Artificial Intelligence-Enabled Echocardiographic Assessment of Right Ventricular Function

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Right ventricular (RV) function is crucial for cardiovascular health but challenging to assess via echocardiography due to complex anatomy and location.
• Limited inter-observer reproducibility hinders accurate RV function evaluation.

Purpose of the Study:

• To develop EchoNet-RV, a novel deep learning model for automated RV segmentation and RV fractional area change (RVFAC) estimation from echocardiographic videos.
• To evaluate EchoNet-RV's performance against expert annotations and existing models on diverse datasets.

Main Methods:

• Trained EchoNet-RV on 7,169 expert-annotated apical 4-chamber view (A4C) echocardiographic videos.
• Validated the model on internal (1,320 videos) and two international external test sets (3,107 and 1,077 videos).
• Assessed the association between predicted RVFAC and clinical outcomes (heart failure hospitalization or all-cause death).

Main Results:

• EchoNet-RV achieved high Dice coefficients for RV segmentation (0.788-0.893) and low mean absolute errors for RVFAC estimation (5.795-6.362 percentage points).
• Prediction error was significantly lower than inter-observer variability (p<0.001).
• The model demonstrated strong performance in identifying RV dysfunction (AUCs 0.684-0.859) and outperformed two multi-task models.

Conclusions:

• EchoNet-RV offers rapid, automated RVFAC assessment with performance exceeding inter-observer variability.
• The model shows significant potential as a valuable tool for RV function evaluation and disease surveillance in clinical practice.
• Predicted RVFAC values were inversely associated with adverse clinical outcomes, highlighting prognostic value.