Performance Comparison of Deep Learning Approaches for Left Atrium Segmentation from LGE-MRI Data

Quantification of viable left atrial (LA) tissue is a reliable information which should be used to support therapy selection in atrial fibrillation (AF) patients. Late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) is employed for the non-invasive assessment of LA fibrotic tissue. Unfortunately, the analysis of LGE-MRI relies on manual tracing of LA boundaries. This task is time-consuming and prone to high inter-observer variability. Therefore, an automatic approach for LA wall detection would be very helpful. In this study, we compared the performance of different deep architectures - U-Net and attention U-Net (AttnU-Net) - and different loss functions - Dice loss (DL) and focal Tversky loss (FTL) to automatically detect LA boundaries from LGE-MRI data. In addition, AttnU-Net was trained without deep supervision (DS) and multi-scale inputs (MI), with DS and with DS+MI. No statistically significant differences were found training the networks with DL or FTL. U-Net was the best-performing algorithm overall, outperforming significantly AttnU-Net with a Dice Coefficient of 0.9015±0.0308 (mean ± standard deviation). However, no significant differences were found between U-Net and AttnU-Net DS/DS+MI. Based on these results, using a DL or FTL does not affect the performance and U-Net was the best-performing solution.