Oral Abstract Session 04: Early Career Awards 3: Basic Science (2018)-Convolutional recurrent neural networks for dynamic cardiac MR image reconstruction

Convolutional recurrent neural networks for dynamic cardiac MR image reconstruction

Video Transcription

Video Summary

The talk presents a convolutional recurrent neural network (CRNN) for dynamic cardiac MRI reconstruction from undersampled k-space data. Instead of using separate CNNs at each unrolled optimization step, the method shares recurrent parameters across iterations, reducing redundancy and computation. It combines CRNN-based denoising with data consistency layers and uses a bidirectional CRNN to exploit both temporal redundancy and iterative reconstruction history. Evaluated on fully sampled cardiac cine data with simulated undersampling, the approach outperformed classical methods and prior deep learning models in accuracy, speed, and parameter efficiency, especially in motion-affected myocardium regions.

Keywords

convolutional recurrent neural network

dynamic cardiac MRI reconstruction

undersampled k-space data

data consistency layers

bidirectional CRNN