SCMR/ISMRM Workshop: Data-driven image reconstruction and motion compensation-Impact of Choice of Deep Learning Architecures in Accelerated Cardiac T1 Mapping with MyoMapNet

Impact of Choice of Deep Learning Architecures in Accelerated Cardiac T1 Mapping with MyoMapNet

Video Transcription

Video Summary

The study compared deep learning architectures for accelerated cardiac T1 mapping in MyomapNet, which estimates T1 from only four MOLLI/LL4 images. Using data from 746 subjects plus phantom experiments, the authors tested fully connected networks, VGG19, ResNet50, U-Net, and ResUnit. Fully connected and U-Net models produced T1 maps with accuracy and precision similar to standard MOLLI, in both native and post-contrast scans. ResUnit showed anatomically plausible images but larger T1 errors, while VGG19 and ResNet50 failed to estimate T1 reliably. Overall, simpler architectures performed best, challenging the assumption that more complex CNNs are superior.

Keywords

cardiac T1 mapping

deep learning

MOLLI

MyomapNet

ResUnit