Trained on 186,712 sonographer-annotated echocardiographic images from 93,978 studies across 56,855 patients at Cedars-Sinai Medical Center, EchoNet-Segmentation achieved R² of 0.817–0.882 for automated velocity-time integral (VTI) Doppler measurements and R² of 0.675–0.747 for left and right atrial area segmentation. Performance held across temporal splits, an external Kaiser Permanente Northern California cohort, and the public MIMIC-Echo dataset, with the model outperforming an existing open-source medical image foundation model on R², MAE, and Dice scores.
Echocardiography remains the workhorse of cardiovascular diagnostics, yet its interpretive burden is immense — over 70 parameters requiring manual annotation with known inter-observer variability. Prior AI efforts largely tackled left ventricular ejection fraction via 2D B-mode imaging, leaving diastolic and valvular assessment — clinically critical for heart failure classification and valve disease staging — largely unautomated. EchoNet-Segmentation closes that gap by extending automation to spectral Doppler waveforms and atrial chamber sizing, parameters directly tied to diastolic dysfunction grading and stroke risk stratification.
The multi-cohort external validation across different vendor hardware meaningfully strengthens generalizability claims beyond single-institution tools. However, this is a retrospective study with no prospective clinical outcome data — it validates measurement accuracy against sonographer labels, not downstream diagnostic or prognostic impact. Open-source model release is a genuine strength for reproducibility. As a preprint not yet peer-reviewed, performance metrics and conclusions remain subject to revision. If validated clinically, this could represent a meaningful workflow shift in high-volume echo labs.