SCMR-ISMRM Workshop: 05 - Scientific Abstracts (AI and Machine Learning & Workflow)-Lecture 3 - Using Synthetic Grid Tagged Images to Train a Neural Network for Tag Tracking

Lecture 3 - Using Synthetic Grid Tagged Images to Train a Neural Network for Tag Tracking

Video Transcription

Video Summary

The talk presented a synthetic-data approach for training a neural network to track motion in grid-tagged cardiac MRI. The method generates tag-like images from natural images using random deformation fields and cardiac-like motion, then trains a ResNet-based spatiotemporal model on millions of synthetic patches. In testing on a cardiac phantom and 27 in vivo slices from nine patients, the network showed strong agreement with manual tracking, with low displacement and strain errors. The approach may reduce dependence on labor-intensive manual annotation and adapt easily to different tagging protocols.

Keywords

synthetic data

cardiac MRI

motion tracking

neural network

spatiotemporal model