Source-linked AI summary
Attention-guided super-resolution of 4D flow MRI in carotid arteries
Ali Mokhtari, Dominik Obrist
TL;DR
Low resolution and noise limit 4D flow MRI assessment of carotid hemodynamics. The study trains an attention-guided, CFD-supervised super-resolution model on patient-specific carotid data, and reports lower RMSE, more stable predictions, and improved reconstruction of complex velocity contours than a base model.
Problem
Low spatial resolution and noise in 4D flow MRI restrict accurate assessment of hemodynamic biomarkers and blood-flow dynamics.
Method
The study combines multi-scale feature extraction and channel-spatial attention in a super-resolution model trained with patient-specific CFD velocity fields from 120 patients and 240 carotid arteries.
Results
The attention-based model achieved nearly half the RMSE of the base model and better captured complicated velocity contours.
Takeaways & Limitations
The model provides a more dependable approach for reconstructing noisy carotid 4D flow MRI velocity fields.
Takeaways & Limitations
The evaluation relies on CFD ground truth, examines only twofold enhancement, uses single-center data, and omits derived hemodynamic metrics and independent in vivo validation.
Abstract
from arXiv · showhide
Four-dimensional (4D) flow magnetic resonance imaging (MRI) is a powerful non-invasive technique for visualizing and quantifying complex blood flow patterns in vivo. Despite its clinical promise, broader adoption is limited by low spatial resolution and sensitivity to noise, which restrict accurate assessment of critical hemodynamic biomarkers such as wall shear stress, pressure gradients, and turbulent kinetic energy. To overcome these challenges, we propose a deep learning-based super-resolution framework that integrates multi-scale feature extraction and attention mechanisms to enhance the quality of 4D flow MRI data. The model was trained on a dataset of 120 patients with 240 stenosed carotid arteries. High-resolution ground truth data were generated using patient-specific computational fluid dynamics (CFD) simulations based on segmented vascular geometries and physiologically realistic boundary conditions, and the resulting velocity fields served as targets for supervised learning. The proposed architecture uses convolutional block attention modules (CBAM) to guide the network toward clinically relevant spatial features and to suppress noise in low-resolution inputs. Quantitative results show that the attention-guided model substantially reduces the root mean square error (RMSE) compared with a baseline model without attention, and qualitative velocity contour analysis confirms improved reconstruction of intricate flow patterns. These findings highlight the capacity of the model to restore high-fidelity flow fields under noisy conditions and support the use of deep learning to extend the clinical utility of 4D flow MRI for non-invasive hemodynamic assessment.
1. Introduction
Carotid 4D flow MRI provides comprehensive time-resolved velocity information, but resolution, noise, and acquisition artifacts limit reliable hemodynamic assessment. This work addresses these challenges with CFD-supervised, attention-based super-resolution for patient-specific carotid data.
- 1. Introduction: 4D flow MRI captures time-resolved, three-directional velocities throughout volumetric vascular regions.This supports analysis of pulsatile and multidirectional carotid flow patterns.
- 1. Introduction: Low spatial and temporal resolution, spatial averaging, partial volume effects, aliasing, and phase-offset errors degrade 4D flow fidelity.These artifacts can hinder accurate quantification of hemodynamic biomarkers.
- 1. Introduction: Super-resolution methods aim to improve spatial resolution and reduce noise in 4D flow MRI.Prior approaches include CFD-trained networks, physics-informed models, and temporal super-resolution.
- 1. Introduction: The study combines multi-scale feature extraction and channel-spatial attention using 240 patient-specific carotid artery simulations.The cohort contains 120 patients with 240 stenosed carotid arteries.
2. Methods
The study uses patient-specific carotid MRI data and CFD simulations to create high-resolution velocity targets under physiologically based conditions. The simulations resolve carotid flow with detailed meshes, boundary conditions, and numerical assumptions.
- 2. Methods: The cohort comprised 120 patients with 240 stenosed carotid arteries and a mean stenosis of 34 ± 17%.Patients had internal carotid plaque at least 1.5 mm thick and stenosis below 50% by NASCET criteria.
- 2. Methods: 4D flow MRI was acquired at 3T with 0.8 mm isotropic spatial resolution and 52.8 ms temporal resolution.The acquisition used a prospectively ECG-triggered k-t-accelerated 3D phase-contrast sequence.
- 2. Methods: Patient-specific vascular geometries were reconstructed from segmented MRI data, smoothed, centerlined, and meshed for CFD analysis.The workflow used Taubin smoothing, VMTK centerlines, and OpenFOAM meshing with local refinement in carotid bulb and stenotic regions.
- 2. Methods: MRI-derived velocities supplied physiologically based inlet and outlet conditions after registration and interpolation onto the CFD mesh.Flow rates were corrected using multiple cross-sections and synchronized across carotid branches.
- 2. Methods: The q-DNS simulations solved the Navier–Stokes equations without turbulence models using Newtonian blood properties and adaptive time stepping.Each case ran for three cardiac cycles, with the first two excluded from analysis.
3. Proposed Architecture
The proposed network progressively reconstructs high-resolution volumetric velocity data using multi-scale convolutions and CBAM attention. Attention is inserted through the feature-extraction and upsampling stages to emphasize relevant features and suppress noise.
- 3. Proposed Architecture: The architecture uses initial feature extraction, two multi-scale feature-extraction blocks, three upsampling blocks, and a final convolutional layer.Features are progressively refined before producing the high-resolution output.
- 3. Proposed Architecture: The initial feature-extraction block uses three 3D convolutional layers with ReLU activations to increase feature dimensionality while preserving spatial information.This block forms the foundation of the architecture.
- 3. Proposed Architecture: Each multi-scale block processes features with 3D kernels of sizes 3, 5, and 9 in parallel before summing their outputs.Different kernel sizes enable representation of features at multiple spatial scales.
- 3. Proposed Architecture: Depthwise separable convolutions reduce parameters and operations relative to conventional convolutions while supporting enlarged kernels.The reduction is intended to support faster training and inference.
- 3. Proposed Architecture: CBAM combines channel and spatial attention to suppress redundant or noisy features and emphasize relevant regions.It is inserted after the first multi-scale block and after each upsampling block.
4. Results & Discussion
The attention-based model reconstructed noisy 4D flow MRI velocity fields more accurately than the base model, with lower and more stable RMSE and improved velocity-contour recovery. However, the evaluation remains bounded by CFD assumptions, limited upsampling, single-center data, velocity-only metrics, and missing independent validation.
- 4. Results & Discussion: Nearly half the RMSE was achieved by the attention-based model compared with the base model.The base model had higher RMSE and was less accurate at reconstructing velocity fields under noisy inputs.
- 4. Results & Discussion: The attention model showed a narrower RMSE distribution, fewer extreme values, and a lower median than the base model.These patterns indicate reduced variance and more reliable predictions across the dataset.
- 4. Results & Discussion: The attention-based model captured complicated velocity patterns that the base model missed, although predicted values and locations still showed discrepancies.Further adjustment may be needed to better match the actual data.
- 4. Results & Discussion: The results extend CFD-supervised super-resolution to 240 patient-specific carotid arteries and improve robustness to measurement noise.Direct numerical comparison with prior studies is not possible because territories, resolution ratios, error definitions, and evaluation volumes differ.
- 4. Results & Discussion: The network learns CFD velocity fields based on assumptions including rigid walls, Newtonian rheology, and MRI-derived boundary conditions.Agreement with CFD is therefore not equivalent to agreement with true in vivo flow.
- 4. Results & Discussion: The evaluation used only twofold resolution enhancement, one center and scanner, velocity errors, and no independent in vivo validation.Derived quantities such as wall shear stress, pressure gradients, and turbulent kinetic energy were not assessed.
5. Conclusion
The attention-based super-resolution model improves 4D flow MRI velocity-field reconstruction while addressing low resolution and noise. High dataset noise remains a limitation, motivating denoising strategies and attention to temporal resolution.
- The model integrates multi-scale feature extraction and attention mechanisms to improve spatial resolution and reduce noise in 4D flow MRI.
- High dataset noise may obscure fine details and produce artifacts that undermine the reliability and quality of reconstructed images.
- Future improvement should combine super-resolution with denoising while ensuring spatial enhancements do not compromise temporal resolution.
Ethical approval and consent to participate
The underlying imaging studies received ethics approval and obtained written informed consent from all participants.
- The original imaging studies were approved by the University of Freiburg Ethics Committee, and all participants provided written informed consent.
Consent for publication
The manuscript reports a secondary retrospective analysis of fully anonymized data and includes no individual person’s data.
- No individual person’s data are included in this manuscript.
Funding
The research received financial support from the Swiss National Science Foundation.
- The Swiss National Science Foundation financially supported this research through grant #205321L_197189.
Low resolution Ground truth Base model Attention model
Figure 10 compares velocity contour predictions from the base and attention-based models across carotid artery regions, including the bifurcation, ICA/ECA branches, and CCA.
- Velocity contours are compared between the base and attention-based super-resolution models across four carotid artery slices.Slice 1 covers the bifurcation, Slice 2 covers the ICA and ECA branches, and Slices 3 and 4 cover the CCA.