FlowMRI-Net: A generalizable self-supervised 4D flow MRI reconstruction network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Reconstructing 4D flow MRI from undersampled data is slow and can lead to velocity underestimation.
• This limits the clinical applicability of 4D flow MRI.

Purpose of the Study:

• Develop a generalizable, self-supervised deep learning framework for fast and accurate 4D flow MRI reconstruction.
• Demonstrate its utility in aortic and cerebrovascular applications.

Main Methods:

• Proposed FlowMRI-Net, a physics-driven unrolled optimization using a complex-valued convolutional recurrent neural network.
• Trained in a self-supervised manner and evaluated on multi-vendor data with varying undersampling factors (R=8, 16, 24).
• Compared against compressed sensing (CS-LLR) with ablation and quantitative/qualitative analyses.

Main Results:

• FlowMRI-Net outperformed CS-LLR in aortic 4D flow MRI reconstruction.
• Achieved significantly lower errors in velocity measurements in the thoracic aorta.
• Demonstrated generalizability for cerebrovascular 4D flow MRI with reconstruction times of 3-7 minutes.

Conclusions:

• FlowMRI-Net enables rapid and precise reconstruction of highly undersampled 4D flow MRI.
• The framework is effective for both aortic and cerebrovascular applications.
• Potential for broader application in other vascular territories.