SCMR/ISMRM Workshop: Data-driven image reconstruction and motion compensation-Deep Learning-based Reconstruction for Highly Accelerated 4D Flow MRI

Deep Learning-based Reconstruction for Highly Accelerated 4D Flow MRI

Video Transcription

Video Summary

The speaker presented a deep learning method for reconstructing highly accelerated 4D Flow MRI of the thoracic aorta. Compared with compressed sensing, the CNN produced better magnitude image quality and improved agreement with GRAPA ground truth for peak velocity, net flow, and peak flow across acceleration factors 5.7, 7.7, and 10.2. The model used 3D U-Net/DenseNet architecture, MSE plus SSIM loss, and masking to reduce noise effects. Results showed reduced underestimation and better hemodynamic quantification, though some image blurring remained. Future work will improve sharpness and test larger cohorts and higher acceleration factors.

Keywords

4D Flow MRI

thoracic aorta

deep learning reconstruction

3D U-Net DenseNet

hemodynamic quantification