SCMR/ISMRM Workshop: Data-driven image reconstruction and motion compensation-Demystifying Deep Learning CMR Reconstruction

Video Transcription

Video Summary

The session opened with welcome remarks and an overview of a five-talk program on data-driven image reconstruction and motion compensation in cardiac MRI. Florian Knorr’s lecture explained how deep learning can improve CMR reconstruction from undersampled k-space data. He contrasted classic compressed sensing, which uses handcrafted sparsity priors, with neural networks that can better model complex spatial-temporal structures and reduce aliasing and blurring artifacts. He described three main deep learning strategies: image-space post-processing, k-space synthesis, and iterative reconstructions that alternate between data consistency and learned priors. Examples from CT, MRI, and several CMR studies showed strong image-quality gains over traditional methods. He also highlighted useful resources, including public raw data sets, reconstruction challenges, and CMR data from Ohio State University for researchers beginning in this area.

Keywords

cardiac MRI

deep learning reconstruction

undersampled k-space

compressed sensing

motion compensation