An inline deep learning based free-breathing ECG-free cine for exercise cardiovascular magnetic resonance

Figures

• 

  Fig. 1. Deep learning model and generation of training data. A The deep learning model consists of a three-dimensional U-Net trained to filter out streaking artifacts from complex-valued real-time radial cine \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n\times n$$\end{document}n×n images with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{t}$$\end{document}nt time frames. The inputs and outputs to the U-Net are the real and imaginary components concatenated along the spatial dimension, and the loss during training was the mean square error (MSE) between inputs and outputs. During inference, the generated real and imaginary parts with suppressed artifacts are re-combined. B Synthetic real-time image training pairs were generated from electrocardiogram (ECG)-gated segmented Cartesian k-space data acquired at rest in 503…

• 

  Fig. 2. Inline implementation of real-time cine with deep learning-based radial acceleration. Multi-coil raw radial k-space data acquired from the scanner is processed in the Image Reconstruction Environment (ICE) on the vendor reconstruction computer. Using the International Society for Magnetic Resonance in Medicine Raw Data (ISMRMRD) format, the collected data is transferred to the Framework for Image Reconstruction (FIRE) server using a FireEmitter functor. The FIRE server is located in the vendor reconstruction computer. The data is then transferred from the FIRE server to an external server via a connecting Secure Shell Protocol (SSH) tunnel. In the external sever, a Docker containing all Python dependencies such as PyTorch is used to process the raw k-space data in a single 32 GB Graphics Processing Unit (GPU). The deep-learning radial acceleration with parallel reconstruction (DRAPR) technique was implemented in the Docker. First, a non-uniform fast Fourier transform (NUFFT) is used to grid and reconstruct undersampled multi-coil radial k-space data. GPU parallelization is done in PyTorch by treating frames and coils as batch and channel dimensions. This approach enables…

• 

  Fig. 3. Exercise imaging protocol. Rest scans consisted of breath-hold, ECG-gated Cartesian segmented cine followed by free-breathing ECG-free real-time radial cine. After rest imaging, subjects were removed from the scanner bore and were exercised in the supine position using a cycle ergometer. Work rate was started at 25 W and was increased by + 25 W every 2 min while subjects maintained a constant pedaling speed of 75 rpm. After reaching target heart rate or exhaustion, subjects were immediately placed back inside the scanner bore for post-exercise stress imaging. This consisted of a repetition of the real-time cine sequence

• 

  Fig. 4. Bland–Altman comparisons of left-ventricular (LV) parameters. Measures derived from breath-hold ECG-gated segmented cine are compared to those derived from the free-breathing ECG-free real-time cine reconstructed with the GRASP and deep-learning radial acceleration with parallel reconstruction (DRAPR) techniques. Solid and dotted lines represent the mean difference and mean difference ± 1.96 × standard deviation of the difference. A The mean differences in LV end-diastolic and end-systolic volumes with golden angle radial sparse parallel (GRASP) were 4.4 mL and 6.4 mL, while the mean difference in LV ejection fraction was − 4.3%. B The mean differences in LV end-diastolic and LV end-systolic volumes with DRAPR were 3 mL and 7 mL, while the mean difference in ejection fraction was − 5%

• 

  Fig. 5. Subjective evaluation of image quality. Three readers evaluated artifacts across all cine scans. Images were classified as having a 1-non-diagnostic, 2-severe, 3-moderate, and 4-minimal artifact level. The mean score of breath-hold ECG-gated segmented cine images was 3.0. The mean score of free-breathing ECG-free real-time cine images reconstructed with GRASP was 3.7 at rest, with 9.1% of images graded as severe. The mean score during post-exercise stress was 3.5, and the percentage of images graded as severe increased to 17.9%. Real-time cine images at rest reconstructed with DRAPR had a mean score of 3.3, with 10.6% of images graded as severe. During post-exercise stress, images had a mean score of 3.1. The percentage of images graded as severe increased to 15.4%. None of the real-time radial cine images reconstructed with GRASP or DRAPR were labeled as non-diagnostic. *p < 0.01. GRASP: golden-angle radial sparse parallel; DRAPR: deep-learning radial acceleration with parallel reconstruction

• 

  Fig. 6. Representative real-time cine images of subjects at rest. The NUFFT was used to grid and reconstruct radial k-space data. Artifacts were subsequently suppressed using the GRASP and DRAPR techniques. Images were classified as having a 1-non-diagnostic, 2-severe, 3-moderate, and 4-minimal artifact level. All NUFFT images were 1-non-diagnostic. The five subjects shown at end-diastole represent the range of mean scores at rest. Subjects 1–4 are patients. The mean artifact levels with DRAPR for subjects 1–5 were 3.0, 3.3, 3.3, 4.0 and 4.0, accordingly. The body mass index (BMI) for subjects 1–5 was 38, 33, 26, 23 and 21 lbs/in2, accordingly

• 

  Fig. 7. Representative real-time cine images of subjects at post-exercise stress. The NUFFT was used to grid and reconstruct radial k-space data. Artifacts were subsequently suppressed using the GRASP and DRAPR techniques. Images were classified as having a 1-non-diagnostic, 2-severe, 3-moderate, and 4-minimal artifact level. All NUFFT images were 1-non-diagnostic. The five subjects shown at end-diastole represent the range of mean scores at post-exercise stress. Subjects 3–4 are healthy. Subjects 2 and 5 correspond to Subjects 1 and 4 in Fig. 6, accordingly. The mean artifact levels with DRAPR for subjects 1–5 were 2.3, 3.0, 3.0, 3.3 and 3.7, accordingly. The body mass index (BMI) for subjects 1–5 was 32, 38, 23, 29 and 24 lbs/in2, accordingly. The peak heart rate post-exercise for subjects 1–5 was 110, 108, 97, 132, and 135 bpm, accordingly

• 

  Fig. 8. Inline real-time cine with deep learning-based radial acceleration in coronary artery disease (CAD). Four patients are shown whose recruitment criteria included being symptomatic and having a stress test for suspicion of CAD. Patients underwent an exercise CMR protocol. Top images were collected at rest, and bottom images were collected after supine exercise using a CMR compatible cycle ergometer. Images are shown at end-diastole. Patient 4 at pre-exercise corresponds to subject 4 in Fig. 6, and at post-exercise corresponds to subject 5 in Fig. 7, accordingly. The BMI for patients 1–4 was 22, 27, 26, and 24 lbs/in2, accordingly. During rest, the mean artifact levels were 3.0, 3.7, 4.0 and 4.0. The resting heart rates were 83, 62, 70 and 49 bpm. During post-exercise stress, the mean artifact levels were 3.3, 3.3, 3.3 and 3.7. The peak heart rates post-exercise were 115, 132, 107 and 134 bpm, accordingly

• 

  Fig. 9. Illustrative clinical case. A 60-year-old male with worsening upper chest heaviness and shortness of breath with exertion. Stress echo showed focal systolic dysfunction with hypokinesis of the left-ventricular basal inferoseptum and inferior walls at both rest and stress. Similarly, during an exercise CMR protocol, real-time cine with deep learning-based radial acceleration showed hypokinesis of the left-ventricular basal inferoseptal wall during rest and at post-exercise stress (arrows). Subsequent coronary angiography revealed a heavily calcified right coronary artery (arrow) with serial subtotal occlusions