CURRENT fills in both space and time for aortic 4D flow MRI with one recurrent network, sharpening relative pressure estimates

A single recurrent network can perform spatial, temporal or joint super-resolution as needed, folding what previously took two dedicated models or a two-stage pipeline into one framework. Spatial super-resolution had been handled frame by frame by 4DFlowNet and temporal super-resolution by T4DFlowNet, so spatiotemporal super-resolution required chaining the two in sequence. On synthetic aortic data, CURRENT reached a directional error of 0.072±0.037 for spatial super-resolution versus 0.097±0.053 for 4DFlowNet, and 0.062±0.031 for temporal super-resolution versus 0.064±0.031 for T4DFlowNet.

The payoff of spatiotemporal super-resolution appears downstream: relative pressure estimated from its reconstructed velocity fields lands closer to the high-resolution reference than estimates from spatial-only or temporal-only super-resolution. Relative pressure depends on velocity derivatives in both space and time, so super-resolving one direction alone recovers only part of the information. Against relative pressure from the spatiotemporal high-resolution reference, the linear fit for spatiotemporal super-resolution had a slope closer to identity, lower bias and higher R², while spatial-only and temporal-only super-resolution underestimated more strongly.

The network operates on 3D-plus-cardiac-phase patches of velocity vectors and passes hidden states between neighboring cardiac phases through bidirectional recurrent layers, exploiting frame-to-frame correlation. Frame-wise convolution sees only a single cardiac phase, leaving temporal redundancy unused. The network uses eight bidirectional convolutional recurrent layers in low-resolution space and four in high-resolution space, with learnable short and long skip connections around a central upsampling layer, at 3.2M parameters.

Trained only on synthetic data, the model was applied directly to in vivo aortic 4D flow MRI from one healthy volunteer and produced less velocity noise than sinc and linear interpolation. In vivo high-resolution reference data cannot be acquired directly, so training relies on paired data made by embedding computational fluid dynamics velocity fields into in vivo backgrounds and downsampling. In vivo data were retrospectively downsampled to 2.5 mm isotropic and 49.2 ms as input; linear regression gave higher R² than 4DFlowNet, though all methods underestimated high velocities.

The demonstrated scope is two-fold spatial, temporal and spatiotemporal super-resolution on synthetic aortic 4D flow MRI, plus a feasibility check on in vivo data from one healthy volunteer; the authors propose distributional learning and fine-tuning on paired in vivo data as next steps for domain generalization.

The synthetic data do not yet include eddy currents, drifts, probe spectra tied to specific encoding gradient waveforms, or spatiotemporal jitter from heart-rate and hemodynamic variation; embedding computational fluid dynamics fields into in vivo backgrounds left some magnitude-phase misalignment along the vessel wall that may affect near-wall velocities; and no physical regularization was built into the network, so reconstructed fields may not obey fluid-mechanical principles.