Focus Session 5: MR Techniques and Methods: Rapid Image Acquisition and Reconstruction (2019)-Ultrafast Reconstruction of UnderSampled (URUS) 3D CMR Imaging via Deep Complex Neural Network

Ultrafast Reconstruction of UnderSampled (URUS) 3D CMR Imaging via Deep Complex Neural Network

Video Transcription

Video Summary

The talk presents a deep complex convolutional network for fast reconstruction of undersampled 3D MRI data. Unlike many deep learning methods that use only magnitude images, this approach preserves phase information, which helps reconstruction quality. The model uses complex convolution, normalization, and activation layers implemented with real-valued operations in PyTorch. Tested on 219 3D-LGE cardiac MRI cases, it outperformed a magnitude-only network in MSE and SSIM, and preserved scar quantification accuracy. It also worked well at higher undersampling rates and reconstructed a 100-slice volume in about 15 seconds, much faster than compressed sensing.

Keywords

3D MRI reconstruction

complex convolutional network

undersampled MRI

phase information

cardiac MRI