Deep learning for temporal super-resolution 4D Flow MRI.

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Artificial Intelligence in Medicine

Background:

• 4D Flow MRI enables volumetric, time-resolved blood flow quantification but faces limitations in temporal resolution, noise, and acquisition time.
• Accurate capture of transient flow variations is crucial for clinical applications but is hindered by coarse temporal resolution.

Purpose of the Study:

• To implement and evaluate a residual data-driven network for temporal super-resolution in 4D Flow MRI.
• To address the largely unexplored area of temporal super-resolution in 4D Flow MRI using deep learning.

Main Methods:

• A residual network (4DFlowNet) was adapted for temporal upsampling by modifying input dimensions and optimizing internal layers.
• The network was trained and tested on synthetic in-silico 4D Flow MRI data and further evaluated on in-vivo clinical datasets.

Main Results:

• The network achieved excellent performance, effectively denoising and temporally upsampling velocities with a mean absolute error of 1.0 cm/s on unseen in-silico data.
• It outperformed deterministic methods like linear (MAE=2.3 cm/s) and sinc interpolation (MAE=2.6 cm/s).
• The network successfully synthesized high-resolution temporal information from low-resolution in-vivo data, showing strong correlation at peak flow frames.

Conclusions:

• Data-driven neural networks show significant potential for temporal super-resolution in 4D Flow MRI.
• This approach enables high-frame-rate flow quantification, overcoming clinical limitations without extending acquisition times.