DEEP LEARNING-ENABLED ECHOCARDIOGRAPHIC ASSESSMENT OF BIVENTRICULAR EJECTION FRACTIONS
Szijártó Ádám
Cardiovascular Medicine and Research Division
Dr. Merkely Béla
SE Városmajori Klinikák Tanterme
2026-06-16 15:00:00
Cardiovascular Disorders: Physiology and Medicine of Ischaemic Circulatory Diseases
Dr. Merkely Béla
Dr. Kovács Attila és Dr. Tokodi Márton
Dr. Ágg Bence
Dr. Goda Márton Áron
Dr. Ferdinandy Péter
Dr. Barabás János Imre
Dr. Kőszegi Zsolt
Accurate assessment of LV and RV function is pivotal in cardiac imaging. To support
clinicians in this routine task, we developed QUEST-EF, a vendor-independent and
segmentation-free DL tool capable of fully automated prediction of LVEF and RVEF
from single-view echocardiographic videos.
QUEST-EF was implemented as a comprehensive end-to-end DL pipeline comprising a
multi-stage preprocessing module and an EF prediction module with two video vision
transformers. The transformers were first pre-trained in a self-supervised manner on a
large set of unlabeled A4C videos, using a novel ROI-aware masking strategy, designed
to improve the suboptimal pre-training performance of conventional tube masking. By
ignoring irrelevant areas of the frames, this approach substantially improved pre-training
efficiency and enhanced performance in both image reconstruction and downstream tasks.
In the subsequent supervised learning phase, one of the transformers was fine-tuned for
LVEF prediction using the publicly available EchoNet-Dynamic dataset and a dual-center
echocardiographic dataset, while the other was trained for RVEF prediction only on the
latter. QUEST-EF was externally validated in patients with acquired or congenital cardiac
diseases from four international centers and healthy adults from six continents enrolled
in the WASE study. Associations between QUEST-EF-predicted EF values and 10-year
all-cause mortality were also analyzed in a community-based cohort.
During internal and external validation, QUEST-EF exhibited robust performance in
predicting LVEF and RVEF and accurately detected LV and RV dysfunction. Among
patients with available outcome data, the predicted EF values were associated with the
composite endpoint of heart failure hospitalization or all-cause death. In the communitybased
cohort, the predictions were also associated with 10-year all-cause mortality,
independent of the Framingham Risk Score and LV diastolic function.
In summary, we successfully developed QUEST-EF, a dual-task DL model that enables
rapid, automated, and accurate assessment of biventricular EFs from A4C
echocardiographic videos, allowing efficient screening for LV and RV dysfunction.
