Rapid Reconstruction of Four-dimensional MR Angiography of the Thoracic Aorta Using a Convolutional Neural Network

Published Online:https://doi.org/10.1148/ryct.2020190205

Abstract

Purpose

To implement an integrated reconstruction pipeline including a graphics processing unit (GPU)–based convolutional neural network (CNN) architecture and test whether it reconstructs four-dimensional non-Cartesian, non–contrast material–enhanced MR angiographic k-space data faster than a central processing unit (CPU)–based compressed sensing (CS) reconstruction pipeline, without significant losses in data fidelity, summed visual score (SVS), or arterial vessel–diameter measurements.

Materials and Methods

Raw k-space data of 24 patients (18 men and six women; mean age, 56.8 years ± 11.8 [standard deviation]) suspected of having thoracic aortic disease were used to evaluate the proposed reconstruction pipeline derived from an open-source three-dimensional CNN. For training, 4800 zero-filled images and the corresponding CS-reconstructed images from 10 patients were used as input-output pairs. For testing, 6720 zero-filled images from 14 different patients were used as inputs to a trained CNN. Metrics for evaluating the agreement between the CNN and CS images included reconstruction times, structural similarity index (SSIM) and normalized root-mean-square error (NRMSE), SVS (3 = nondiagnostic, 9 = clinically acceptable, 15 = excellent), and vessel diameters.

Results

The mean reconstruction time was 65 times and 69 times shorter for the CPU-based and GPU-based CNN pipelines (216.6 seconds ± 40.5 and 204.9 seconds ± 40.5), respectively, than for CS (14 152.3 seconds ± 1708.6) (P < .001). Compared with CS as practical ground truth, CNNs produced high data fidelity (SSIM = 0.94 ± 0.02, NRMSE = 2.8% ± 0.4) and not significantly different (P = .25) SVS and aortic diameters, except at one out of seven locations, where the percentage difference was only 3% (ie, clinically irrelevant).

Conclusion

The proposed integrated reconstruction pipeline including a CNN architecture is capable of rapidly reconstructing time-resolved volumetric cardiovascular MRI k-space data, without a significant loss in data quality, thereby supporting clinical translation of said non–contrast-enhanced MR angiograms.

Supplemental material is available for this article.

Keywords: Adults, Angiography, MR-Angiography, Vascular

© RSNA, 2020

Summary

A graphics processing unit–based convolutional neural network architecture is capable of fast reconstruction (<5 minutes) of accelerated four-dimensional non-Cartesian, non–contrast-enhanced MR angiographic k-space data without losses in data fidelity, summed visual score, or vessel-diameter measurements.

Key Points

  • ■ The mean reconstruction time was 65 times shorter for the central processing unit (CPU)–based convolutional neural network (CNN) pipeline (mean = 216.6 seconds ± 40.5, range = 179.1–302.5 seconds, P < .01) and 69 times shorter for the graphics processing unit–based CNN pipeline (mean = 204.9 seconds ± 40.5, range, 167.6−291.4 seconds, P < .01) compared with the CPU-based compressed sensing (CS) pipeline (mean reconstruction time = 14 152.3 seconds ± 1707.6, range = 11 709.3–18 905.0 seconds).

  • ■ The average median reader summed visual scores (3 = nondiagnostic, 9 = clinically acceptable, 15 = excellent) for the CNN were deemed above the acceptable (11.6) cut point and were not significantly different (P > .25) compared with CS (12.1).

  • ■ The mean vessel diameters were not significantly different (P > .27) between the CNN and CS images for all seven standardized locations of the thoracic aorta, except at the proximal descending aorta (P = .02), where the percentage difference in diameter was only 3% (ie, clinically irrelevant).

Introduction

Contrast material–enhanced MR angiography is used routinely in radiology practice for evaluation of thoracic aortic aneurysms. Unfortunately, contrast-enhanced MR angiography is contraindicated in patients with renal insufficiency because of the risk for nephrogenic systemic fibrosis (1) and in patients with poor intravenous access. In addition, there is a growing concern for the unknown health effects of gadolinium accumulation in the brain after repeated exposure to a gadolinium-based contrast agent (2). This concern is particularly important for younger patients with genetic disorders (Marfan, Ehler Danlos, or Loeys-Dietz syndrome) predisposing them to pathologic aortic conditions and for patients with bicuspid aortic valve disease who require frequent monitoring of thoracic aortic dimensions. Non–contrast-enhanced MR angiography (3,4) is an alternative test that has several advantages over contrast-enhanced MR angiography, including not requiring a gadolinium-based contrast agent and permitting rescanning during the same MRI examination as needed.

A previous study (5) extended the “coronary” MR angiographic pulse sequence by incorporating radial stack-of-star k-space sampling and extra motion-state golden-angle radial sparse parallel (XD-GRASP) MRI (6) reconstruction with self-navigation of respiratory motion to produce high-quality respiratory motion–resolved thoracic non–contrast-enhanced MR angiograms at high spatial resolution (1.5 × 1.5 × 1.5 mm) and with predictable scan time (6 minutes) in patients suspected of having thoracic aortic disease. Although the XD-GRASP framework is clinically acceptable from an image-acquisition perspective, its lengthy image reconstruction (eg, 58 minutes for four-dimensional [4D] whole-heart coronary MR angiography [7], 174 minutes for 4D whole-heart MRI [8], and 250 minutes for 4D thoracic non–contrast-enhanced MR angiography [5]) makes it impractical for clinical translation of time-resolved volumetric cardiovascular MRI. For widespread clinical adoption, there is a need to implement a rapid image-reconstruction system associated with the aforementioned thoracic non–contrast-enhanced MR angiographic pulse sequence (5).

Although the most obvious approach for accelerating an XD-GRASP image reconstruction is using a graphics processing unit (GPU), compressed sensing (CS) is inherently inefficient because the algorithm uses iterative nonlinear conjugate-gradient optimization. A more promising approach to overcome the computational efficiency challenges associated with CS is deep learning (9). Unlike CS, which requires multiple iterations, a trained neural network with optimized weights requires only a single iteration during testing. Two recent studies on cardiac cine MRI reconstruction describe a cascade of convolutional neural networks (CNNs) (10) and a residual U-Net (11) trained on retrospectively undersampled and fully sampled small-size three-dimensional (3D; two spatial dimensions plus time) data sets as input-output pairs. Technical challenges for reconstructing 4D non-Cartesian k-space data using deep learning include the requisite for a computationally intensive gridding algorithm such as nonuniform fast-Fourier transform (12), difficulty in obtaining reference k-space data sets that have not been corrupted by variations in cardiac and respiratory motion over the length of the scan, and lacking open-source 4D CNNs with 4D convolutional kernels. Thus, there is a need to engineer a practical solution for rapidly reconstructing 4D, non-Cartesian, thoracic non–contrast-enhanced MR angiographic k-space data using an open-source 3D CNN on reasonably affordable GPU hardware.

In this study, we sought to accelerate the reconstruction pipeline by replacing central processing unit (CPU)–based multicoil CS in XD-GRASP with GPU-based CNNs of coil-combined data. The purpose of this study was to implement an integrated reconstruction pipeline including a GPU-based CNN architecture and test whether it reconstructs 5.5-fold–accelerated 4D non-Cartesian, non–contrast-enhanced MR angiographic k-space data (5) faster than a CPU-based XD-GRASP reconstruction pipeline, without significant losses in data fidelity, image quality, or artery vessel–diameter measurements, to support clinical translation of said thoracic non–contrast-enhanced MR angiograms.

Materials and Methods

We reused existing raw k-space non–contrast-enhanced MR angiographic data of 13 patients (10 men and three women; mean age, 56.2 years ± 11.5 [standard deviation]) suspected of having thoracic aortic disease who had consented to future analysis of their data (5). In this study, those raw k-space data were reconstructed using both XD-GRASP and CNN reconstruction pipelines to evaluate the proposed CNN pipeline.

Patients

This study was conducted in accordance with protocols approved by our institutional review board and was compliant with the Health Insurance Portability and Accountability Act. All participants provided written informed consent. We prospectively enrolled 11 patients (eight men and three women; mean age, 56.9 years ± 12.8) suspected of having thoracic aortic disease undergoing clinical cardiovascular MRI, in which non–contrast-enhanced MR angiography was added as a research scan. We combined those data sets to a pool of existing non–contrast-enhanced MR angiographic raw k-space data of 13 patients. Combining both raw k-space data sets is justifiable because they were acquired using an identical pulse sequence in patients with the same clinical indication. See Table E1 (supplement) for the relevant clinical profiles of our patient population.

MRI Hardware

MRI scans were conducted with two 1.5-T whole-body MRI scanners (Magnetom Avanto and Aera; Siemens Healthineers, Erlangen, Germany) equipped with a gradient system capable of achieving a maximum gradient strength of 45 mT/m and maximum slew rate of 200 T/m/sec. A body coil was used for radiofrequency excitation, and both body-matrix and spine-coil arrays (approximately 15–18 elements for Avanto and approximately 30 elements for Aera) were used for signal reception.

Pulse Sequence

Relevant imaging parameters of non–contrast-enhanced MR angiography included the following: field of view = 288 × 288 × 120 mm; image-acquisition matrix = 192 × 192 × 80; spatial resolution = 1.5 × 1.5 × 1.5 mm; electrocardiogram-triggering; receiver bandwidth = 745 Hz per pixel; flip angle = 80–90°; echo time msec/repetition time msec = 1.8/3.5; T2 preparation time = 50 msec; fat saturation, α/2 preparation pulse, total number of rays for the outer 90% along partition encoding = 175; total number of rays for the inner 10% along partition encoding = 350; 50 rays (48 for data and two for the navigator) acquired per heartbeat; readout duration per shot = 168 msec; oblique sagittal orientation; and scan time = 350 heartbeats.

Image Reconstruction

We implemented an integrated reconstruction pipeline that replaces the CS part of XD-GRASP with CNN architecture, as shown in Figure 1. To access the TensorFlow code stored in GitHub repository, please use the following: https://github.com/HassanHaji90/CNN-for-dealiasing. Similar to the schematic shown in Figure 1, we implemented an integrated XD-GRASP reconstruction pipeline using the same computer hardware. For additional details on the CNN and XD-GRASP reconstruction pipelines, see Appendix E1 (supplement).

A schematic of an integrated reconstruction pipeline including a 3D                         CNN architecture with four hidden layers and a 3 × 3 x 3 convolutional                         kernel. In this system, raw k-space data files are directly sent to the                         Yarra server, where the data are reconstructed in three sequential steps                         (preprocessing in Matlab, de-aliasing in Python TensorFlow, and                         postprocessing in Matlab). Subsequently, the resulting images in DICOM                         format are sent automatically to a PACS system and retrieved later for                         analysis. Preprocessing includes parallel imaging and self-navigation of                         respiratory motion for binning data into six respiratory phases. Prior to                         de-aliasing, a 4D data set is separated into 80 sets of 3D (two dimensions                         plus time) data in the image domain. Immediately after de-aliasing using a                         trained 3D CNN, the resulting images are combined back into the original 4D                         configuration. Postprocessing involves BM3D filtering to remove residual                         aliasing artifacts. For the XD-GRASP reconstruction pipeline, CNN is                         replaced with compressed sensing. Batch Norm = batch normalization,                         BM3D = block-matching and 3D filtering, CNN = convolutional neural                         network, DICOM = Digital Imaging and Communications in Medicine, 4D                         = four-dimensional, GPU = graphics processing unit, NUFFT =                         nonuniform fast-Fourier transform, PACS = picture archiving and                         communication system, Recon = reconstruction, Relu = rectified                         linear unit, SENSE =sensitivity encoding, 3D = three-dimensional,                         XD-GRASP = extra motion-state golden-angle radial sparse                         parallel.

Figure 1: A schematic of an integrated reconstruction pipeline including a 3D CNN architecture with four hidden layers and a 3 × 3 x 3 convolutional kernel. In this system, raw k-space data files are directly sent to the Yarra server, where the data are reconstructed in three sequential steps (preprocessing in Matlab, de-aliasing in Python TensorFlow, and postprocessing in Matlab). Subsequently, the resulting images in DICOM format are sent automatically to a PACS system and retrieved later for analysis. Preprocessing includes parallel imaging and self-navigation of respiratory motion for binning data into six respiratory phases. Prior to de-aliasing, a 4D data set is separated into 80 sets of 3D (two dimensions plus time) data in the image domain. Immediately after de-aliasing using a trained 3D CNN, the resulting images are combined back into the original 4D configuration. Postprocessing involves BM3D filtering to remove residual aliasing artifacts. For the XD-GRASP reconstruction pipeline, CNN is replaced with compressed sensing. Batch Norm = batch normalization, BM3D = block-matching and 3D filtering, CNN = convolutional neural network, DICOM = Digital Imaging and Communications in Medicine, 4D = four-dimensional, GPU = graphics processing unit, NUFFT = nonuniform fast-Fourier transform, PACS = picture archiving and communication system, Recon = reconstruction, Relu = rectified linear unit, SENSE =sensitivity encoding, 3D = three-dimensional, XD-GRASP = extra motion-state golden-angle radial sparse parallel.

Quantitative Metrics of Image Quality

Although the reconstruction pipeline processed all six respiratory phases, only a single frame corresponding to end expiration was used for data analysis. Given that XD-GRASP, CNN, and zero-filled non–contrast-enhanced MR angiograms are perfectly registered, we calculated the normalized root-mean-square error (NRMSE) and structural similarity index (SSIM) (13), apparent signal-to-noise ratio (SNR), and edge profiles to infer data fidelity with XD-GRASP as the reference. For both SSIM and NRMSE calculations, we excluded signal-free regions (eg, air) and outer parts of the field of view, as shown in Figure 2. For evaluation of image blurring, we measured an intensity profile through the aorta and background and calculated the distance between the 25th and 75th percentiles of the peak-intensity value. To increase precision in calculating the edge profiles, we interpolated each intensity profile by a factor of 20 (see Fig 2). Finally, we also calculated the apparent SNR on the maximum intensity projection as defined by the mean signal of the aorta covering all seven standardized locations divided by the standard deviation of the signal-free background (see Fig 2), as previously described (14).

Left: An ROI (red) encapsulating the aorta was used to calculate the                         SSIM and NRMSE. Middle: Edge thickness was measured from intensity profiles                         across the aorta and background (red line), defined as the distance between                         the 25th and 75th percentiles (blue dotted lines) of peak-intensity values                         for a given profile. Blue dots represent uninterpolated intensity values.                         Right: An ROI (red) covering seven standardized locations and another ROI                         encapsulating the signal-free background (blue) were used to calculate the                         apparent SNR as shown. NRMSE = normalized root-mean-square error, ROI                         = region of interest, SNR = signal-to-noise ratio, SSIM =                         structural similarity index.

Figure 2: Left: An ROI (red) encapsulating the aorta was used to calculate the SSIM and NRMSE. Middle: Edge thickness was measured from intensity profiles across the aorta and background (red line), defined as the distance between the 25th and 75th percentiles (blue dotted lines) of peak-intensity values for a given profile. Blue dots represent uninterpolated intensity values. Right: An ROI (red) covering seven standardized locations and another ROI encapsulating the signal-free background (blue) were used to calculate the apparent SNR as shown. NRMSE = normalized root-mean-square error, ROI = region of interest, SNR = signal-to-noise ratio, SSIM = structural similarity index.

Visual Metrics of Image Quality

To evaluate the diagnostic confidence of reconstruction types, one radiologist trainee (A.M.S.) with 2 years of clinical experience and another radiologist attendee (R.J.A.) with 8 years of clinical experience graded the image quality of non–contrast-enhanced MR angiographic sets. In total, 28 MR angiographic sets (14 patients times two reconstruction types) were randomized and de-identified for display on a 3D workstation (Leonardo; Siemens). Prior to visual evaluation, the readers were given training data sets with poor-to-excellent quality to calibrate readers’ scores (see Fig E4 [supplement] for examples). Following this training session, the readers were blinded to image reconstruction and clinical history. The readers independently graded the images on a five-point Likert scale for three categories: conspicuity of vessel lumen (1 = nondiagnostic, 2 = poor, 3 = clinically acceptable, 4 = good, and 5 = excellent), artifact (1 = nondiagnostic, 2 = poor, 3 = clinically acceptable, 4 = mild, and 5 = minimal), and noise (1 = nondiagnostic, 2 = severe, 3 = clinically acceptable, 4 = mild, and 5 = minimal). We also computed the summed visual score (SVS) as the sum of the three scores (ranging from 3 to 15), for which 9 was defined as clinically acceptable.

Aortic-Diameter Measurements

To evaluate the impact of data fidelity on the clinical task, one 5th-year graduate student (H.H.V.) and another 4th-year graduate student (D.S.) were trained by the clinical reader (A.M.S.) to identify appropriate planes and measured the wall-to-wall vessel diameters at seven standardized locations of the thoracic aorta (15) (1 = aortic sinuses of Valsalva, 2 = sinotubular junction, 3 = mid-ascending aorta, 4 = proximal aortic arch, 5 = mid-aortic arch, 6 = proximal descending thoracic aorta, and 7 = mid-descending aorta) on a 3D workstation (Leonardo). Anatomic landmarks were used to reformat the MR angiographic studies and find appropriate views.

Statistical Analysis

The statistical analyses were conducted by one investigator (H.H.V.) using an SPSS software package (version 23.0, IBM, Armonk, NY). We performed the Shapiro-Wilk test (16) to determine whether SVS were normally distributed. We calculated Cohen κ coefficient (17) and used previously described interpretation definitions (18) to determine the interreader variability in individual visual scores and the intraclass correlation coefficient with model 2, form 1 (19) to determine interreader variability in SVS and vessel diameters. We used average reader scores for comparison between the two reconstruction pipelines. Assuming normal distribution for SVS, SNR, and edge sharpness measurements, the paired t test was used to detect differences between the two groups. Assuming nonnormal distribution for the individual visual scores, the Wilcoxon signed rank test (20) was used to detect differences between the two groups. The Bland-Altman (21) and linear regression analyses were conducted for vessel diameters to determine their agreement and association, respectively. Unless otherwise stated, continuous variables are reported as mean ± standard deviation. A P value less than .05 was considered statistically significant for each statistical test.

Results

Integrated Image-Reconstruction Time

Excluding the pre- and postprocessing steps, the mean image filtering time was 621.7 and 1322.4 times shorter for the CPU-based CNN (mean = 22.1 seconds ± 0.3, range = 21.5−22.6 seconds) and the GPU-based CNN (mean = 10.4 seconds ±0.2, range = 10.0−10.9 seconds), respectively, than the CPU-based CS part of XD-GRASP (13 739.0 seconds ± 1675.5, range = 11 365.4–18 387.3 seconds) (P < .001). Including the pre- and postprocessing steps, the mean reconstruction time was 65 and 69 times shorter for the CPU-based CNN pipeline (mean = 216.6 seconds ± 40.5, range = 179.1–302.5 seconds) and the GPU-based CNN pipeline (mean = 204.9 seconds ± 40.5, range = 167.6–291.4 seconds), respectively, than the CPU-based XD-GRASP pipeline (mean = 14 152.3 seconds ± 1707.6, range = 11 709.3–1905.0 seconds) (P < .001).

Quantitative Metrics of Image Quality

Figure 3 shows three reformatted thin maximum intensity projections of three different patients with diseases (coarctation, aneurysm, and bicuspid aortic valve) produced by the CNN and XD GRASP reconstruction pipelines (see Movies 16 [supplement] for the dynamic display of images shown in Fig 3). With respect to XD-GRASP, the SSIM, NRMSE, apparent SNR, and edge-profile metrics progressively improved after each step in the reconstruction pipeline: zero-filled images (after step 1 in the pipeline) produced an SSIM = 0.72 ± 0.04, an NRMSE = 7.4% ± 1.3, an SNR = 14.0 ± 3.1, and edge sharpness = 2.1 mm ± 1.6; CNN filtering (after step 2 in the pipeline) produced an SSIM = 0.92 ± 0.02, an NRMSE = 3.4% ± 0.5, an SNR = 29.0 ± 9.3, and edge sharpness = 1.4 mm ± 0.3 ; and CNN plus postprocess filtering (after step 3 in the pipeline, see Appendix E1 [supplement] for more details on postprocess filtering) produced an SSIM = 0.94 ± 0.02, an NRMSE = 2.8% ± 0.4, an SNR = 34.0 ± 11.0, and edge sharpness = 1.3 mm ± 0.2. For reference, zero-filled images with postprocess filtering (without step 2 in the pipeline) produced an SSIM = 0.79 ± 0.4, an NRMSE = 6.2% ± 1.2, an SNR = 17.5 ± 4.6, and edge sharpness = 2.1 mm ± 1.2, and XD-GRASP produced an SNR = 32.9 ± 8.5 and edge sharpness = 1.3 mm ± 0.2. Therefore, we used CNN plus postprocess filtering thereafter; CNN refers to CNN plus postprocess filtering throughout.

Thin maximum intensity projections from three different patients                         produced by the XD-GRASP (left side of each pair) and CNN (right side of                         each pair) pipelines: A, patient with coarctation, B, patient with aneurysm,                         and, C, patient with bicuspid aortic valve. In each bottom row, difference                         images with respect to XD-GRASP are displayed with a                         five-times–narrower gray-scale to bring out differences as shown. For                         dynamic display in respiratory and slice dimensions, see Movies 1Movies                         –Movies 6 (supplement). a.u. = arbitrary units, CNN =                         convolutional neural network, XD-GRASP = extra motion-state                         golden-angle radial sparse parallel.

Figure 3: Thin maximum intensity projections from three different patients produced by the XD-GRASP (left side of each pair) and CNN (right side of each pair) pipelines: A, patient with coarctation, B, patient with aneurysm, and, C, patient with bicuspid aortic valve. In each bottom row, difference images with respect to XD-GRASP are displayed with a five-times–narrower gray-scale to bring out differences as shown. For dynamic display in respiratory and slice dimensions, see Movies 16 (supplement). a.u. = arbitrary units, CNN = convolutional neural network, XD-GRASP = extra motion-state golden-angle radial sparse parallel.

Movie 1: Video display of respiratory motion in MIP of a patient with coarctation in the descending aorta (corresponding to Fig 3a): XD-GRASP (left), and CNN (right).

Movie 2: Video display of end-expiratory images along the slice dimension of a patient with coarctation in the descending aorta (corresponding to Fig 3a): XD-GRASP (left), and CNN (right).

Movie 3: Video display of respiratory motion in MIP of a patient with aneurysm in the ascending aorta (corresponding to Fig 3b): XD-GRASP (left), and CNN (right).

Movie 4: Video display of end-expiratory images along the slice dimension of a patient with aneurysm in the ascending aorta (corresponding to Fig 3b): XD-GRASP (left), and CNN (right).

Movie 5: Video display of respiratory motion in MIP of a patient with bicuspid aortic valve (corresponding to Fig 3c): XD-GRASP (left), and CNN (right).

Movie 6: Video display of end-expiratory images along the slice dimension of a patient with bicuspid aortic valve (corresponding to Fig 3c): XD-GRASP (left), and CNN (right).

Visual Metrics of Image Quality

According to the Shapiro-Wilk test (W = 0.90, P = .13), SVS was normally distributed. As summarized in Table E2 (supplement), the average median reader vessel conspicuity, artifact, and noise scores, as well as SVS, were not significantly different (P > .25) between the pair, except for the conspicuity score (P = .02). Nevertheless, both reconstruction pipelines produced individual scores (≥ 3.5) and SVS (≥ 11.6) that were greater than the clinically acceptable cut points of 3 and 9, respectively. Although the individual visual scores assessed by two readers ranged from slight agreement (κ = 0.08) to fair agreement (κ = 0.39), SVS values were strongly correlated (intraclass correlation coefficient = 0.97 for both XD-GRASP and CNN).

Clinical Task: Aortic-Diameter Measurements

As summarized in Table E3 (supplement), the mean vessel diameters were not significantly different (P > .27) between the two groups across all locations, except at the proximal descending aorta (P = .02). However, the percentage difference in diameter at this location was only 3%, which is clinically irrelevant. Figure 4 shows the scatterplots demonstrating that the vessel diameters were strongly correlated (r = 0.99) and in good agreement for both CNNs compared with XD-GRASP (mean value = 3.29 cm, mean difference = −0.01 cm, and upper and lower limits of agreement = 0.19 cm and −0.20 cm, respectively). The vessel diameters measured by two readers were strongly correlated (intraclass correlation coefficient = 0.97, mean value = 0.08 cm, mean difference = 3.29 cm, and upper and lower limits of agreement = 0.22 cm and −0.39 cm, respectively).

Scatterplots summarize (left) linear regression and (right)                         Bland-Altman analyses on average reader aortic-diameter measurements between                         CNN and XD-GRASP. The vessel diameters were strongly correlated (r =                         0.99) and in excellent agreement between CNN and XD-GRASP (mean value =                         3.29 cm, mean difference = −0.01 cm, and upper and lower limits                         of agreement = 0.19 cm and −0.20 cm, respectively). CNN =                         convolutional neural network, XD-GRASP = extra motion-state                         golden-angle radial sparse parallel.

Figure 4: Scatterplots summarize (left) linear regression and (right) Bland-Altman analyses on average reader aortic-diameter measurements between CNN and XD-GRASP. The vessel diameters were strongly correlated (r = 0.99) and in excellent agreement between CNN and XD-GRASP (mean value = 3.29 cm, mean difference = −0.01 cm, and upper and lower limits of agreement = 0.19 cm and −0.20 cm, respectively). CNN = convolutional neural network, XD-GRASP = extra motion-state golden-angle radial sparse parallel.

Discussion

This study described the implementation, training, and testing of an integrated reconstruction pipeline including a GPU-based CNN architecture that is 69 times faster than a CPU-based XD-GRASP reconstruction pipeline. The average median reader SVS for CNN was deemed above the acceptable (11.6) cut point and was not significantly different (P > .25) compared with CS (12.1). The mean vessel diameters were not significantly different (P > .27) between the CNN and CS images for all seven standardized locations of thoracic aorta, except at the proximal descending aorta (P = .02), where the percentage difference in diameter was only 3% (ie, clinically irrelevant).

First, our study describes an engineering approach to address an unmet need in the deep learning community for processing high-dimensional data. Specifically, we separated the 4D data set into multiple series of 3D (two spatial dimensions plus time) data sets, trained an open-source 3D CNN architecture to filter out aliasing artifacts for all 3D series, and combined the results back into the original 4D configuration. To our knowledge, this is the first work describing CNN-based reconstruction of 4D MRI data. This engineering solution may be valuable for other investigators working on high-dimensional MRI, such as 3D perfusion, 3D cine, 4D coronary MR angiography, 4D flow, and time-resolved MR angiography. Second, our proposed CNN was trained using zero-filled, coil-combined, and undersampled images as inputs and XD-GRASP–reconstructed images as outputs. This decision was a practical one because a fully sampled stack-of-star acquisition with six respiratory states at a spatial resolution of 1.5 × 1.5 × 1.5 mm and with 120-mm volume would require approximately 33 minutes of scanning, thereby making it impractical to acquire fully sampled data from patients undergoing a clinical MRI that is already 45–60 minutes long. Third, this study used coil-combined, zero-filled magnitude images for CNN architecture. It may be possible to enhance the performance of CNNs using multicoil data and/or complex data, but it would be challenging to include a data-fidelity term (10,22,23) because of the memory limits of GPU cards for processing complex, multicoil, 4D non-Cartesian k-space data. Fourth, our study goes beyond previous studies using neural networks for cardiac MRI data (10,11) because we used undersampled k-space data for training and testing. Our study goes beyond a previous study using a deep neural network with undersampled Cartesian k-space data (24) because we used non-Cartesian k-space data. Fifth, the reconstruction time, particularly for XD-GRASP, varied among patients because the size of the raw k-space data depends on how many radiofrequency coil elements were used to receive the signal. Sixth, future improvement in computational hardware and/or GPU-accelerated software such as CUDA programming (Nvidia, Santa Clara, Calif) or gpuArray functionality in Matlab (MathWorks, Natick, Mass) will further accelerate CS processing in XD-GRASP reconstruction, potentially making it clinically practical. Seventh, the proposed CNN architecture was trained exclusively with non–contrast-enhanced MR angiographic images produced by a 5.5-fold–accelerated stack-of-star k-space sampling pattern, and it therefore remains unknown how the proposed CNN framework will handle other 4D image types, particularly those acquired with different k-space undersampling patterns. For optimal performance, the proposed CNN framework may need to be adapted (eg, through transfer learning) for different data types. Eighth, this study addresses image-reconstruction efficiency, which is largely focused on improving image quality. For clinical translation, it is crucial to support high-dimensional cardiovascular MRI acquisitions with a rapid image-reconstruction system. In this study, we propose a learning-based image-reconstruction system implemented in the Yarra framework, which automatizes the transfer of raw data from the MRI scanner to a GPU server with the resulting reconstructed results appearing in a digital-imaging and communications-in-medicine format to a picture archiving and communication system. This study demonstrates that a 4D non-Cartesian MR angiographic data set with imaging parameters described in the Materials and Methods section may be reconstructed within less than 4 minutes using the said reconstruction system.

This study had several limitations that warrant further discussion. First, we trained our CNN architecture using data from only 10 patients. This was partially addressed by separating each 4D data set into 80 sets of 3D (two spatial dimensions plus time) data. It may be possible to achieve even higher data fidelity by training the network with more data and/or with data augmentation. Second, we used a GPU-based nonuniform fast-Fourier transform in the preprocessing step to accelerate the gridding process. Despite best efforts, the preprocessing step, including gridding, derivation of autocalibrated sensitivity profiles for parallel imaging, and self-navigation of respiratory motion for binning data, was approximately 18 times longer than the CNN filtering time. A future study is warranted to implement nonuniform fast-Fourier transform in TensorFlow (https://github.com/mikgroup/tf-nufft) to further reduce the preprocessing time. Third, the apparent SNR reported in this study is not equivalent to the intrinsic SNR, which is difficult to estimate in reconstruction frameworks, such as parallel imaging, CS, and neural networks in which noise distribution is spatially varying, poorly defined, and heavily influenced by weights. We used large regions of interests to minimize this effect but acknowledge that the apparent SNR should be interpreted with caution. Fourth, vessel diameters were measured by two graduate students who were trained by a clinical radiologist. Despite their inexperience, the two graduate students achieved high interrater agreement (intraclass correlation coefficient = 0.97). The visual scores, however, were evaluated by two clinical radiologists. Although this preliminary analysis is encouraging, a thorough evaluation in a large cohort of patients is necessary to validate the accuracy of vessel-diameter measurements made from CNN-reconstructed images. Fifth, we used both Matlab (pre- and postprocessing) and Python (de-aliasing) in the reconstruction pipeline because our laboratory has access to extensive computational tools in Matlab. We acknowledge that mixing computer languages is less efficient because of time wasted on the saving and retrieval of data between computer languages. A future study is warranted to code the entire reconstruction pipeline in Python.

In summary, this study describes implementation, training, and testing of an integrated reconstruction pipeline with a GPU-based CNN architecture that is capable of reconstruction of 5.5-fold–accelerated 4D non-Cartesian, non–contrast-enhanced MR angiographic k-space data at a rate 69 times faster than a CPU-based XD-GRASP reconstruction pipeline, without significant losses in data fidelity, image quality, or vessel-diameter measurements, thereby supporting clinical translatability of said non–contrast-enhanced MR angiograms.

Disclosures of Conflicts of Interest: H.H.V. disclosed no relevant relationships. D.S. disclosed no relevant relationships. R.J.A. disclosed no relevant relationships. A.M.S. disclosed no relevant relationships. F.A.S. disclosed no relevant relationships. A.K.K. disclosed no relevant relationships. O.S.C. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: employed by Northwestern University. Other relationships: disclosed no relevant relationships. D.K. Activities related to the present article: institution receives grants from NIH (R01HL116895, R01HL138578, R21EB024315, R21AG055954) and American Heart Association (19IPLOI34760317) (funding from NIH and AHA supported some of the authors while working on this project). Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships.

Author Contributions

Author contributions: Guarantors of integrity of entire study, H.H.V., R.J.A., A.K.K., D.K.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, H.H.V., A.M.S., F.A.S., A.K.K., D.K.; clinical studies, H.H.V., R.J.A., A.M.S., D.K.; statistical analysis, H.H.V., D.S., F.A.S., D.K.; and manuscript editing, all authors.

Supported in part by the National Institutes of Health (grants R01HL116895, R01HL138578, R21EB024315, and R21AG055954) and the American Heart Association (grant 19IPLOI34760317).

References

  • 1. Broome DR, Girguis MS, Baron PW, Cottrell AC, Kjellin I, Kirk GA. Gadodiamide-associated nephrogenic systemic fibrosis: why radiologists should be concerned. AJR Am J Roentgenol 2007;188(2):586–592. Crossref, MedlineGoogle Scholar
  • 2. Olchowy C, Cebulski K, Łasecki M, et al. The presence of the gadolinium-based contrast agent depositions in the brain and symptoms of gadolinium neurotoxicity - a systematic review. PLoS One 2017;12(2):e0171704. Crossref, MedlineGoogle Scholar
  • 3. Wheaton AJ, Miyazaki M. Non-contrast enhanced MR angiography: physical principles. J Magn Reson Imaging 2012;36(2):286–304. Crossref, MedlineGoogle Scholar
  • 4. Miyazaki M, Lee VS. Nonenhanced MR angiography. Radiology 2008;248(1):20–43. LinkGoogle Scholar
  • 5. Haji-Valizadeh H, Collins JD, Aouad PJ, et al. Accelerated, free-breathing, noncontrast, electrocardiograph-triggered, thoracic MR angiography with stack-of-stars k-space sampling and GRASP reconstruction. Magn Reson Med 2019;81(1):524–532. Crossref, MedlineGoogle Scholar
  • 6. Feng L, Axel L, Chandarana H, Block KT, Sodickson DK, Otazo R. XD-GRASP: golden-angle radial MRI with reconstruction of extra motion-state dimensions using compressed sensing. Magn Reson Med 2016;75(2):775–788. Crossref, MedlineGoogle Scholar
  • 7. Piccini D, Feng L, Bonanno G, et al. Four-dimensional respiratory motion-resolved whole heart coronary MR angiography. Magn Reson Med 2017;77(4):1473–1484. Crossref, MedlineGoogle Scholar
  • 8. Feng L, Coppo S, Piccini D, et al. 5D whole-heart sparse MRI. Magn Reson Med 2018;79(2):826–838. Crossref, MedlineGoogle Scholar
  • 9. Kyong Hwan Jin, McCann MT, Froustey E, Unser M. Deep Convolutional Neural Network for Inverse Problems in Imaging. IEEE Trans Image Process 2017;26(9):4509–4522. Crossref, MedlineGoogle Scholar
  • 10. Schlemper J, Caballero J, Hajnal JV, Price AN, Rueckert D. A deep cascade of convolutional neural networks for dynamic MR image reconstruction. IEEE Trans Med Imaging 2018;37(2):491–503. Crossref, MedlineGoogle Scholar
  • 11. Hauptmann A, Arridge S, Lucka F, Muthurangu V, Steeden JA. Real-time cardiovascular MR with spatio-temporal artifact suppression using deep learning-proof of concept in congenital heart disease. Magn Reson Med 2019;81(2):1143–1156. Crossref, MedlineGoogle Scholar
  • 12. Fessler JA, Sutton BP. Nonuniform fast Fourier transforms using min-max interpolation. IEEE Trans Signal Process 2003;51(2):560–574. CrossrefGoogle Scholar
  • 13. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 2004;13(4):600–612. Crossref, MedlineGoogle Scholar
  • 14. Dietrich O, Raya JG, Reeder SB, Reiser MF, Schoenberg SO. Measurement of signal-to-noise ratios in MR images: influence of multichannel coils, parallel imaging, and reconstruction filters. J Magn Reson Imaging 2007;26(2):375–385. Crossref, MedlineGoogle Scholar
  • 15. Hiratzka LF, Bakris GL, Beckman JA, et al. 2010 ACCF/AHA/AATS/ACR/ASA/SCA/SCAI/SIR/STS/SVM guidelines for the diagnosis and management of patients with thoracic aortic disease. A report of the American College of Cardiology Foundation/American Heart Association Task Force on Practice Guidelines, American Association for Thoracic Surgery, American College of Radiology, American Stroke Association, Society of Cardiovascular Anesthesiologists, Society for Cardiovascular Angiography and Interventions, Society of Interventional Radiology, Society of Thoracic Surgeons, and Society for Vascular Medicine. J Am Coll Cardiol 2010;55(14):e27–e129. Crossref, MedlineGoogle Scholar
  • 16. Shapiro SS, Wilk MB. An analysis of variance test for normality (complete samples). Biometrika 1965;52(3-4):591–611. CrossrefGoogle Scholar
  • 17. Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas 1960;20(1):37–46. CrossrefGoogle Scholar
  • 18. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics 1977;33(1):159–174. Crossref, MedlineGoogle Scholar
  • 19. Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing rater reliability. Psychol Bull 1979;86(2):420–428. Crossref, MedlineGoogle Scholar
  • 20. Wilcoxon F. Individual comparisons by ranking methods. Biom Bull 1945;1(6):80–83. CrossrefGoogle Scholar
  • 21. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1(8476):307–310. Crossref, MedlineGoogle Scholar
  • 22. Hammernik K, Klatzer T, Kobler E, et al. Learning a variational network for reconstruction of accelerated MRI data. Magn Reson Med 2018;79(6):3055–3071. Crossref, MedlineGoogle Scholar
  • 23. Yang G, Yu S, Dong H, et al. DAGAN: deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction. IEEE Trans Med Imaging 2018;37(6):1310–1321. Crossref, MedlineGoogle Scholar
  • 24. Chen F, Taviani V, Malkiel I, et al. Variable-density single-shot fast spin-echo MRI with deep learning reconstruction by using variational networks. Radiology 2018;289(2):366–373. LinkGoogle Scholar

Article History

Received: Sept 30 2019
Revision requested: Dec 2 2019
Revision received: Feb 4 2020
Accepted: Mar 2 2020
Published online: June 25 2020