Portal de Periódicos da CAPES

Deep Learning for Optimization of Abdominopelvic 4D Flow MRI Analysis

2021; Radiological Society of North America; Volume: 302; Issue: 3 Linguagem: Inglês

10.1148/radiol.212702

1527-1315

Alejandro Roldán‐Alzate, Thomas M. Grist,

Advanced MRI Techniques and Applications

HomeRadiologyVol. 302, No. 3 PreviousNext Reviews and CommentaryFree AccessEditorialDeep Learning for Optimization of Abdominopelvic 4D Flow MRI AnalysisAlejandro Roldán-Alzate , Thomas M. GristAlejandro Roldán-Alzate , Thomas M. GristAuthor AffiliationsFrom the Departments of Radiology (A.R., T.M.G.) and Mechanical Engineering (A.R.), University of Wisconsin–Madison, 1111 Highland Ave, Madison, WI 53705.Address correspondence to A.R. (e-mail: [email protected]).Alejandro Roldán-Alzate Thomas M. GristPublished Online:Nov 30 2021https://doi.org/10.1148/radiol.212702MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by You and Masutani et al in this issue.Alejandro Roldán-Alzate, PhD, is an assistant professor in the Departments of Radiology and Mechanical Engineering at the University of Wisconsin–Madison. He is the director of the Cardiovascular Fluid Dynamics Laboratory, which has the main interest of coupling engineering tools with medical imaging to noninvasively characterize the hemodynamics in different physiologic and pathologic conditions.Download as PowerPointOpen in Image Viewer Thomas M. Grist, MD, is the John H. Juhl professor of radiology and medical physics and is the chair of the Department of Radiology at the University of Wisconsin–Madison. He is a past president of the International Society for Magnetic Resonance in Medicine and past chair of the Radiological Society of North America R&E foundation.Download as PowerPointOpen in Image Viewer Assessment of abdominopelvic vasculature is crucial for many clinical diagnoses, but it is challenging because of the tremendous variability and complexity of anatomy, physiology, and disease. Doppler US is the most commonly used diagnostic tool for assessment of abdominopelvic vasculature (1). Doppler US may be limited by overlying intestinal gas and a limited acoustic window, which can lead to an incomplete or inaccurate characterization of the abdominopelvic hemodynamics. Two-dimensional (2D) phase-contrast MRI is an alternative to US. Selection of individual vessels for 2D phase-contrast MRI requires operator experience; the technologist needs to coordinate MRI scan acquisition with patient breath holding. Numerous 2D planes are needed for comprehensive flow evaluation of the abdominal vasculature, leading to long scanning times and often requiring a radiologist at the scanner. Therefore, 2D phase-contrast MRI flow is often not part of the clinical routine.An alternative to 2D phase-contrast MRI is time-resolved three-dimensional (3D) phase-contrast MRI with three-directional flow encoding, also referred to as four-dimensional (4D) flow MRI. The 4D flow MRI technique can be used to evaluate vascular anatomy and the rate and direction of blood flow (2–4). The 4D flow technique is attractive in concept because, unlike 2D phase-contrast measurements, images can be acquired in one acquisition and then retrospectively processed to measure flow in any vessel and any orientation.Unfortunately, the quantitative flow information from 4D flow MRI can be degraded by several artifacts, including magnetic eddy current–related background phase errors (5). The correction of this error remains a challenge. One approach is to use a stationary phantom requiring an additional scan after the patient scan using identical imaging parameters. That approach is impractical for clinical implementation (6). Previous studies have explored the effects of different parameters on the accuracy of phase error correction using manual segmentation of images. These studies showed that the quality of the correction is directly affected by the percentage of static tissue segmented for the correction.Because of these uncorrected background phase errors, the entire quantitative information from 4D flow MRI becomes suspect—the numbers may simply not make sense. For example, the sum of the flow in the proximal common iliac arteries may not properly add up to be equal to the flow in the distal aorta. This frustrating situation reduces clinician confidence in measurements obtained with 4D flow MRI. Likewise, the process of manual background phase correction lends itself to bias; the technologist or radiologist can choose various background regions, but which is “correct?” How is the region of interest defined? As a result, this incredibly promising 4D flow technique, which is simple to apply, becomes complex to implement for abdominal MRI.In this issue of Radiology, You and Masutani et al present a proof-of-concept study evaluating the feasibility of using a fully automated deep learning algorithm to perform image-based background phase error correction in 4D flow MRI and compared its effectiveness relative to the tedious process of manual image-based correction (7). Specifically, a convolutional neural network was trained to automatically generate estimates of background phase error maps to replace the current standard workflow that relies on semiautomatic segmentation of static tissue and exclusion of vessels. This approach would potentially remove the need for technologist or radiologist intervention and thus simplify the analysis and interpretation of abdominopelvic 4D flow MRI studies.You and Masutani et al completed 139 clinical abdominopelvic 4D flow MRI studies at 3 T using a 3D cartesian pulse sequence. Manual correction of background phase error was performed via segmentation of static tissue followed by patchwise linear regression of static tissue velocities. Internal consistency analysis was performed by using the principle of conservation of total flow at bifurcations (inflow-outflow consistency).Using the principle of conservation of total flow (vessel inflow must be equal to outflow), the authors found poor consistency in the flow measurements prior to phase error correction (mean percentage difference, 37% ± 26). The consistency improved after manual correction (mean, 14% ± 10), as expected. The performance of the proposed convolutional neural network–based approach was comparable to that of the standard manual approach. The average total time for automated phase error correction was 12 seconds per case on average, while manual correction by an experienced user took about 2.5 minutes per case. Direct comparison of flow measurements using manual and automated correction demonstrated excellent correlation (ρ = 0.98, P < .001).This new study by You and Masutani et al is an interesting application of a deep learning technique to improve the postprocessing workflow by reducing the need for manual segmentation. As explained previously, current correction methods for 4D flow require extensive time and expertise, given the need for manual anatomic segmentation, making it a challenge to implement in a clinical setting. Automatic correction of the phase errors of 4D flow MRI data in a fraction of the time as manual correction is a substantial advance; this approach may finally open the door to more routine clinical use of 4D flow methods.Other studies on deep learning techniques applied to 4D flow MRI have been focused on improvement of the resolution and quality of 4D flow estimates in the different vascular territories (8). Augmentation of 4D flow MRI data with computational fluid dynamics–informed training networks produces enhanced physiologic flow fields by using high-spatial-resolution and high-temporal-resolution patient-specific computational fluid dynamics data to train a neural network. The trained network is then used to create more accurate MRI-derived cerebrovascular velocity fields. Through testing on simulated images, phantom data, and in vivo human subject data, the trained network successfully denoised flow images and reduced velocity error, particularly near vessel walls (9). Such image correction has the potential to significantly improve both experimental and clinical (qualitative and quantitative) phase-contrast MRI analysis of cerebrovascular flow.Although phase error correction is a key step in 4D flow data preprocessing, other factors make it difficult to implement 4D flow MRI in a clinical setting. These factors include relatively long scanning times (up to 10 minutes). The flow velocities must also be selected ahead of time to prevent velocity aliasing when set too low or poor signal-to-noise ratio when set too high. In the study by You and Masutani et al, only phase error correction was taken into account when training the neural network. In theory, deep learning may help correct aliasing artifacts and reduce noise in phase-contrast images. In terms of the methods used for comparison, the principle of conservation of total flow (inflow vs outflow) was used as the main comparison parameter. One limitation is that this method does not account for collateral vessels that might be present in certain abnormal conditions in the abdominal circulation or blood flow in small vessels not depicted by MRI.Convolutional neural networks can be further developed to shorten MRI scan times and correct the velocity aliasing seen in settings of unexpectedly high flow. As stated by the authors, additional studies are needed to investigate results for other body territories using different scanners from other vendors and at other institutions.Disclosures of conflicts of interest: A.R. disclosed no relevant relationships. T.M.G. stock in Elucent, Histosonice, and Shine Medical; research support from GE, Bracco, Siemens, Hologic, and Change Healthcare; past chair of the Radiological Society of North America R&E foundation.References1. Dyverfeldt P, Bissell M, Barker AJ, et al. 4D flow cardiovascular magnetic resonance consensus statement. J Cardiovasc Magn Reson 2015;17(1):72. Crossref, Medline, Google Scholar2. Markl M, Frydrychowicz A, Kozerke S, Hope M, Wieben O. 4D flow MRI. J Magn Reson Imaging 2012;36(5):1015–1036. Crossref, Medline, Google Scholar3. Roldán-Alzate A, Francois CJ, Wieben O, Reeder SB. Emerging Applications of Abdominal 4D Flow MRI. AJR Am J Roentgenol 2016;207(1):58–66. Crossref, Medline, Google Scholar4. Rutkowski DR, Barton GP, François CJ, Aggarwal N, Roldán-Alzate A. Sex Differences in Cardiac Flow Dynamics of Healthy Volunteers. Radiol Cardiothorac Imaging 2020;2(1):e190058. Link, Google Scholar5. Callaghan FM, Burkhardt B, Geiger J, Valsangiacomo Buechel ER, Kellenberger CJ. Flow quantification dependency on background phase correction techniques in 4D-flow MRI. Magn Reson Med 2020;83(6):2264–2275. Crossref, Medline, Google Scholar6. Hofman MBM, Rodenburg MJA, Markenroth Bloch K, et al. In-vivo validation of interpolation-based phase offset correction in cardiovascular magnetic resonance flow quantification: a multi-vendor, multi-center study. J Cardiovasc Magn Reson 2019;21(1):30. Crossref, Medline, Google Scholar7. You S, Masutani EM, Alley MT, et al. Deep Learning Automated Background Phase Error Correction for Abdominopelvic 4D Flow MRI. Radiology 2022;302(3):584–592. Link, Google Scholar8. Retson TA, Besser AH, Sall S, Golden D, Hsiao A. Machine Learning and Deep Neural Networks in Thoracic and Cardiovascular Imaging. J Thorac Imaging 2019;34(3):192–201. Crossref, Medline, Google Scholar9. Rutkowski DR, Roldán-Alzate A, Johnson KM. Enhancement of cerebrovascular 4D flow MRI velocity fields using machine learning and computational fluid dynamics simulation data. Sci Rep 2021;11(1):10240. Crossref, Medline, Google ScholarArticle HistoryReceived: Oct 25 2021Revision requested: Nov 1 2021Revision received: Nov 8 2021Accepted: Nov 12 2021Published online: Nov 30 2021Published in print: Mar 2022 FiguresReferencesRelatedDetailsAccompanying This ArticleDeep Learning Automated Background Phase Error Correction for Abdominopelvic 4D Flow MRINov 30 2021RadiologyRecommended Articles Deep Learning Automated Background Phase Error Correction for Abdominopelvic 4D Flow MRIRadiology2021Volume: 302Issue: 3pp. 584-592Phase-Contrast MRI: Physics, Techniques, and Clinical ApplicationsRadioGraphics2020Volume: 40Issue: 1pp. 122-140Optimizing Image Quality When Evaluating Blood Flow at Doppler US: A TutorialRadioGraphics2019Volume: 39Issue: 5pp. 1501-1523How Well Does an Automated Approach Calculate and Visualize Blood Flow Vorticity at 4D Flow MRI?Radiology: Cardiothoracic Imaging2020Volume: 2Issue: 13D Super-Resolution US Imaging of Rabbit Lymph Node Vasculature in Vivo by Using MicrobubblesRadiology2019Volume: 291Issue: 3pp. 642-650See More RSNA Education Exhibits Understanding Doppler Ultrasound: Technique, Artifacts And InterpretationDigital Posters2021Sonography Made Simple: A Case-based ReviewDigital Posters2021Doppler Artifacts and Pitfalls: Made Easy to UnderstandDigital Posters2020 RSNA Case Collection Cerebral Arteriovenous MalformationRSNA Case Collection2021Cystic adventitial disease of the popliteal arteryRSNA Case Collection2020Subclavian Stenosis with Pre-StealRSNA Case Collection2021 Vol. 302, No. 3 Metrics Altmetric Score PDF download