Super Resolution of Cardiac Cine MRI Sequences Using Deep Learning

摘要

Cardiac cine MRI facilitates structural and functional analysis of the heart through the dynamic aspect of the sequences. Clinical acquisitions consist of sparse 2D images instead of 3D volumes, taken at landmark points of the ECG to cover the whole heartbeat. A stack of short axis images and a small number of long axis views are generally acquired. Efforts have been made to accelerate acquisitions at the acquisition stage as well as at post-processing. A major part of current research in medical image processing focuses on deep learning approaches driven by large datasets. However, most of those methods leave out the dynamic aspect of temporal data and treat frames of cine MRI sequences individually. We propose a super resolution network based on the U-net and long short-term memory layers to exploit the temporal aspect of the dynamic cardiac cine MRI data. When given a sequence of low resolution long axis images, our method is able to render a high resolution sequence. Results on synthetic data simulating a stack of short axis images show quantitative and qualitative improvements over traditional interpolation methods or the equivalent machine learning method using a single frame, including the ability of the network to recover important image features such as the apex.