniques can be used for faster acquisition and thinner slices, making it especially useful for free-breathing sequences, like single-shot MDE and anatomical 2D images.
Deep-learning image reconstruction at 3.0T
The American Hospital of Paris performs more than 1,000 cardiac MR studies per year. The hospital installed a 3.0T magnet in early 2020 and added AIR™ Recon DL in November that year. Marc Sirol, MD, PhD, a radiologist at the hospital, says 3.0T cardiac MR imaging with AIR™ Recon DL offers improvements across the board.
"We gain contrast and sharpness in the image. It’s clear the artificial intelligence and deep learning help us in that manner," he says.
Dr. Sirol uses AIR™ Recon DL images for segmentation and quantification, as well as for classification, reporting and predicting prognoses.
In dark blood imaging in a patient without any edema, standard reconstruction images are of diagnostic quality, but slightly blurry with some noise in the image. However, images reconstructed with AIR™ Recon DL are crisp and clear. "The sharpness of the image is much better, and the CNR and the SNR are even better," he says.

In a subject with suspected giant cell arteritis, he compared traditionally reconstructed T1 and T2 images to images with AIR™ Recon DL of the same patient, taken just minutes apart (Figure 3).
Figure 3. Patient with giant cell arteritis. (A, B) Black blood T1 with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (C, D) Black blood T2 FatSat with AIR™ Recon DL, 1.8 x 1.8 x 7 mm3. (E) Contrast-enhanced MRA, 1.0 x 1.8 x 2.4 mm3. (F, G) Post-contrast black blood T1 FatSat, 1.8 x 1.8 x 7 mm3. Images courtesy of American Hospital of Paris.
"Even though the conventional reconstruction is pretty good because the high field strength of a 3.0T scanner provides a good signal, the deep-learning reconstruction helped us gain more sharpness of the image and signal," he says.
The hospital performs many stress cardiac MR studies every day. In a 56-year-old male patient with hypertension and Type-2 diabetes, with a history of proximal LAD angioplasty, the AIR™ Recon DL images were compared to standard images.
"In the AIR™ Recon DL images, the image is even sharper, the contrast is even better and we could see that the ischemia is wider now than previously. You can be more confident of your diagnosis on inferior and lateral infractions. Because of this, we referred this patient to the cath lab, and it confirmed that he had a secluded marginal branch."
Seeing these cardiac details clearly can help clinicians differentiate ischemia from inflammation, for example, to confidently diagnose challenging cases, even when breath-holds are difficult or in cases of arrhythmias.
"AIR™ Recon DL can decrease noise in the image and increase the image sharpness. It reduces breath-hold time for difficult patients," he says. "All the MDE images we are performing are not requiring patient breath-holding. It’s a single shot image that you can perform free breathing."
His experience demonstrates the value of deep learning on 3.0T imaging for cardiac MR. His department has now used AIR™ Recon DL on more than 1,000 cases. "It’s feasible to perform cardiac imaging at 3.0T, and now we do it routinely at American Hospital in Paris. AIR™ Recon DL is a real-time reconstruction based on the raw data, which is important so you don’t have to wait for the image," says Dr. Sirol.
Deep learning for pediatric cardiac scans
In pediatric cardiac MR imaging (CMR), there’s a need for efficient protocols because patients have a limited ability to stay still and hold their breath. Enhancing the efficiency of these scans with free-breathing, rapid and automated approaches is critical to capturing diagnostic-quality images and providing care for these precious patients.
DL Cine is a work-in-progress from GE Healthcare that is being developed to support rapid and automated cardiac MR imaging. It is an approach that combines conventional 2D Cine imaging with deep-learning reconstruction to leverage respiratory triggering to avoid breathholds for the patient. DL Cine includes automated segmentation‡ to easily and quickly capture quantitative metrics.
DL Cine is being designed to enable freebreathing scanning that’s easier for the patient. "From the patient’s perspective, it will happen automatically without a lot of breath-holds," says Shreyas Vasanawala, MD, PhD, the Director of MRI at Stanford Children’s Health in San Francisco, where he’s also the Division Chief of Pediatric Radiology, Associate Chair of Radiology, and the Radiologistin-chief for Pediatric Radiology.

In a standard cardiac exam, prescribing scan planes is time-consuming. It’s also inefficient to obtain a short axis with multiple breath holds, as they lead to long exams that are uncomfortable for patients and can lead to inaccurate quantification of ejection fractions due to the use of multiple breathholds, where the heart can be in different positions during the scan acquisition, as well as cardiac motion.
"In order to overcome this, we’ve been working on a deep-learning approach that will allow for greatly accelerated acquisitions, and at the same time may enable respiratory triggering of the slices of those acquisitions so that you don’t have to ask for breath-holds," he says (Figure 4).
IP_Cardiac Figure 4 Image A.jpg
A
IP_Cardiac Figure 4 Image B.jpg
B
IP_Cardiac Figure 4 Image C.jpg
C
IP_Cardiac Figure 4 Image D.jpg
D
IP_Cardiac Figure 4 Image E.jpg
E
IP_Cardiac Figure 4 Image F.jpg
F
IP_Cardiac Figure 4 Image G.jpg
G
IP_Cardiac Figure 4 Image H.jpg
H
IP_Cardiac Figure 4 Image I.jpg
I
IP_Cardiac Figure 4 Image J.jpg
J
IP_Cardiac Figure 4 Image K.jpg
K
IP_Cardiac Figure 4 Image L.jpg
L
Figure 4. Images acquired from a pediatric patient with abnormal trabeculations in the right ventricle. Images show multiple slices from the (A-D) standard clinical scan acquired with seven breath-holds, and a single breath-hold acquisition using (E-H) compressed sensing and (I-L) deep-learning reconstruction. The deep-learning images provide finer definition of RV trabeculations over compressed sensing (arrows). Images courtesy of Stanford Children’s Health.
For example, DL Cine may enable automatic views of the short access plane to quantify ventricular volume ejection fractions. It will start with the axial localizer and then will use deep learning to automatically display pseudo two-chamber and four-chamber views, followed by a short access view. "This will all happen automatically and greatly accelerate the speed at which these exams can be acquired," Dr. Vasanawala explains.
In addition, it could simplify and speed cardiac output and ejection fraction measurements. In an internal evaluation of DL Cine on 30 patients, Dr. Vasanawala reports that DL Cine enabled a faster scan time of 2 minutes compared to 10 minutes with parallel imaging.
END_MARK_In_Practice-01.png