EchoMed Agent needs only 20 labeled cases to estimate ejection fraction, cutting median error to 4.33%

Left ventricular segmentation and ejection fraction estimation in echocardiography can be done with very little site-specific annotation, and a clinician can obtain a structured report by asking in natural language. Fully supervised models previously assumed abundant labeled data at the target institution, while zero-shot transfer to a new site left ejection fraction errors well outside clinical bounds. On the 98-case Mayo Clinic cohort, adaptation with only 20 labeled cases gave a median absolute ejection fraction error of 4.33% with 83.3% of cases within ±10 percentage points, versus 5.47% and 80.6% without that site adaptation.

Segmentation feeds directly into functional quantification: frame-by-frame masks become a cavity-area curve, which locates end-diastole and end-systole. The single-plane area-length method needs both mask area and the long-axis length from apex to mitral valve center, and deriving that length by ellipse fitting or the mask major axis is sensitive to irregular boundaries. AxisHead regresses the two long-axis endpoint coordinates from the mask and image features; the module has about 33K parameters and is trained on EchoNet-Dynamic pseudo-labels, then calibrated under the same few-shot protocol.

The interaction layer maps a clinician's natural-language request onto registered analysis functions instead of letting the model freely generate analysis code. Earlier segmentation and quantification pipelines required programmatic calls, so clinicians could not use them through conversation. Across four representative query categories, end-diastolic and end-systolic phase identification and ejection fraction computation were routed correctly; a semantically adjacent request for pumping efficiency triggered clarification before re-routing, and a request for ventricular wall strain analysis was declined as unsupported.

The demonstrated scope is left ventricular segmentation and single-plane area-length ejection fraction estimation, evaluated on EchoNet-Dynamic, a 98-case Mayo Clinic cohort and the public CAMUS benchmark. CAMUS provides frame-level contours without full-cycle video, so only end-diastolic and end-systolic segmentation is evaluated there, not ejection fraction. The interface is aimed at clinicians and invokes registered functions for segmentation, phase extraction and ejection fraction computation through natural language.

Ejection fraction errors are heavy-tailed, driven by a small subset of failure cases, and qualitative inspection placed most discrepancies at end-systole and in reduced-ejection-fraction cases with abnormal wall motion. No prespecified disease-specific or image-quality subgroup analysis was performed, so it is unknown whether these outliers are associated with particular diagnoses, apical structural abnormalities or poor acoustic windows. The routing layer uses structured hard-coded decision logic rather than open-ended reasoning, the Mayo cohort comes from a single institution, and generalization across sites and vendors still needs validation.