Automatic left-ventricular view detection and ejection fraction assessment by artificial intelligence models in echocardiography

Background Left-ventricular (LV) size and ejection fraction (LVEF) play a crucial role in the diagnosis and risk stratification of several cardiovascular diseases. Their current assessment on echocardiography images has substantial inter-observer variability, possibly impacting patients management. Full-automatization by artificial intelligence (AI) models may improve LV size and LVEF reproducibility and permit their quantification by non-experts. Purpose To develop AI models for the identification of relevant echocardiography views, segment the LV in 2, 3 and 4-chamber views and compute LVEF. Methods Fifteen thousand echocardiography studies obtained during patients care were retrospectively identified, retrieved and anonymized. Via commercial clinical software, 619 videos (14082 frames) of 2-, 3- and 4-chamber views were annotated for LV internal and external borders, creating three regions of interest (LV cavity, LV wall and overall LV), and divided into independent training (465 videos) and testing (154) sets. LV volumes on 4-chamber views were used to assess LVEF, which was validated against clinical report data in an internal cohort of 488 patients and in an external cohort of 500 patients from the CAMUS open dataset. Results Demographic and clinical characteristics of the 488 internal cohort patients are included in Table 1. View detection was obtained with 93% accuracy. The segmentation of LV cavity, overall LV and LV wall were good in 2-chamber (Dice score of 0,86[0,79;0,90], 0,91[0,86;0,93], 0,79[0,74;0,83], respectively), 3-chamber (0,88[0,84;0,91], 0,91[0,90;0,93], 0,81[0,77;0,83]) and 4-chamber (0,90[0,86;0,93], 0,92[0,88;0,94], 0,82[0,79;0,85]) views. Error analysis revealed that segmentation performance was lower in images with low quality and in patients with atrial fibrillation, with no differences between sexes. Similarly, performance of these segmentation tasks was good in the external validation cohort, with Dice score of 0,91[0,87;0,94] and 0,80[0,73;0,84] for whole LV and LV cavity in 2 and 4-chamber views, respectively. LVEF predictions showed an acceptable linear association (p<0.001) but substantial underestimation (mean error = 12%) in the internal validation set, and a good linear association (p<0.001) and minimal underestimation (mean error = 2.2%) in the external validation set. Conclusions AI models perform well in echocardiography views identification and LV segmentation, resulting in LVEF predictions with errors in the order of inter-observer variability. Biases may be present in patients with atrial fibrillation or in videos of limited image quality.Table 1.Demographic and clinical data