Deep learning for temporal super-resolution 4D Flow MRI

TL;DR

This work introduces a residual CNN redesigned for temporal super-resolution of 4D Flow MRI (extending beyond prior spatial-only approaches) and validates it on synthetic CFD-derived data and in-vivo acquisitions. The model denoises and up-samples temporal frames using 2D+t input patches and a tailored loss combining $L_{MSE}$ and a mutually projected $L_{1}$ term, achieving an average MAE of around $1.0$ cm/s in unseen in-silico data and strong agreement with high-resolution references in vivo. A realistic CFD-to-MRI pipeline and comprehensive patch-based training enable generalization across patient anatomies, supporting potential clinical deployment to obtain high-frame-rate flow quantification without longer scans. The study also establishes a baseline temporal SR framework for 4D Flow MRI and discusses limitations and directions for future enhancement, including more diverse datasets and advanced spatiotemporal architectures.

Abstract

4D Flow Magnetic Resonance Imaging (4D Flow MRI) is a non-invasive technique for volumetric, time-resolved blood flow quantification. However, apparent trade-offs between acquisition time, image noise, and resolution limit clinical applicability. In particular, in regions of highly transient flow, coarse temporal resolution can hinder accurate capture of physiologically relevant flow variations. To overcome these issues, post-processing techniques using deep learning have shown promising results to enhance resolution post-scan using so-called super-resolution networks. However, while super-resolution has been focusing on spatial upsampling, temporal super-resolution remains largely unexplored. The aim of this study was therefore to implement and evaluate a residual network for temporal super-resolution 4D Flow MRI. To achieve this, an existing spatial network (4DFlowNet) was re-designed for temporal upsampling, adapting input dimensions, and optimizing internal layer structures. Training and testing were performed using synthetic 4D Flow MRI data originating from patient-specific in-silico models, as well as using in-vivo datasets. Overall, excellent performance was achieved with input velocities effectively denoised and temporally upsampled, with a mean absolute error (MAE) of 1.0 cm/s in an unseen in-silico setting, outperforming deterministic alternatives (linear interpolation MAE = 2.3 cm/s, sinc interpolation MAE = 2.6 cm/s). Further, the network synthesized high-resolution temporal information from unseen low-resolution in-vivo data, with strong correlation observed at peak flow frames. As such, our results highlight the potential of utilizing data-driven neural networks for temporal super-resolution 4D Flow MRI, enabling high-frame-rate flow quantification without extending acquisition times beyond clinically acceptable limits.

Figures (8)

• Figure 1: Illustration of the proposed temporal super-resolution network for 4D Flow MRI, based on the previously described 4DFlowNet Ferdian20204DFlowNet:Dynamics. The residual convolutional neural network takes a sequence of 2D low resolution, noisy input velocities $\Tilde{v}_x$, $\Tilde{v}_x$, $\Tilde{v}_x$, magnitude, speed and a phase-contrast magnetic resonance angiogram (PC-MRA) mask, and outputs denoised and super-resolved velocities $\hat{v}_x$, $\hat{v}_x$, $\hat{v}_x$.
• Figure 2: Overview of in-silico data and patch generation for one of six data models. Left to right: voxelized geometry (2 mm resolution), velocity field (0–68 cm/s), and patch selection. Patches generate 2D slice sequences ($2D +t$) in all three Cartesian directions across multiple time frames.
• Figure 3: Overview of the synthetic MRI data preparation pipeline. (A) Combination of phase and magnitude images into a complex image. (B) Coil sensitivity maps are generated from Biot-Savart simulations, each of which is multiplied by the complex image. (C) Application of FFT and addition of complex Gaussian noise in the k-space. (D) Sampling of frequency domain data with a Cartesian variable-density phyllotaxis sampling pattern. Temporal subsampling is achieved by accumulating the readout pattern for consecutive frames. (E) Compressed sensing reconstruction of the undersampled k-space data (F) Extraction of velocity and magnitude data from the reconstructed image.
• Figure 4: In-silico evaluation on the test set. From left to right: (A) Comparison of $V_y$ between low resolution (LR), high-resolution (HR), super-resolution (SR) and sinc interpolation during a few selected time frames at early diastole. (B) Linear regression plot for one of the peak synthesized frames at early diastole. (C) Evolution of linear regression parameters $k$ and $R^2$ over the cardiac cycle.
• Figure 5: Flow evaluation over the defined aortic and mitral valve planes of the in-silico test set. Qualitative evaluation comparing LR, HR and SR results for aortic and mitral valve planes (A, C). Mean velocity [m/s] across outlets plotted as a function of time (B, D).