Generating Synthetic Late-Gadolinium Enhancement Images for Training Scar Segmentation Networks

Motivation: The quality and accuracy of scar segmentation networks is heavily dependent on the size of the training dataset. Goal(s): To generate a synthetic dataset of LGE images to address the scarcity of annotated in-vivo datasets. Approach: Two conditional GANs generated synthetic labels (healthy myocardium and scar) and realistic LGE signals, forming a dataset to train and evaluate nnU-Net models against in-vivo and combined data. Results: The network trained on only synthetic data and tested on in-vivo LGE images, achieved a Dice score of 0.77, higher than that of the network trained on only in-vivo data (0.75). Impact: This study presents a GAN-based framework to generate high-variability synthetic LGE images for scar segmentation, achieving realistic scar patterns and outperforming real-image networks and standard augmentation.