Automatic background offset correction of cardiovascular 4D flow MRI data using Deep Learning

Motivation: Several approaches for background offset correction in 4D flow MRI exist, but none has proven fully effective. Best results are obtained with a post in-vivo phantom measurement as correction, but this, however, doubles the scan time. Goal(s): To develop an automatic, deep learning-based background offset correction method for cardiovascular 4D flow MRI data, that uses static phantom measurements as ground truth. Approach: The method consists in training a convolutional neural network with static-phantom measurement as ground truth. Results were compared to polynomial fit correction methods. Results: The proposed method outperformed the conventional polynomial fit methods, and was comparable to the ground-truth phantom-based correction. Impact: The proposed fully automated CNN-based background offset correction method outperformed the conventional background offset correction methods that use a polynomial fit to static tissue. This method has potential for significantly improving the data quality of cardiovascular 4D flow MRI.