Convolutional LSTM: A Deep learning approach for Dynamic MRI Reconstruction

Abstract

Dynamic Magnetic Resonance Imaging (MRI) has been a choice of modality in capturing time-varying anatomical structures of different organs within the body in a sequential format. However, its applications are limited by slower acquisition time because of both physical and physiological constraints. Dynamic MRI is proved to have spatio-temporal redundancy in its frequency domain(k-space). The acquisition period can be minimized significantly by reducing the number of k-space samples but at the cost of introduction of artifacts in the corresponding image domain. The proposed work develops a cascaded Convolutional Long Short Term Memory (ConvLSTM) architecture for reconstructing T2-weighted dynamic MRI sequences from highly undersampled k-space data to accelerate the overall acquisition process. In particular, fully sampled data acquired from the ADNI database will be undersampled using a Cartesian undersampling mask. ConvLSTM architecture proposed is then used to remove the aliasing artifacts introduced by undersampling. In addition, ConvLSTM model learns both spatial and temporal dependencies of the imagery to reconstruct it efficiently while outperforming the Convolutional Neural Network (CNN) based reconstruction in terms of reconstruction accuracy.