Comparison of synthetic LGE with optimal inversion time vs. conventional LGE via representation learning:...

Late Gadolinium Enhancement (LGE) images are crucial elements of CMR protocols for evaluating myocardial infarct (MI) severity and size. However, these images rely on signal intensity changes and manual inversion time (TI) settings, leading to suboptimal lesion/remote contrast in many cases. Here, we propose an original approach to evaluate the impact of suboptimal TI on the retrospective analysis of ST-elevation MI (STEMI) patients, using a representation learning methodology tailored to consider infarct- and image-based characteristics across the studied population.

We analyzed 133 pairs of conventional and synthetic LGE short-axis images from the HIBISCUS-STEMI cohort (ClinicalTrials ID: NCT03070496). Optimal TI was identified among co-registered synthetic LGE images, using a mixture of the Mann-Whitney U-test, standard deviation, and saturation of pixel values, while the TI used for conventional LGE image generation was extracted from the DICOM header. Images were realigned to a reference for pixel-wise inter-subject comparisons. Population analysis relied on Attribute-based Regularized Variational Autoencoders which provide a latent representation of the population that is both easier to analyze (lower dimensionality) and ordered by infarct-relevant attributes.

Despite visual quality control in the clinic, our study demonstrates that nearly 50% of conventional LGE slices may include a suboptimal TI setting, mostly related to TI settings shorter than the optimal TI determined from synthetic LGE. Additionally, our findings showed that when isolating contrast effects and suboptimal TI settings, contrast had a minimal impact on infarct lesion metrics such as infarct size or transmurality in the latent space. This suggests that other factors than contrast setting are leading (for both cases) to systematic and proportional bias (p<0.05) and loss of precision (respectively ρ=0.42 and ρ=0.43) in the latent space.

Suboptimal TI undermines the analysis of infarct patterns in populations. Representation learning is a powerful method to retro-analyze cohorts, enabling the identification of imperfect settings, a crucial step for accurately characterizing representative patterns of a population. Our strategy can be considered a promising candidate for monitoring longitudinal changes and evaluating therapy outcomes on broader populations.