Studying the relation between post-infarction ventricular arrhythmia and left ventricular myocardium thinning on computed tomography images using explainable deep learning. (19th May 2022)

• Record Type: Journal Article
• Authors: Ly, B; Finsterbach, S; Nunez Garcia, M; Bacoyannis, T; Cedilnik, N; Sridi, S; Jais, P; Sermesant, M; Cochet, H
• Is Part Of: Europace. Volume 24:Supplement 1(2022)
• Publication Date: 2022-05-19
• Subjects: Arrhythmia -- Treatment -- Periodicals; Cardiac pacing -- Periodicals; Catheter ablation -- Periodicals; Heart -- Physiology -- Periodicals; Electrophysiology -- Periodicals; 617.4120645
• DOI: 10.1093/europace/euac053.560
• Languages: English
• ISSNs: 1099-5129

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: Public Institution(s). Main funding source(s): National Research Agency, 3IA Côte d'Azur (ANR-19-P3IA-0002), IHU Liryc (ANR- 10-IAHU-04) Background: Infarct heterogeneity plays a critical role in the development of scar-related ventricular arrhythmias (VA). Wall thickness (WT) from CT was shown to correlate with arrhythmogenic sites in the context of ablation. Purpose: To analyze the relationship between WT distribution and the presence VAs in patients with history of myocardial infarction (MI). Methods: From 2010 to 2020, we retrospectively included consecutive patients who underwent CT more than 1 year after MI.Automated LV wall segmentation, reorientation, WT computation and flattening methods were applied to obtain 2D WT bullseye maps.The population was divided into a training set (3/4) and a testing set (1/4). On the training set, a conditional variational autoencoder (CVAE) model was trained to encode the WT map in its latent representation, which was then used by a classifier model to predict VA. For each prediction, a gradient back-projection method was used to generate attention maps highlighting the bullseye region most influential in the model's decision. The ability of the trained CVAE to identify patients with VA was then assessed on the test population, and compared to that of other clinical variables (age, gender, LVEF, scar size). Results: 641 patients were included (age 73±7 years, 83% males, LVEF 46±10%), including 166 (26%) with history of VA. From original CT images, automated processing methods allowed for the obtention of a WT bullseye, a VA prediction and an attention map in less than 2 min. On the testing population, univariable correlates of VA were LVEF (P<0.001), scar size defined as WT area (P<0.001), CVAE prediction (P<0.001), and male gender (P=0.007). Multivariable analysis identified CVAE prediction and male gender as independent VA correlates (P<0.001 and P=0.01, R2=0.364), while LVEF and scar size were not (P=0.052 and P=0.60). The CVAE model identified patients with VA with a sensitivity/specificity of 0.87/0.74. The analysis of attention maps confirmed that the CVAE actually used the MI region to predict VA, and showed that the CVAE was filtering out areas of physiological WT, thereby outperforming simple WT thresholding. Conclusions: Fully automated analysis of WT from CT images is feasible in patients with MI, and allows for the extraction of robust markers of VA, outperforming conventional risk stratifiers