AIR™ Recon DL PROPELLER for motion-insensitive and distortion-free imaging

AIR™ Recon DL is a commercially available, deep-learning-based MR imaging reconstruction that delivers higher SNR and sharper images while enabling shorter scan times. AIR™ Recon DL is available for all anatomies and supports most 2D and 3D Cartesian acquisitions of k-space with and without partial Fourier and parallel-imaging acceleration.

Unlike conventional DICOM image filters, AIR™ Recon DL improves image quality at the foundational level by making use of raw data to remove image noise and truncation or ringing artifacts. It also allows the user to set a preferred level of SNR improvement and generate images directly at the MR console on scan completion. The clinical benefits of AIR™ Recon DL have been extensively described and evaluated.1,2

AIR™ Recon DL for PROPELLER

AIR™ Recon DL is now US FDA approved for use with 2D PROPELLER (Periodically Rotated Overlapping Parallel Lines with Enhanced Reconstruction). In this article, we focus on the clinical benefits of AIR™ Recon DL with PROPELLER, where SNR, in-plane resolution and scan time are often considered the primary competing factors.

GE Healthcare’s 2D PROPELLER application is a radial (non-Cartesian) FSE-like acquisition technique that can deliver all of the same contrasts and weightings as 2D FSE, including T1-, T2- and PD-weighted imaging. Rather than filling k-space line by line, it completes the k-space acquisition through rotated FSE echo trains, so-called “blades,” acquired at a particular kx/ky angle. The result is a motion-insensitive acquisition that has proven successful in applications that are prone to motion, such as the shoulder and abdomen. In addition, PROPELLER supports DWI, producing the least amount of distortion compared to other methods such as EPI.

In Cartesian acquisitions, k-space is sampled using fixed frequency and phase-encoding directions. Each phase-encoding line is usually acquired only once, making it sensitive to any patient motion that occurs between successive phase-encoding steps. In Cartesian acquisitions, motion results in replicative ghosting artifacts along the phase-encoding direction. When AIR™ Recon DL is used, these ghosting artifacts can become more apparent, due to the denoising performance of the algorithm.

The conventional (Cartesian) FSE images in Figure 1A and 1F show evidence of patient-related motion artifacts in the brain and shoulder due to cerebrospinal fluid pulsation and respiratory motion, respectively. In this example, the artifacts are noticeable but not deleterious. When AIR™ Recon DL is used with the Low setting (Figure 1B and 1G), the motion artifacts become more apparent due to the denoising. The motion artifacts become increasingly conspicuous with the AIR™ Recon DL Medium and High settings (Figure 1C, 1D, 1H, 1I). When 2D PROPELLER is used on the same patient, no motion artifacts are evident (Figure 1E, 1J) and AIR™ Recon DL provides the expected improvement to SNR and image sharpness.

 

Figure 1.

Representative FSE and PROPELLER images in the brain and shoulder in the presence of motion. (A, F) Cartesian 2D FSE shows evident motion artifacts, which become increasingly conspicuous with AIR™ Recon DL (B-D, G-I) with Low, Medium and High settings. No motion artifacts are observed with 2D PROPELLER (E, J). Images A-D and F-I were reconstructed with the same raw data.  

In DWI, an EPI acquisition is commonly used due to its higher SNR and faster acquisition, but it suffers from susceptibility artifacts. As a result, geometric distortions and chemical shift artifacts are commonly observed in DW EPI images in the presence of large B0 inhomogeneities. Multi-shot EPI techniques can significantly reduce susceptibility artifacts, but at the cost of longer scan times. With even longer scan times, DW PROPELLER delivers the least amount of distortion and susceptibility artifacts, such as in the skull base, head/neck and near MR-Conditional metal implants.3 DW PROPELLER’s distortion performance is a result of the FSE acquisition that is less sensitive to B0 inhomogeneities.

Although PROPELLER has many advantages over other acquisition methods for its motion-insensitivity and reduced diffusion-weighted distortion, PROPELLER users need to carefully balance spatial resolution, scan time and SNR. Moreover, when compared to Cartesian FSE and EPI-based techniques, PROPELLER scan times are generally longer with slightly lower spatial resolution, which can hinder clinical adoption.

SNR, spatial resolution and scan time

In the context of 2D FSE, the most common ways to reduce PROPELLER scan times are to reduce the number of signal averages (NEX) and to increase the parallel-imaging (ARC) acceleration factors.4 But, these changes come at the expense of SNR. Generally, the SNR of an MR image varies with the square root of scan time, so a reduction of scan time by a factor of 2 leads to a reduction in SNR by approximately 1.4.

Figure 2 describes the SNR of PROPELLER as a function of scan time as measured in a phantom. In Figure 2D, we see how the measured SNR varies with the expected square root behavior with conventional and the Low, Medium and High settings of AIR™ Recon DL. In this example, the conventional image acquired in 503 seconds (Figure 2A) has an equivalent SNR as the images reconstructed with AIR™ Recon DL Low (286 seconds) and Medium (149 seconds). From the chart in Figure 2D, we might extrapolate to assert that an equivalent SNR could be attained with AIR™ Recon DL High in approximately 50 seconds, thus representing an approximate 10-fold reduction in scan time compared to the conventional-reconstructed image.

In addition, the results in Figure 2D demonstrate that, with a fixed scan time, SNR is significantly improved with AIR™ Recon DL, increasing by an approximate factor of 4, when compared to conventional reconstruction with AIR™ Recon DL High.

 

Figure 2.

PROPELLER phantom image showing SNR as a function of scan time. Shown are (A) conventional reconstruction with a 503 sec. scan time, (B) AIR™ Recon DL Low reconstruction with a 286 sec. scan time, and (C) AIR™ Recon DL Medium reconstruction of a 149 sec scan time, all showing equivalent SNR. (D) The SNR of the four different reconstruction methods with varying scan times. The red horizontal dashed line indicates the SNR level of the image that was acquired in 503 seconds and reconstructed with a conventional reconstruction method.

In addition, the results in Figure 2D demonstrate that, with a fixed scan time, SNR is significantly improved with AIR™ Recon DL, increasing by an approximate factor of 4, when compared to conventional reconstruction with AIR™ Recon DL High.

As described in an earlier article,1 in conventional reconstruction, low-pass or truncation filters are usually applied to suppress Gibbs ringing and noise, but at a cost in image sharpness or spatial resolution. AIR™ Recon DL not only reduces noise, but also suppresses truncation artifacts or Gibbs ringing, leading to sharper images with better defined edges. A resolution phantom was scanned using PROPELLER with different matrix sizes (in-plane resolutions) and reconstructed with both conventional reconstruction and AIR™ Recon DL High (Figure 3). The PROPELLER resolution is dramatically improved with AIR™ Recon DL and found to be comparable with FSE. For example, the AIR™ Recon DL 0.52 x 0.52 mm in-plane voxel image approximately matches the conventional 0.31 x 0.31 mm image.

PROPELLER, Conventional

 

PROPELLER, AIR™ Recon DL

 

Figure 3.

PROPELLER image sharpness improvement with AIR™ Recon DL. Shown are (A) PROPELLER with conventional reconstruction and (B) PROPELLER with AIR™ Recon DL reconstruction at different in-plane resolutions. The diameter of the pin holes is approximately 1.1 mm.

AIR™ Recon DL for DW PROPELLER

One of the challenges in FSE-based DWI is that the random phase arising from motion-sensitizing diffusion-weighting gradients causes a violation of the so-called Carr-Purcell-Meiboom-Gill (CPMG) condition, resulting in destructive interference of spin-echo and stimulated echo signals, leading to unwanted shading and artifacts. Phase-insensitive preparation, or splitting CPMG and non-CPMG components (DW Duo), can be applied to mitigate this issue with DW FSE acquisitions but reduces SNR.

The reduced SNR in DW PROPELLER leads to an increase in the ADC measurement variation. As such, users must increase NEX (scan time) or reduce in-plane resolution to improve the SNR, which either increases the scan time or image blurriness. AIR™ Recon DL provides another alternative to improve the SNR of DW images and ADC measurements without compromising in-plane resolution or increasing scan time.

The benefit of AIR™ Recon DL on the SNR of DW PROPELLER is depicted in the ADC measurements of a phantom (Figure 4). Compared to conventional reconstruction, no statistically significant difference was found in the mean ADC. However, the standard deviations of the ADC measurements were significantly smaller with AIR™ Recon DL SNR improvement levels, thus demonstrating an improvement to the ADC measurement precision.

 

Figure 4.

The effect of AIR™ Recon DL on DW PROPELLER ADC measurements. Shown are regions of interest (ROIs) for ADC analysis of a phantom for conventional and AIR™ Recon DL reconstructions. AIR™ Recon DL reconstructions demonstrate improved ADC measurement precision, compared to conventional reconstruction of the same raw data.

The in-plane resolution of DWI is often compromised to reduce scan time or increase SNR, resulting in blurry images. However, AIR™ Recon DL’s Intelligent Ringing Suppression can improve the sharpness of diffusion-weighted images. As shown in Figure 5, the original diffusion-weighted image shows strong Gibbs ringing, which is not visually evident in the AIR™ Recon DL reconstruction of the same raw data. To quantify this improvement in spatial resolution, the signal profile across the ventricle is plotted (Figure 5C), showing sharper signal variations at ventricle boundaries with AIR™ Recon DL. From the same data in Figure 5C, a slope or gradient can be calculated to further demonstrate that AIR™ Recon DL delivers sharper structures, compared to conventional reconstruction.

 

Figure 5.

The effect of AIR™ Recon DL on DW PROPELLER image sharpness. Representative diffusion-weighted PROPELLER images in the brain (A) without and (B) with AIR™ Recon DL. (C) Line profiles through the indicated region in the brain, and the (D) gradient of the line profile demonstrates improved spatial resolution with AIR™ Recon DL.

Clinical benefits and early adopter feedback

AIR™ Recon DL PROPELLER has been evaluated internally and externally across many different anatomies, showing considerable clinical benefit over conventional reconstruction with improved image quality and reduced scan time. Overall, the clinical benefits are best demonstrated with images that span multiple anatomies and systems.

In Figure 6, noise and Gibbs ringing compromise the visualization of gray matter along the central canal in the conventional T2-weighted cervical spine image, while it is better visualized in the AIR™ Recon DL image due to efficient noise and Gibbs ringing suppression.

 

Figure 6.

AIR™ Recon DL PROPELLER in the cervical spine at 3.0T. Shown are (A) conventional and (B) AIR™ Recon DL High PROPELLER reconstructed images in the same patient. The AIR™ Recon DL images show significant SNR improvement and Gibbs ringing suppression.

As previously demonstrated in Figure 3, AIR™ Recon DL can improve image sharpness and the visualization of small structures. Figure 7 shows representative abdominal PROPELLER images at 1.5T, showing higher spatial resolution in the pancreatic duct with AIR™ Recon DL in addition to improved SNR.

 

Figure 7.

AIR™ Recon DL PROPELLER in the abdomen at 1.5T. Shown are (A) conventional and (B) AIR™ Recon DL High PROPELLER reconstructed images demonstrating sharper images and higher SNR with improved visualization of the pancreatic duct.

With optimized protocols, AIR™ Recon DL can reduce the scan time, leading to shorter exam times and increased patient throughput. Figure 8 shows representative shoulder images acquired with a routine protocol and fast protocol at 3.0T. The scan time was reduced from 3:15 min to 1:30 min by reducing NEX from 2.5 to 1.5 and increasing ARC parallel-imaging factor from 2 to 3. The overall image quality of the fast protocol with AIR™ Recon DL is slightly better than the long routine protocol with conventional reconstruction.

 

Figure 8.

Scan time reduction with AIR™ Recon DL optimized protocols. Shown is an axial PD with fat suppression at 3.0T with (A) conventional reconstruction with 3:15 min scan time and (B) AIR™ Recon DL High with 1:30 min scan time, using an optimized protocol. The faster AIR™ Recon DL protocol shows a subchondral lesion (yellow arrow) and a small glenoid cartilage defect (blue arrow) that were not well defined on the routine protocol.

For high-resolution 7.0T imaging, AIR™ Recon DL further improves the definition of small structures compared to the conventional reconstruction by improving SNR and sharpness (Figure 9).

 

Figure 9.

High-resolution 7.0T imaging with AIR™ Recon DL PROPELLER. T2w PROPELLER images with 200 um in-plane resolution and 1 mm slice thickness, reconstructed with (A) conventional reconstruction and (B) AIR™ Recon DL reconstruction, demonstrating substantially improved SNR and anatomical conspicuity.

Compared to multi-shot DW EPI (MUSE) (Figure 10A), DW PROPELLER results in less geometric distortion (Figure 10B). AIR™ Recon DL significantly improves the sharpness and SNR of DW PROPELLER (Figure 10C).

b0

b1000

 

Figure 10.

Image quality improvement of DW PROPELLER with AIR™ Recon DL. Multi-shot DW EPI showed significant geometric distortion in the skull base (red arrows). The conventional DW PROPELLER suffers from low SNR and in-plane resolution. AIR™ Recon DL significantly improves the SNR and in-plane resolution of DW PROPELLER.

Summary

AIR™ Recon DL is now compatible with PROPELLER and delivers significantly reduced sensitivity to motion artifacts and eliminates diffusion-related distortions in DWI. AIR™ Recon DL improves the image quality of PROPELLER and DW PROPELLER with increased SNR and sharpness, shorter scan time and improved DWI,5-16 as demonstrated in a phantom and in vivo. These benefits were also confirmed by early adopter clinical sites with access to a research prototype used within institution review board guidelines.

This latest development represents GE’s continued emphasis to expand proven technologies to clinical areas where the clinical need is evident.

References

1. The clinical benefits of AIR™ Recon DL for MR image reconstruction. September 2020. https://www.gehealthcare.com/-/jssmedia/c943df5927a049bb9ac95a9f0349ad8c.pdf.
2. Lebel RM. Performance characterization of a novel deep learning-based MR image reconstruction pipeline. August 2020. http://arxiv.org/abs/2008.06559.
3. Diffusion imaging demystified. SIGNA Pulse 2018, Autumn. 25:60-65.
4. Peters RD, Blahnik H, Harris H, Lawson S. Practical protocol conversion and optimization with AIR Recon DL. GE SIGNA Pulse of MR, Autumn 2020. 29:81-83.
5. Wang X, Litwiller D, Lebel M, et al. High Resolution T2W imaging using Deep Learning Reconstruction and Reduced Field-of-View PROPELLER. Proceedings of the ISMRM. 2020.
6. Wang X, Litwiller D, Ersoz A, et al. Diffusion Weighted Imaging using PROPELLER Acquisition and a Deep Learning based Reconstruction. Proceedings of the ISMRM. 2020.
7. Poujol J, Henry C, Barrau V, et al. Learning how to adapt T2 PROPELLER MR prostate imaging: going beyond PIRADS requirements with MR Deep Learning reconstruction.
8. Wang X, Bayram E, Litwiller D, et al. PROPELLER Diffusion-Weighted Imaging of the Prostate with Deep-Learning Reconstruction. Proceedings of the ISMRM. 2020.
9. Pirasteh A, Estkowski L, Litwiller D, Bayram E, Wang X. Motion Robust High-Resolution Rectal Imaging using PROPELLER and Deep Learning Reconstruction. Proceedings of the ISMRM. 2021.
10. Saleh M, Javadi S, Mathew M, et al T2-weighted Pelvic MR Imaging Using PROPELLER with Deep Learning Reconstruction for Improved Motion Robustness. Proceedings of the ISMRM. 2021.
11. Wiesinger F, Kaushik S, Engstrom M, et al. Deep Learning-based MR-only Radiation Therapy Planning for Head&Neck and Pelvis. Proceedings of the ISMRM. 2021.
12. Song Y-S, Lee I-S, Hwang M-J, Jang K, Fung M, Wang X. Axial T2 weighted MR Imaging of the Cervical Spin using PROPELLER with Deep Learning Reconstruction to improve image quality. Proceedings of the ISMRM. 2022.
13. Wang X, Ersoz A, Litwiller D, Ma J, Stafford J, Bayram E. Robust Diffusion Weighted Imaging with Deep Learning-Based DW PROPELLER Reconstruction. Proceedings of the ISMRM. 2022.
14. Yi J, Lee H-J, Hahn S, Lee J, Wang X, Fung M. Shoulder PROPELLER MRI with Deep learning-based reconstruction: image quality and agreement between standard and accelerated sequence. Proceedings of the ISMRM. 2022.
15. Litwiller DV, Wang X, Lebel RM, et al. Improving Motion-Robust Structural Imaging at 7T with Deep Learning-Based PROPELLER Reconstruction. Proceedings of the ISMRM. 2022.
16. Carretero L, Wang X, Sánchez E, et al. Improving diagnostic confidence using a Deep- Learning Reconstructed Fast Motion-Robust PROPELLER protocol for Shoulder Imaging. Proceedings of the ISMRM. 2022.