SCMR-ISMRM Workshop: 05 - Scientific Abstracts (AI and Machine Learning & Workflow)-Discussion

Video Transcription

Video Summary

The session featured Q&A for several CMR AI talks. For the LGE segmentation study, outliers were mainly caused by thrombi, outflow tract errors, apex partial-volume artifacts, pericardial effusion, and poor-quality images, especially in multicenter data from different vendors and protocols. Regarding clinical impact, the speaker noted that human inter-observer variability is already non-negligible, so small machine-learning differences may fall within that range.  <br /><br />For synthetic tag-tracking data, the team varied tag spacing, resolution, TR/TE, and motion, and found that mixing synthetic with real data improved performance slightly.  <br /><br />For cardiac view classification, the team used CNNs rather than simpler models, achieving strong results with lots of data. They discussed using the system for offline quality control now, with future potential for real-time acquisition feedback.  <br /><br />The final discussion highlighted future AI applications in CMR: end-to-end analysis and enabling more complex sequences to be used clinically.

Keywords

CMR AI

LGE segmentation

outliers

synthetic tag-tracking

cardiac view classification