A deep learning model to enhance temporal information for the cardiac cine imaging

Motivation: Cine CMR is essential in the diagnosis of cardiac structure and function, but its long breath-holding time can be a burden for patients. Reducing temporal resolution allows reducing scan time but introduces artifacts. Goal(s): This study proposes a deep learning model to improve image quality of fast, low temporal resolution cine CMR. Approach: A cascaded neural network is implemented with the main cascade to recover signals in the y-t domain and the second cascade to eliminate artifacts in the x-y domain. Results: The model can restore image quality of the low temporal resolution series on a variety of field strengths(0.6T, 1.5T, and 3.0T). Impact: The proposed deep-learning based temporal interpolation for the cine CMR is independent of and compatible with existing acceleration strategies such as SENSE, Compressed Sensing and GRAPPA allowing a further increase in acceleration rate without the need for the raw data.