SCMR-ISMRM Workshop: 05 - Scientific Abstracts (AI and Machine Learning & Workflow)-Lecture 1 - Deep Learning-based Synthetic Data Generation Improves the Robustness of Automatic Cardiac Function Estimation

Lecture 1 - Deep Learning-based Synthetic Data Generation Improves the Robustness of Automatic Cardiac Function Estimation

Back to course

Video Transcription

Video Summary

The talk presents a method to improve automatic cardiac function estimation by generating synthetic cardiac MRI data with deep learning. Because annotated CMR data are limited and EF values are imbalanced, the authors manipulate segmentation masks to create about 20,000 synthetic subjects with more uniform ejection fraction distribution. They use GauGAN to translate masks into realistic images, then train an EF prediction network in three stages: on real data, on synthetic data, and finally fine-tuned on real data. Pretraining on synthetic data reduces errors, especially for low EF cases, and the final model outperforms the Kaggle competition winner.

Keywords

cardiac MRI

synthetic data

ejection fraction

deep learning

GauGAN