3D Segmentation of Multi-Contrast Cardiac Magnetic Resonances With Topological Correction and Synthetic Data Augmentation

Purpose: Automatic segmentation of cardiac magnetic resonance (CMR) images improves the evaluation of heart structure and function, helping clinical diagnosis and the generation of in silico models. Recent advances have introduced synthetic augmentation (SA) using generative adversarial networks (GANs) and topological correction (TC) via persistent homology to enhance segmentation with convolutional neural networks (CNNs). However, their combined effectiveness remains unexplored. Here, we extend and systematically evaluate these techniques, individually and in combination, for the first time in the context of three-dimensional (3D) CMR segmentation across challenging multi-vendor, multi-center, multi-class and multi-contrast data sets. Methods: Data involved anisotropic, topologically inconsistent cine and late gadolinium-enhanced (LGE) CMRs, and isotropic, topologically consistent ex vivo CMRs. Topological priors were defined in each data set from ground truth label (GTL) assessments, and TC was applied by retraining the baseline 3D CNN with a loss function accounting for topological discrepancies. For SA, deformed GTLs were used to generate synthetic images using trained 3D GANs. Results: Consistent segmentation improvements were observed for the ex vivo data in both overlap with GTLs and topological precision when applying TC and SA individually and in combination. Notably, an enhanced identification of the infarction was obtained when SA and TC were used in the LGE data. Overall, SA increased the predictions overlap with GTLs, while TC reduced the topological discrepancies across all data sets. Conclusion: TC and SA demonstrate strong potential for improving 3D CMR segmentation on complex, real-world data sets, especially when topologically consistent data are available for training.