A digital twin approach for automated left ventricular wall thickness assessment from 2D echocardiography

Abstract Introduction The left ventricle (LV) thickness is a critical parameter for diagnosing cardiac conditions such as hypertension or hypertrophy. Nevertheless, the current assessment of LV thickness on 2D echocardiographic images exhibits high variability, as it relies on manual user-dependent distance measurements 1. Moreover, while the LV has an intrinsic 3D structure, thickness is measured on 2D views, dependent on acquisition quality, and assessed in only a limited number of manually drawn distances. Purpose This work aims to (1) provide an enhanced assessment of LV wall thickness on standard 2D echocardiography by automating the analysis, via artificial intelligence (AI) methods; and (2) illustrate the potential of the digital twins for LV thickness 3D mapping. Methods The study included a validation cohort of 68 patients (age: 64 ± 12 years, 67% male), in whom LV thickness was manually assessed by 6 sonographers (8.5 ± 3.8 years of experience) on parasternal long axis (PLAX) and short axis (PSAX) views on 6 regions: respectively, interventricular septum (IVS) and posterior wall section (PW); and anterolateral (AL), inferior(LVI), inferolateral (LVIL) and septal section LVS), resulting in n=379 annotations. An AI segmentation model, based on the state-of-the-art nnUNet framework and trained on an independent cohort of 780 studies 2, was used to infer epicardium and endocardium borders on the validation cohort. Automated measurements of thickness were calculated minimizing distances between endocardium and epicardium, which were respectively validated against manual annotations. Linear regression and Blant-Altman analyses were performed to quantify agreement. A 3D digital twin of the patient-specific heart is computed to illustrate the potential of the automated method 3. Wall thickness maps are automatically generated in the 3D model and extracted to a representation in the form of an AHA 17-segment model bullseye plot. Results There was excellent agreement between manually annotated wall thickness and the respective automated quantified wall thickness, as seen in Figure 1 (R2 = 0.98, bias of -0.02cm and limits of agreement of -0.09 cm to 0.05 cm). This trend was also sustained for each of the 6 regions assessed (R²(IVS) = 0.96; R²(PW) = 0.990; R²(LVAL) = 0.997; R²(LVI) = 0.996; R²(LVIL) = 0.997; R²(LVS) = 0.998). Furthermore, the feasibility of the 3D digital twin assessment of myocardial thickness is shown on Figure 2 for a control, a hypertrophic (HCM) and an ischemic case (MI). The thickening and thinning patters of the HCM and MI, respectively, can be visualised with 360° fine-detail, enhancing the standard approach limited to a few 2D measurements. Conclusions The proposed AI-based automated assessment of myocardial wall thickness exhibits excellent agreement with 2D manual experts’ measurements. Furthermore, it illustrates the potential of enhancing the analysis from a few 2D measurements to a 3D personalized assessment of LV.