Diffusion weighted imaging (DWI) has become a workhorse sequence for research and clinical imaging. In research, diffusion provides a unique ability to shed light on microscopic processes underlying normal brain development and neurodegenerative diseases. In the clinic, DWI is a crucial diagnostic tool to allow fast detection of malignant lesions that exhibit high diffusion restriction and tend to appear brighter than surrounding tissue at high b-values (Figure 1). As such, DWI is now a fundamental sequence recommended in clinical guidelines for many indications (e.g., stroke, cancer).1-3
Figure 1.
A 60-year-old female experiencing memory decline for several weeks was referred to MR. Isotropic 1.6 mm DWI with AIR™ Recon DL Phase Correction at (A) b1000, (B) b3000 and (C) b10000. Note the thin-slice, high b-value diffusion shows diffuse cortical high signal (C, arrows). Patient was diagnosed with Creutzfeldt-Jakob disease.
Diffusion happens at various scales in vivo and the maximum “b-value” dictates the range at which these microscopic processes can be resolved. With the advent of high- and ultra-high-performance gradient systems, higher b-values can now be routinely achieved either for advanced neuroimaging (SIGNA™ MAGNUS‡), or for common clinical applications (SIGNA™ Premier 3.0T). Regardless of the MR system being used, diffusion has inherently low SNR, especially when used with high b-value or high spatial resolution acquisitions.
To address this, multiple excitations (NEX) are performed to generate multiple images, which are then averaged to improve SNR and lesion conspicuity. Each MR image is intrinsically complex, with each pixel value composed of a magnitude and phase component. However, because the images are acquired separately, each image will have a different phase, primarily due to bulk motion during diffusion sensitizing gradients and, to a lesser extent, systemic noise which, if left uncorrected, can lead to destructive signal interferences in the final combined image.
For conventional reconstructions, this phase is estimated using a low-pass filtering approach, which can be limiting as it does not properly account for high-frequency phase variations. This could lead to phase cancellation artifacts (Figure 2), or an increased noise bias. In diffusion imaging, this phase estimation is further complicated by the higher noise stemming from the high b-values and concomitantly longer echo times, which work to lower the effective signal. The lower SNR can obscure the high-frequency phase variations in the images, without which, the averaging of images will create a loss of signal and lead to a “wormhole” artifact that appears as dark signal “encroaching” into areas where it should not be seen (Figure 3B, arrow). Additionally, if the high-frequency phase cannot be distinguished from the noise prior to phase correction, the noise statistics take on a Rician distribution, rather than Gaussian. Consequently, when the final magnitude images are generated, the mean of the noise will be non-zero, creating a “noise floor” that can hide pathology and reduce contrast. This increased noise floor will impact the accuracy of apparent diffusion coefficient (ADC) measurements, which can have broader clinical consequences.
Figure 2.
(A) Conventional phase correction failed to correct respiratory motion-induced high-frequency phases in the pancreatic head, causing signal dropouts in the b700 image and, therefore, (C) the overestimation of ADC. (B) The robust AIR Recon DL Phase Correction minimized the signal dropouts and (D) corrected the ADC bias.5
Figure 3.
DWI neuro images at b5000 with complex averaging using filter-based and AIR Recon DL Phase Correction. (A) The original DWI image with filter-based phase correction and complex averaging shows a higher noise floor than (C) AIR Recon DL Phase Correction and complex averaging. (B, D) In the zoomed images, (B) the original DWI image suffers from wormhole artifacts due to imperfect phase correction (arrow), while (D) AIR Recon DL Phase Correction reduces wormhole artifacts (arrow).4
Why use deep learning?
AIR™ Recon DL is GE HealthCare’s deep-learning-based reconstruction algorithm for MR imaging that improves image quality at the foundational level by making use of the raw data to remove image noise and ringing, increasing SNR and image sharpness by up to 60%, and enabling shorter scan times. It has been heralded as a significant innovation in MR technology that addresses the compromise between SNR and scan time.
AIR Recon DL Phase Correction‡‡ builds on the success of previous AIR Recon DL releases and consists of two components—a deep-learning-based estimate of image phase (DLPC), and a neural network to remove the noise and truncation artifacts in the images (ARDL). The DLPC network was trained to cover a wide variety of linear and non-linear image phase variations that may arise from a multitude of sources (motion, eddy currents, field inhomogeneities, etc.). Neural networks are very efficient in recognizing these linear and non-linear phase patterns, better than traditional signal processing techniques. Once the neural network is trained, inferencing with the GPU is extremely fast. Consequently, the DLPC network creates an accurate phase map to help solve the critical problem of accurate complex image combination. This helps keep the noise distribution Gaussian and lowers the noise floor to reveal features that were previously buried in the noise. This process further improves the efficiency of AIR Recon DL, which is used in combination to remove the noise and truncation artifacts. The improved SNR in the diffusion images could be traded off to acquire images at higher b-values4 or even at lower field strengths, which we believe increase the broad clinical utility of the technique.
Technical validation
To validate AIR Recon DL Phase Correction, we first tested on digital phantoms, where it was possible to simulate a range of phase variations (Figure 4). With this ground truth, AIR Recon DL Phase Correction was compared with the current product using a low-pass filter-based method. For all variations of the simulated phase, the test demonstrated that AIR Recon DL Phase Correction always outperformed the current method.4
Figure 4.
Digital phantom results demonstrating the bias of phase correction using (A) a conventional phase correction method compared to (B) AIR Recon DL Phase Correction.
The technique was further tested using a NIST diffusion phantom.5 This phantom has vials with known temperature-controlled ADCs, which permits accurate testing of diffusion images after the application of AIR Recon DL Phase Correction. Figure 5 shows the log-linear signal decay for different b-values. For conventional reconstructions—using either magnitude averaging or a low-pass filter-based phase estimate—the increased noise floor causes a deviation from the expected linear decay. This “noise bias” results in underestimated ADC maps. With AIR Recon DL Phase Correction, this log-linear trend was maintained at high b-values, which minimizes the noise bias and produces more accurate estimates of the ADC.5
Figure 5.
NIST phantom results. Signal curves and ADC values of NIST diffusion phantom samples at (A, B) 0% and (C, D) 10%. The ADC values were estimated from different b-value pairs that were reconstructed with magnitude averaging, conventional complex averaging and DLPC complex averaging with AIR Recon DL denoising.
After validating the technique in phantoms, it was tested in vivo for body and neuro diffusion imaging. The neuro example (Figure 6) shows a conventional high b-value diffusion image with a high noise floor. The images acquired with AIR Recon DL Phase Correction reduce the noise floor and provide significantly improved SNR and overall image quality for the diffusion images. For the diffusion images acquired of the female pelvis (Figure 7)—an often challenging anatomy to acheive high image quality—capturing the high-frequency phase avoids signal cancellation and wormhole artifacts, thus providing an improvement in contrast and image sharpness with AIR Recon DL Phase Correction.
/
Figure 6.
(A) Conventional phase correction on a Discovery™ MR750w 3.0T, b5000, and (B) same image with AIR Recon DL reconstruction and AIR Recon DL Phase Correction. (C) Conventional phase correction on SIGNA MAGNUS 3.0T system, b20000, and (D) same image with AIR Recon DL reconstruction and AIR Recon DL Phase Correction.
Figure 7.
(A, B) Female pelvis, 2 x 2 x 4 mm, b1000, 6 NEX, 2 shots, (A) with AIR Recon DL reconstruction and (B) same image with AIR Recon DL Phase Correction. (C, D) Male prostate, 1.3 x 1.3 x 3 mm, b1000, 3 NEX, 4 shots, (C) with AIR Recon DL reconstruction and (B) same image with AIR Recon DL Phase Correction.
Summary
Deep-learning-based phase correction significantly improves the robustness and image quality of diffusion-weighted MR. It enhances diffusion contrast by enabling more accurate complex signal averaging, reducing the noise floor and minimizing artifacts. These benefits afford the acquisition of images at higher b-values, higher quality synthetic diffusion images and ADC maps of higher accuracy. This feature significantly improves the robustness of diffusion imaging, giving radiologists a better diagnostic tool.
‡ Not yet CE marked. Not available for sale in all regions.
‡‡ 510(k) pending at U.S. FDA. Not yet CE marked. Not available for sale.
References
- Latchaw, Richard E., Mark J. Alberts, Michael H. Lev, et al. 2009. “Recommendations for Imaging of Acute Ischemic Stroke: A Scientific Statement From the American Heart Association.” Stroke 40 (11). https://doi.org/10.1161/STROKEAHA.108.192616.
- Cogswell, Petrice M., Jerome A. Barakos, Frederik Barkhof, et al. 2022. “Amyloid-Related Imaging Abnormalities with Emerging Alzheimer Disease Therapeutics: Detection and Reporting Recommendations for Clinical Practice.” American Journal of Neuroradiology 43 (9): E19–E35. https://doi.org/10.3174/ajnr.A7586.
- Boss, Michael A., Dariya Malyarenko, Savannah Partridge, et al. 2024. “The QIBA Profile for Diffusion-Weighted MRI: Apparent Diffusion Coefficient as a Quantitative Imaging Biomarker.” Radiology 313 (1): e233055. https://doi.org/10.1148/radiol.233055.
- Wang, Xinzeng, Patricia Lan, Arnaud Guidon. 2024. “DL-based Phase Correction Enables Robust Real Diffusion-Weighted MRI with Increased Diffusion Contrast.” Proceedings of the International Society for Magnetic Resonance in Medicine 2414. https://archive.ismrm.org/2024/2414.html.
- Wang, Xinzeng, Patricia Lan, Kang Wang, Ante Zhu, Abad Nastaren, Arnaud Guidon. 2025. “Deep Learning based Phase Correction and Denoising for Accurate ADC Quantification.” Proceedings of the International Society for Magnetic Resonance in Medicine 4197. https://archive.ismrm.org/2025/4197_6nYyDQQTd.html.
