Figure 1.
AIRTM Recon DL is integrated directly into the MR image reconstruction pipeline to intelligently reconstruct a final image with high SNR and sharpness.
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.
C
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.
A
Figure 3.
Spontaneous median neuropathy of the elbow. Axial PD 2D FSE. (A) Image acquired with standard protocol, 512 x 352, 12 cm FOV, 2 NEX; (B) image reconstructed with AIRTM Recon DL at maximum SNR improvement. Note the clarity of the median nerve in the AIRTM Recon DL image. Images courtesy of HSS
B
Figure 3.
Spontaneous median neuropathy of the elbow. Axial PD 2D FSE. (A) Image acquired with standard protocol, 512 x 352, 12 cm FOV, 2 NEX; (B) image reconstructed with AIRTM Recon DL at maximum SNR improvement. Note the clarity of the median nerve in the AIRTM Recon DL image. Images courtesy of HSS
A
Figure 4.
Coronal PD 2D FSE image of the elbow depicts normal ulnotrochlear cartilage. (A) Standard protocol, 512 x 352, 14 cm FOV, 1 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly demonstrates the superficial cartilage layer (lamina splendens) and subchondral bone. Images courtesy of HSS
B
Figure 4.
Coronal PD 2D FSE image of the elbow depicts normal ulnotrochlear cartilage. (A) Standard protocol, 512 x 352, 14 cm FOV, 1 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly demonstrates the superficial cartilage layer (lamina splendens) and subchondral bone. Images courtesy of HSS
A
Figure 5.
Axial PD 2D FSE image through the arm in a patient with a severe, spontaneous median neuropathy. (A) 512 x 352, 12 cm FOV, 2 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly depicts fascicular detail and enlargement. Images courtesy of HSS
B
Figure 5.
Axial PD 2D FSE image through the arm in a patient with a severe, spontaneous median neuropathy. (A) 512 x 352, 12 cm FOV, 2 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly depicts fascicular detail and enlargement. Images courtesy of HSS
A
Figure 6.
Standard protocol. Coronal T2* GRE, 0.3 x 0.6 x 1.7 mm; (B) AIRTM Recon DL at maximum SNR improvement. Images courtesy of HSS
B
Figure 6.
Standard protocol. Coronal T2* GRE, 0.3 x 0.6 x 1.7 mm; (B) AIRTM Recon DL at maximum SNR improvement. Images courtesy of HSS
A
Figure 7.
Elbow. (A) Unfiltered, axial 2D FSE, 256 x 180, 1 NEX, 1:10 min. (B) AIRTM Recon DL at maximum-plus SNR improvement, 256 x 180, 1 NEX, 1:10 min. (C) Reference unfiltered, 512 x 352, 2 NEX, 4:09 min. Images courtesy of HSS
B
Figure 7.
Elbow. (A) Unfiltered, axial 2D FSE, 256 x 180, 1 NEX, 1:10 min. (B) AIRTM Recon DL at maximum-plus SNR improvement, 256 x 180, 1 NEX, 1:10 min. (C) Reference unfiltered, 512 x 352, 2 NEX, 4:09 min. Images courtesy of HSS
C
Figure 7.
Elbow. (A) Unfiltered, axial 2D FSE, 256 x 180, 1 NEX, 1:10 min. (B) AIRTM Recon DL at maximum-plus SNR improvement, 256 x 180, 1 NEX, 1:10 min. (C) Reference unfiltered, 512 x 352, 2 NEX, 4:09 min. Images courtesy of HSS
‡ 510(k) pending at the US FDA. Not yet CE marked. Not available for sale.
AIR™ Recon DL‡ recovers a high-quality image from an otherwise noisy thin-slice axial T2 prostate image acquired in only 1:07 min.
‡ 510(k) pending at the US FDA. Not yet CE marked. Not available for sale.
AIR™ Recon DL‡ recovers a high-quality image from an otherwise noisy thin-slice axial T2 prostate image acquired in only 1:07 min.
‡ 510(k) pending at the US FDA. Not yet CE marked. Not available for sale.
Original image
With AIR™ Recon DL, images can be reconstructed with mild, medium or maximum SNR improvement for visibly improved image quality compared to the original image.
Mild SNR improvement
With AIR™ Recon DL, images can be reconstructed with mild, medium or maximum SNR improvement for visibly improved image quality compared to the original image.
Medium SNR improvement
With AIR™ Recon DL, images can be reconstructed with mild, medium or maximum SNR improvement for visibly improved image quality compared to the original image.
Maximum SNR improvement
With AIR™ Recon DL, images can be reconstructed with mild, medium or maximum SNR improvement for visibly improved image quality compared to the original image.
510(k) pending at the US FDA. Not yet CE marked. Not available for sale.
A
Sagittal STIR lumbar spine reconstructed without (A) and with (B) AIRTM Recon DL. Image quality in the spinal cord is clearly improved through reduced noise and ringing. 0.8 x 0.9 x 1.5 mm 2:47 min.
A
Sagittal STIR lumbar spine reconstructed without (A) and with (B) AIRTM Recon DL. Image quality in the spinal cord is clearly improved through reduced noise and ringing. 0.8 x 0.9 x 1.5 mm 2:47 min.
B
Sagittal STIR lumbar spine reconstructed without (A) and with (B) AIRTM Recon DL. Image quality in the spinal cord is clearly improved through reduced noise and ringing. 0.8 x 0.9 x 1.5 mm 2:47 min.
B
Sagittal STIR lumbar spine reconstructed without (A) and with (B) AIRTM Recon DL. Image quality in the spinal cord is clearly improved through reduced noise and ringing. 0.8 x 0.9 x 1.5 mm 2:47 min.
Subscribe Now
Manage Subscription
FOLLOW US
contact usprivacy policyterms & conditions
© 2023 GE HealthCare. GE is a trademark of the General Electric Company used under trademark license.
next arrow_PURPLE.svgprevious arrow_PURPLE.svg
dc_Darryl_Sneag-2015-BW_c.jpg
Darryl Sneag, MD The Hospital for Special Surgery New York, New York
dc_Hollis_Potter-2015-BW_c.jpg
Hollis Potter, MD The Hospital for Special Surgery New York, New York
dc_Pascal_Roux-BW_c.jpg
Pascal Roux, MD Centre Cardiologique du Nord Paris, France


SPOTLIGHT

A new era of deep-learning image reconstruction

dc_Darryl_Sneag-2015-BW_c.jpg
Darryl Sneag, MD
The Hospital for Special Surgery New York, New York
dc_Hollis_Potter-2015-BW_c.jpg
Hollis Potter, MD
The Hospital for Special Surgery New York, New York
dc_Pascal_Roux-BW_c.jpg
Pascal Roux, MD
Centre Cardiologique du Nord Paris, France

Radiologists and technologists are intimately familiar with the traditional compromise in MR between image quality and scan time. Higher image quality — through higher SNR and/or spatial resolution needed to resolve anatomical detail — necessitates long scan times, whereas faster scans — desired for patient comfort and productivity — compromise image quality and diagnostic confidence. AIRTM Recon DL, an innovative new reconstruction technology from GE Healthcare based on deep learning, offers a fundamental shift in this balance between image quality and scan time, resulting in TrueFidelityTM MR images that elevate the science of image reconstruction for clinical excellence without conventional compromises.

Conventional MR image reconstruction gives rise to well-known image artifacts as a direct result of the data acquisition and reconstruction process. For example, thermal and electrical noise during data sampling translates into random image noise that reduces SNR, while incomplete sampling of high spatial frequencies creates partial volume and edge ringing (i.e., Gibbs ringing) artifacts in the final reconstructed image.
By operating on raw data within the online reconstruction pipeline, AIRTM Recon DL benefits from access to the full set of acquired source data to generate an image, compared to post DICOM image conversion where important information has been lost.
AIRTM Recon DL uses a feed-forward deep convolutional neural network trained on over 10,000 images using GE’s Edison AI Platform. Supervised learning was performed by using data pairs of high SNR, high-resolution images and low-SNR, low-resolution images. The trained network employs a cascade of over 100,000 unique filters that recognize patterns characteristic of noise and low resolution to reconstruct only the ideal object image. The network includes a tunable SNR improvement level expressed as mild, medium and maximum to accommodate user preference. AIRTM Recon DL includes an innovative ringing suppression technology: rather than simply removing Gibbs ringing, the network recognizes where ringing occurs and recasts this former artifact into improved image detail. The result is an image with high SNR and spatial resolution that is virtually free of truncation artifacts.
Figure 1. AIRTM Recon DL is integrated directly into the MR image reconstruction pipeline to intelligently reconstruct a final image with high SNR and sharpness.
IS-AIRecon_Figure_1.jpg
Traditional methods to address these artifacts include hardware, software and acquisition approaches. Hardware solutions such as higher field strength magnets and more RF coil elements can improve SNR. Software filters are commonly applied in the data reconstruction pipeline to mitigate noise and ringing; however, these are only partially effective and can have the undesired impact of reducing effective spatial resolution. In the acquisition protocol, scan parameters can be adjusted to improve image quality, but this comes at a high cost. For example, SNR can be improved by increasing the number of signal averages (NEX) with a proportional increase in scan time; truncation artifacts can be mitigated by increasing spatial resolution, which in turn typically increases scan time and also reduces SNR. This costly SNR/spatial resolution/scan time interdependency forces clinicians to make difficult trade-offs between image quality and scan time for a given patient and clinical need. Though there has been some success easing this MR trade-off with existing technologies, the reality is that many images today still suffer from low SNR and artifacts, which can lead to decreased diagnostic confidence and reduced radiologist productivity. Patients may be called back for re-scans, which leads to fewer daily scan slots available for scheduling new patients. It can also lead to lower patient throughput due to repeated scans during the exam, further backlogging the schedule and leading to a poorer patient experience. Artificial intelligence now offers an exciting new means to mitigate traditional image artifacts and generate clearer, higher-quality images than previously obtainable from the same MR data. AIRTM Recon DL represents a revolution in MR image reconstruction by introducing a deep learning-based convolutional neural network to intelligently reconstruct a final MR image with high SNR and image sharpness. AIRTM Recon DL is not a post processing technique but rather is embedded directly in the reconstruction pipeline, where the neural network model is applied to acquired input data to remove noise and ringing artifacts prior to final image formation (Figure 1).
With AIRTM Recon DL, the potential is for technologists to acquire higher SNR without a time penalty and for radiologists to have more consistency and quality in the images they interpret. Alternatively, scan time may be reduced without compromising detail or SNR.
For example, if an MR technologist decreases slice thickness or in plane pixel size, the amount of signal is proportionately reduced, which typically leads to noisier images. With AIRTM Recon DL, the result is higher SNR images and this may enable radiologists to be more confident in their reading and reporting.
The best of both worlds
Pascal Roux, a radiologist at Centre Cardiologique du Nord (CCN), one of the first global clinical sites to evaluate AIRTM Recon DL for GE, believes that AIRTM Recon DL is a solution that offers a dramatic improvement over existing image reconstruction techniques. “In my experience, AIRTM Recon DL demonstrated high-resolution images with no truncation artifact, imperceptible noise and depiction of sharp structure,” Dr. Roux says. As of the end of August 2019, CCN had performed nearly 1,000 exams with a prototype version of AIRTM Recon DL.
In one case, he was able to detect a lesion on a spinal cord exam that was difficult to appreciate on the images processed without AIRTM Recon DL. In Dr. Roux’s opinion, the lesion was more clearly visible on the images processed with AIRTM Recon DL (Figure 2).
IS-AIRecon_Figure_2_Image_A.jpg
A
IS-AIRecon_Figure_2_Image_B.jpg
B
IS-AIRecon_Figure_2_Image_C.jpg
C
Figure 2. AIRTM Recon DL improves SNR to help depict lesions. (A) Existing protocol: sagittal T2 FSE, 0.9 x 1.0 x 3.5 mm, 4 NEX, 2:50 min. (B) Revised protocol: sagittal T2 FSE, 0.9 x 1.0 x 3.5 mm, 2 NEX, 1:28 min. (C) Image in 2B reconstructed with AIRTM Recon DL at maximum noise reduction to enable shorter scan time without sacrificing SNR. Images courtesy of CCN
“Anytime a new technology can help improve resolution, it will help us to better analyze lesions.”
Dr. Pascal Roux
Reading an AIRTM Recon DL image is very natural and comfortable for Dr. Roux. He expects to be more confident in his diagnosis because AIRTM Recon DL is designed to help improve SNR and image sharpness, which can enhance spatial resolution as well as help remove artifacts and reduce acquisition time.
“I have the best of both worlds. I do not have to choose between improving the quality of the exam and shortening the exam time,” he says.
AIRTM Recon DL is an excellent tool to improve workflow. If Dr. Roux’s department can increase the number of exams even by a fraction each hour, the cumulative result at the end of the day could be significant. With a three- exam-each-hour schedule, Dr. Roux believes it is possible to add five to six more patients in a 12-hour day. The shorter acquisition time also means that when he needs to capture an additional image for a difficult case, he can do it without worrying about the schedule. “Sometimes a sequence fails, or you get great information and want to add something,” Dr. Roux explains. “It is hard to do that when an MR exam is 20-30 minutes. However, if we can go faster because we can reconstruct it with a deep-learning solution such as AIRTM Recon DL, then we have sufficient time to do this in the scan room.”
Finding the “sweet spot”
The Hospital for Special Surgery (HSS) is another of several global sites evaluating AIRTM Recon DL and its impact on image quality, spatial resolution and acquisition scan time. Darryl Sneag, MD, Director of Peripheral Nerve MRI, Erin Argentieri, senior lead research specialist and Hollis Potter, MD, Chairman, Department of Radiology and Imaging, examined the use of AIRTM Recon DL in peripheral nerve and musculoskeletal (MSK) imaging.
“AIRTM Recon DL provides the added resolution that we need when looking at musculoskeletal structures, such as ligaments, tendons, nerves and the trabecular detail of the bones,” says Dr. Sneag.

The difference is like ‘night and day’ for Dr. Potter, particularly when using a 512 x 512 matrix with one excitation (1 NEX). With AIRTM Recon DL, trabecular detail is not blurred and the individual nerve fascicles are clearly demonstrated (Figures 3 and 4). Previously, at a 512 x 512 matrix, SNR would be a challenge, but with AIRTM Recon DL, Drs. Potter and Sneag can push the MR system to a higher matrix and achieve impressive imaging results.
IS-AIRecon_Figure_3_Image_A.jpg
A
IS-AIRecon_Figure_3_Image_B.jpg
B
Figure 3. Spontaneous median neuropathy of the elbow. Axial PD 2D FSE. (A) Image acquired with standard protocol, 512 x 352, 12 cm FOV, 2 NEX; (B) image reconstructed with AIRTM Recon DL at maximum SNR improvement. Note the clarity of the median nerve in the AIRTM Recon DL image. Images courtesy of HSS
IS-AIRecon_Figure_4_Image_A.jpg
A
IS-AIRecon_Figure_4_Image_B.jpg
B
Figure 4. Coronal PD 2D FSE image of the elbow depicts normal ulnotrochlear cartilage. (A) Standard protocol, 512 x 352, 14 cm FOV, 1 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly demonstrates the superficial cartilage layer (lamina splendens) and subchondral bone. Images courtesy of HSS
IS-AIRecon_Figure_5_Image_A.jpg
A
IS-AIRecon_Figure_5_Image_B.jpg
B
Figure 5. Axial PD 2D FSE image through the arm in a patient with a severe, spontaneous median neuropathy. (A) 512 x 352, 12 cm FOV, 2 NEX; (B) AIRTM Recon DL at maximum SNR improvement more clearly depicts fascicular detail and enlargement. Images courtesy of HSS
IS-AIRecon_Figure_6_Image_A.jpg
A
IS-AIRecon_Figure_6_Image_B.jpg
B
Figure 6. (A) Standard protocol. Coronal T2* GRE, 0.3 x 0.6 x 1.7 mm; (B) AIRTM Recon DL at maximum SNR improvement. Images courtesy of HSS
“In our experience, this tool enables us to back off on the number of averages or achieve a higher matrix, to either save on scan time or achieve a higher resolution image.”
Dr. Hollis Potter
“There is more detail in the image, especially at a lower matrix. In some conventionally-processed MR images, the trabecular pattern is poor, the nerves are blurred and there is a lot of noise in the image. With AIRTM Recon DL, the difference is striking,” Dr. Potter says (Figure 7).
IS-AIRecon_Figure_7_Image_A.jpg
A
IS-AIRecon_Figure_7_Image_B.jpg
B
IS-AIRecon_Figure_7_Image_C.jpg
C
Figure 7. Elbow. (A) Unfiltered, axial 2D FSE, 256 x 180, 1 NEX, 1:10 min. (B) AIRTM Recon DL at maximum-plus SNR improvement, 256 x 180, 1 NEX, 1:10 min. (C) Reference unfiltered, 512 x 352, 2 NEX, 4:09 min. Images courtesy of HSS
Dr. Potter adds that with the high- resolution AIRTM Recon DL images, she can confidently evaluate the internal architecture of the nerve — something she couldn’t routinely see before.
“In my opinion, we are seeing better image quality and faster radiology reads. This will help us be more confident in our diagnosis,” she adds.
With AIRTM Recon DL, the power of deep learning and neural networks is unleashed in MR image reconstruction. AIRTM Recon DL was designed to improve SNR and image sharpness, thereby improving image quality in MR exams.
Beyond enhancing image quality, AIRTM Recon DL complements GE’s AIR xTM automatic prescription and AIR TouchTM workflow tools to help improve scan consistency and usability, and potentially help facilitate shorter scan times.
Based on initial evaluations at HSS and CCN, AIRTM Recon DL demonstrates that it can provide high-quality images across a variety of anatomies and scan protocols and has the potential to reduce scan times while preserving high image quality for more efficient exams.
Editor’s note: The editors gratefully acknowledge the assistance of R. Marc Lebel, PhD, Lead Scientist, Julie Poujol, PhD, Research Scientist and Anja C.S. Brau, PhD, General Manager, MR Collaboration & Development, in the development of this article.
AIRTM Recon DL gallery
IS-AIRecon_Gallery_Air_Recon_DL_LEFT.jpg
IS-AIRecon-Gallery-Air-Recon-DL-RIGHT.jpg
AIR™ Recon DL recovers a high-quality image from an otherwise noisy thin-slice axial T2 prostate image acquired in only 1:07 min.
IS-AIRecon_Gallery_Air_Recon_DL_ORIGINAL_IMAGE.jpg
Original image
IS-AIRecon_Gallery_Air_Recon_DL_MILD_SNR_IMPRO.jpg
Mild SNR improvement
IS-AIRecon_Gallery_Air_Recon_DL_MEDIUM_SNR_IMP.jpg
Medium SNR improvement
IS-AIRecon_Gallery_Air_Recon_DL_MAXIMUM_SNR_IM.jpg
Maximum SNR improvement
With AIR™ Recon DL, images can be reconstructed with mild, medium or maximum SNR improvement for visibly improved image quality compared to the original image.
IS-AIRecon_Gallery_SAGITTAL_STIR_LEFT.jpg
A
IS-AIRecon_Gallery_SAGITTAL_STIR_LEFT_INSET.jpg
A
IS-AIRecon_Gallery_SAGITTAL_STIR_RIGHT_NEW.jpg
B
IS-AIRecon_Gallery_SAGITTAL_STIR_RIGHT_INSET.jpg
B
Sagittal STIR lumbar spine reconstructed without (A) and with (B) AIRTM Recon DL. Image quality in the spinal cord is clearly improved through reduced noise and ringing. 0.8 x 0.9 x 1.5 mm 2:47 min.
END_MARK_Spotlight.png