Accelerated 4D-flow MRI with 3-point encoding enabled by machine learning.

Area of Science:

• Medical Imaging
• Machine Learning in Radiology

Background:

• 4D-flow MRI is crucial for cardiovascular and neurovascular imaging.
• Acquisition time is a limitation in 4D-flow MRI, often requiring four flow encodings.
• Accelerating 4D-flow MRI could improve patient comfort and throughput.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for accelerating 4D-flow MRI.
• To enable velocity mapping using only three flow encodings, eliminating the need for a fourth reference scan.
• To investigate CNN-based acceleration for 4D-flow MRI without background phase correction.

Main Methods:

• A 3D U-net architecture CNN was trained using 144 neurovascular 4D-flow MRI scans.
• The CNN processed complex images from three flow encodings to predict three velocity components.
• Loss function optimization (magnitude, complex difference, uniform velocity weighting) and cross-validation were performed.

Main Results:

• CNN-derived 3-point velocities showed excellent agreement with 4-point velocities (R² = 0.992).
• Optimized training focused on vessel velocities, improving intra-vessel correlation.
• Uniform velocity weighting in the loss function yielded the lowest error; denoising effect observed in accelerated data.

Conclusions:

• Machine learning, specifically CNNs, can significantly accelerate 4D-flow MRI.
• Acquisition of three-directional velocity maps is feasible with only three flow encodings.
• This method reduces errors and scan time, offering a practical acceleration strategy.