Source-linked AI summary
Real-time Cardiovascular MR with Spatio-temporal Artifact Suppression using Deep Learning - Proof of Concept in Congenital Heart Disease
Andreas Hauptmann, Simon Arridge, Felix Lucka, Vivek Muthurangu, Jennifer A. Steeden
TL;DR
Real-time ventricular-volume assessment needs highly accelerated imaging, but undersampling creates artifacts and breath-holding can be difficult for patients. The paper trains a 3D residual CNN on synthetic paired data, tests radial sampling patterns, and applies it to real-time CHD data. CNN reconstruction was faster and observed to outperform GRASP while producing ventricular volumes not significantly different from cardiac-gated breath-hold reference techniques.
Problem
Real-time ventricular-volume assessment requires high acceleration, while breath-holding is difficult for some patients and undersampling causes reconstruction artifacts.
Method
A 3D residual U-Net was trained on synthetic paired artifact-free and undersampled data, evaluated across radial sampling patterns, and applied to real-time CHD data.
Results
The residual U-Net produced superior image quality and ventricular-volume measurement relative to CS reconstructions, with volumes not statistically significantly different from gold-standard cardiac-gated breath-hold techniques.
Takeaways & Limitations
3D CNN deep de-aliasing showed potential for clinical reconstruction of real-time radial data in patients with CHD.
Takeaways & Limitations
Training data were acquired during breath-hold, whereas real-time data were acquired during free-breathing.
Abstract
from arXiv · showhide
PURPOSE: Real-time assessment of ventricular volumes requires high acceleration factors. Residual convolutional neural networks (CNN) have shown potential for removing artifacts caused by data undersampling. In this study we investigated the effect of different radial sampling patterns on the accuracy of a CNN. We also acquired actual real-time undersampled radial data in patients with congenital heart disease (CHD), and compare CNN reconstruction to Compressed Sensing (CS). METHODS: A 3D (2D plus time) CNN architecture was developed, and trained using 2276 gold-standard paired 3D data sets, with 14x radial undersampling. Four sampling schemes were tested, using 169 previously unseen 3D 'synthetic' test data sets. Actual real-time tiny Golden Angle (tGA) radial SSFP data was acquired in 10 new patients (122 3D data sets), and reconstructed using the 3D CNN as well as a CS algorithm; GRASP. RESULTS: Sampling pattern was shown to be important for image quality, and accurate visualisation of cardiac structures. For actual real-time data, overall reconstruction time with CNN (including creation of aliased images) was shown to be more than 5x faster than GRASP. Additionally, CNN image quality and accuracy of biventricular volumes was observed to be superior to GRASP for the same raw data. CONCLUSION: This paper has demonstrated the potential for the use of a 3D CNN for deep de-aliasing of real-time radial data, within the clinical setting. Clinical measures of ventricular volumes using real-time data with CNN reconstruction are not statistically significantly different from the gold-standard, cardiac gated, BH techniques.
2. Computational Imaging, Centrum Wiskunde & Informatica (CWI), Science Park 123, 1098 XG Amsterdam
The listed affiliations include the UCL Centre for Cardiovascular Imaging and the Institute of Cardiovascular Science in London, alongside a corresponding author.
- The UCL Centre for Cardiovascular Imaging is affiliated with University College London in London.
- Vivek Muthurangu is identified as the corresponding author.
- The UCL Centre for Cardiovascular Imaging is part of the Institute of Cardiovascular Science.
30 Guildford Street, London. WC1N 1EH
The supplied passages contain contact information, a running title, keywords, and a congenital-heart-disease reference.
- The paper lists a corresponding author telephone number and fax number.
- The running title is “Deep Artifact Suppression for Ventricular Volumes.”
- Keywords include deep artifact suppression, deep learning, convolutional neural networks, real-time imaging, and ventricular volumes.
- The study concerns congenital heart disease (CHD).
INTRODUCTION
The introduction motivates free-breathing real-time CMR for patients who struggle with breath-holding, while undersampling creates artifacts requiring advanced reconstruction. It proposes CNN-based artifact removal and evaluates radial sampling patterns in CHD data.
- Breath-hold CMR can be difficult for many patients, motivating a rapid free-breathing alternative.
- Real-time imaging requires substantial undersampling to maintain spatial and temporal resolution, making artifact-free reconstruction challenging.
- Compressed sensing reconstruction is computationally intensive, time-consuming, and sensitive to parameter optimization that can produce unnatural-looking images.
- Tiny golden-angle radial sampling produces spatially incoherent artifacts, making undersampling artifacts amenable to a denoising perspective and motivating CNN removal.
- The study investigates how radial sampling patterns affect CNN reconstruction and compares tGA real-time CHD data with a state-of-the-art compressed-sensing algorithm.
METHODS
The methods develop a 3D residual U-Net for spatio-temporal de-aliasing, train it on paired synthetic undersampled data, and evaluate sampling patterns before applying the network to real-time CHD data.
- Residual U-Net Architecture and Training: A residual U-Net uses multi-scale downsampling and upsampling to produce a cleaned version of an artifact-contaminated image.
- Residual U-Net Architecture and Training: The network processes 3D data comprising 2D images through time to enforce temporal consistency and prevent slice-wise flickering.
- Synthetic Data and Evaluation: Training used paired artifact-free ground-truth and undersampled artifact-contaminated images, followed by synthetic testing and reconstruction of real-time CHD data.
- Assessing the Effect of Sampling Strategy: The patterns generated distinct artifact behavior, ranging from temporally invariant regular aliases to spatially incoherent aliases that varied through time.
Analysis of In-vivo Data
In-vivo image quality and ventricular function were assessed by comparing real-time radial reconstructions with gold-standard BH-bSSFP images using qualitative, quantitative, and statistical analyses.
- Image quality and ventricular function: Real-time radial data were compared with standard BH-bSSFP images using image-quality scoring and ventricular-volume measurements.A CMR specialist reviewed randomized mid-ventricular cine loops on a 5-point Likert scale, while ventricular sections were manually segmented.
- Image quality and ventricular function: Image-quality scoring evaluated endocardial-border sharpness, temporal fidelity of wall motion, and residual artifacts.Sharpness, motion fidelity, and artifacts were each graded from non-diagnostic or poor to excellent or minimal, as applicable.
- Image quality and ventricular function: Quantitative edge sharpness was estimated from the maximum normalized-intensity gradient across the septal border.Pixel intensities were fit with a tenth-order polynomial before differentiation, and values were averaged across six septal positions and all cardiac phases.
- Image quality and ventricular function: Left- and right-ventricular end-diastolic and end-systolic volumes and ejection fractions were calculated after manual segmentation.End-diastolic and end-systolic phases were selected by visual inspection, with papillary muscles and trabeculae included in the blood pool.
- Statistical analysis: BH-bSSFP data served as the reference standard for assessing agreement of ventricular volumes and function.Continuous variables were compared using repeated-measures ANOVA with Bonferroni correction, while ordinal image scores used the Wilcoxon signed-rank test.
RESULTS
The results show that radial sampling pattern affected synthetic reconstruction quality, while the CNN reconstruction was faster and generally better than GRASP in prospective data.
- Synthetic test data: tGArot produced significantly lower RMSE and higher SSIM than the other radial sampling patterns.The differences were statistically significant for reconstruction accuracy, with p<0.0001 reported for the relevant comparisons.
- Synthetic test data: tGArot also reduced visible papillary-muscle or trabeculation defects, poor motion fidelity, and endocardial-border blurring compared with other patterns.Non-rotating trajectories produced greater blurring and could lose papillary muscles relative to the truth data.
- Robustness assessment: As SNR decreased from 20 dB to 10 dB, SSIM decreased by 6.8% and RMSE increased by 1.3% for the tGArot network.SSIM changed from 0.862 at 20 dB to 0.794 at 10 dB, while RMSE changed from 0.043 to 0.056.
- Robustness assessment: As acceleration increased from 10x to 16x, SSIM decreased by 6% and RMSE increased by 0.9%.SSIM changed from 0.89 at 10x to 0.83 at 16x, while RMSE changed from 0.041 to 0.050.
- Prospective in-vivo study: CNN reconstructions had better image quality than GRASP, with less residual aliasing, less temporal blurring, and superior qualitative scores.The CNN had similar myocardial delineation, motion fidelity, and artifact scores to BH-bSSFP, while edge sharpness ranked BH-bSSFP, CNN, then GRASP.
- Ventricular volume quantification: The CNN showed fewer significant ventricular-volume biases than GRASP against BH-bSSFP.GRASP had significant biases for LV ESV, LV EF, and RV EDV, whereas the CNN had only a significant RV EDV bias.
DISCUSSION
The study shows that a residual U-Net can suppress artifacts in synthetic and actual real-time radial CMR, with performance strongly dependent on the sampling pattern. Compared with GRASP, the residual U-Net reconstructed images faster and provided better image quality and ventricular-volume quantification, while several training and generalization limitations remain.
- Synthetic training and translation: The residual U-Net successfully suppressed artifacts in actual radial real-time data acquired from new patients with congenital heart disease.The network was trained using synthetic data and then applied to acquired real-time data.
- Comparison with GRASP: >5x shorter total reconstruction times were achieved with the residual U-Net than with GRASP for the same raw data.The comparison included creation of aliased images for the CNN reconstruction.
- Comparison with GRASP: Residual U-Net reconstructions had superior image quality and more accurate ventricular-volume measurements than GRASP reconstructions.Better endocardial-border sharpness and temporal fidelity likely contributed to the improved volume quantification.
- Synthetic training and translation: A residual U-Net successfully performed deep artifact suppression on synthetic data generated from retrospectively acquired Cartesian breath-hold images.Synthetic data provided known ground truth but could not reproduce acquisition errors or respiratory motion in free-breathing real-time data.
- Sampling patterns: Sampling pattern strongly affected reconstruction accuracy, with non-rotating trajectories producing significant RMSE and SSIM errors and qualitative artifacts.Non-rotating trajectories reused the same spokes across frames and caused missing or deformed papillary muscles or trabeculations in over 88% of synthetic test cases.
- Sampling patterns: Continuously rotating strategies performed better, and tGArot produced noise-like aliases that were associated with superior synthetic reconstruction quality.The authors suggest that less temporally coherent aliasing explains the improved performance of continuously rotating trajectories, particularly tGArot.
- Limitations: Generalization remains limited because single-ventricle patients and the full diversity of congenital heart disease were not studied.The authors also note constraints from breath-hold training data, fixed input and training dimensions, magnitude-only images, and unmodeled acquisition or respiratory effects.
- Clinical translation: The study supports clinical translation of residual U-Net reconstruction because ventricular volumes were not statistically significantly different from the cardiac-gated breath-hold reference standard.The conclusion frames the method as potentially useful for adopting heavily undersampled real-time techniques in clinical workflows.
SUPPORTING INFORMATION
Supporting information details the datasets, reconstruction comparisons, robustness assessments, and visual examples used to evaluate the residual U-Net and GRASP methods.
- Reconstruction architecture: A 2D U-Net produced residual flickering artifacts on dynamic data, whereas the 3D U-Net used in the study did not.The comparison illustrates why the reconstruction architecture incorporates time.
- Datasets and workflow: The supporting materials document training, synthetic test, and prospective real-time datasets used in the study.Dataset labels include training data, synthetic test data, and prospective real-time data; the flow diagram describes conversion of retrospectively cardiac-gated BH-bSSFP data into synthetic real-time training data.
- Robustness analyses: Supporting figures assess reconstruction robustness across acceleration factors, cropping positions, signal-to-noise ratios, and loss functions.The materials report RMSE and SSIM across these conditions and compare networks trained with L1- and L2-loss functions.
- Sampling patterns: tGArot preserved papillary muscles better than other trajectories, while non-rotating trajectories produced ventricular blurring.These examples compare reconstructions from different sampling patterns using correspondingly trained residual U-Nets.
- Clinical comparisons: Prospective examples compare BH-bSSFP with real-time radial reconstructions from GRASP and the residual U-Net, including left- and right-ventricular volume agreement.The Bland–Altman plots show mean differences and ±2 standard deviations for both ventricles.