Automated left ventricular ejection fraction estimation from echocardiogram video using deep learning segmentation
Every automated ejection fraction measurement begins with a single ultrasound clip: the apical 4-chamber (A4C) view, acquired by placing the transducer near the cardiac apex and imaging up through all four chambers in one plane. The AI model needs nothing more than this raw grayscale video loop — but the loop must contain at least one complete, well-framed cardiac cycle to be usable.
The apical 4-chamber window is one of the core acoustic windows in a standard transthoracic echocardiogram, obtained by angling the transducer at the point of maximal cardiac impulse near the apex. It simultaneously visualizes the left ventricle, right ventricle, left atrium, and right atrium in a single long-axis plane, with the LV apex at the top of the image and the mitral valve annulus at the base.
This view is preferred for LV functional assessment because it captures the LV's long axis nearly end-to-end, minimizing foreshortening — a common pitfall where an angled probe makes the ventricle appear artificially short and volumes are underestimated. Sonographers are trained to maximize LV length in this view specifically because downstream volume and EF calculations are highly sensitive to it.
Many protocols pair the A4C view with an orthogonal apical 2-chamber (A2C) view, since the LV is not perfectly symmetric — combining both planes (biplane assessment) captures shape irregularities that a single view alone would miss.
A single resting cardiac cycle lasts roughly 0.8 seconds (at 75 bpm), split into a shorter systolic ejection phase (~1/3 of the cycle) and a longer diastolic filling phase (~2/3). At typical clinical frame rates of 30–60 fps, this cycle is represented by only 25–50 individual frames — a remarkably sparse sampling of a continuous mechanical process.
Because AI models like EchoNet-Dynamic are trained on fixed-length clips (commonly downsampled to 32 frames), acquisition protocols specify a minimum clip duration sufficient to guarantee at least one full R-wave-to-R-wave interval is captured, often gated to the ECG trace recorded alongside the video. Clips that are too short, too shaky, or that cut off mid-cycle degrade every downstream measurement, since the model can only find true end-diastolic and end-systolic frames if they actually appear in the video.
A single cardiac cycle at rest lasts about 0.8 seconds — captured in as few as 25–30 ultrasound frames — yet this brief loop contains all the spatial and temporal information an AI model needs to estimate global left ventricular function.
Once a usable clip is acquired, a deep convolutional neural network — typically a U-Net or DeepLabv3-style encoder-decoder — classifies every pixel of every frame as either "LV blood pool" or "not LV," producing a segmentation mask. The boundary of that mask becomes the automatically traced endocardial border, replacing what used to require manual, frame-by-frame tracing by a sonographer.
The segmentation network takes each raw ultrasound frame as input and outputs a per-pixel probability map indicating how likely each pixel belongs to the LV blood pool. An encoder progressively downsamples the image to learn coarse anatomical context (chamber shape, position, scale), while a decoder upsamples back to full resolution, fusing in fine-grained edge information via skip connections — the defining feature of the U-Net architecture.
Thresholding this probability map yields a binary mask; the mask's outer boundary is then extracted as a smooth contour (e.g., via marching-squares style contour extraction), giving a pixel-accurate endocardial trace on every single frame — something a human sonographer would only trace on two frames per clip (end-diastole and end-systole) due to time constraints.
Segmenting a single static frame is comparatively easy; segmenting all 30–50 frames of a beating, translating, foreshortening heart consistently is the harder engineering problem. Speckle noise, dropout frames, shadowing from ribs, and off-axis probe angulation can all cause a per-frame model to "flicker" — producing a contour that jumps in size between adjacent frames even though the true anatomy changes smoothly.
Modern pipelines address this with temporal smoothing, weak supervision across neighboring frames, and by training on the assumption that LV volume should trace out a smooth, physiologically plausible curve across the cycle. The result is a full frame-by-frame area (and therefore volume) curve, from which the true end-diastolic (maximum) and end-systolic (minimum) frames are automatically identified — rather than relying on a single manually chosen frame that may not be the true extremum.
On the EchoNet-Dynamic benchmark, automated LV segmentation achieved a Dice similarity coefficient of approximately 0.92 against expert-drawn contours — within the 0.90–0.95 range typically seen between two independent human sonographers tracing the same heart.
A traced 2D contour is not yet a volume. To turn an endocardial outline into a clinically meaningful chamber volume, echocardiography relies on the modified biplane Simpson's rule (method of discs) — the technique recommended by the American Society of Echocardiography (ASE) and European Association of Cardiovascular Imaging (EACVI) for quantitative LV assessment.
The Simpson's rule approach divides the traced LV chamber, along its long axis from apex to mitral annulus, into a stack of parallel discs of equal height. Each disc is treated as a short cylinder whose volume is π × (diameter/2)² × height; summing all discs approximates the total chamber volume without assuming any single idealized geometric shape (unlike older single-plane ellipsoid formulas).
Because the LV is not perfectly radially symmetric, the "biplane" version combines disc diameters measured from two orthogonal views — the apical 4-chamber and apical 2-chamber planes — so that asymmetries in one plane are captured by the corresponding measurement in the other. This is why comprehensive AI pipelines increasingly process paired A4C/A2C clips rather than a single view alone.
With a frame-by-frame endocardial contour already available from the segmentation stage, the volumetric calculation stage is largely automated arithmetic: the AI model scans the full area-versus-time curve across the cycle to identify the true end-diastolic frame (maximum traced area, occurring just after mitral valve closure) and the true end-systolic frame (minimum traced area, at aortic valve closure). The method of discs is then applied directly to each of those two contours.
When only a single apical view is available, a single-plane area-length formula is used as a fallback, trading some accuracy for feasibility — this is common in point-of-care or emergency settings where a full biplane acquisition isn't practical.
Because a single apical view cannot fully capture LV asymmetry, biplane Simpson's method (combining 4-chamber and 2-chamber discs) keeps automated volume estimates within roughly 10% of cardiac MRI — the accepted volumetric reference standard.
With EDV and ESV in hand, ejection fraction follows directly: EF = (EDV − ESV) / EDV × 100%. But a single beat is a noisy sample of a dynamic system — respiration, mild arrhythmia, and probe motion all perturb individual cycles. Robust AI pipelines therefore compute EF across every beat in the clip and report a stabilized, averaged value alongside its variability.
Ejection fraction expresses the fraction of end-diastolic blood volume that is ejected with each heartbeat: EF = (EDV − ESV) / EDV × 100. It is a measure of global systolic pump function rather than a direct measure of contractility, muscle mass, or coronary flow — a heart can have preserved EF and still suffer from significant diastolic dysfunction, or reduced EF from causes ranging from prior infarction to dilated cardiomyopathy to valvular disease.
The American Society of Echocardiography classifies EF as normal (55–70%), mildly reduced (41–54%), or reduced (≤40%), and these cutoffs directly define the heart failure phenotypes HFpEF, HFmrEF, and HFrEF used throughout cardiology guidelines and clinical trials.
Because a single cardiac cycle can be affected by a premature beat, a deep breath, or slight transducer drift, EF measured from one beat versus the next can differ by several percentage points even in the same healthy heart. This intrinsic biological and acquisition noise is compounded by measurement noise — the same clip re-traced by two different sonographers, or even the same sonographer on two different days, can produce EF estimates that differ by 7–10 percentage points.
Video-based AI models sidestep some of this by evaluating every beat available in a clip (rather than one manually chosen beat) and reporting a beat-to-beat trend or averaged EF, which is both more reproducible and less sensitive to any single unrepresentative cycle — a key reason AI-derived EF often shows tighter test-retest agreement than manual single-beat measurement.
Human inter-observer variability for LVEF by 2D echocardiography has been reported as high as ±7–10 percentage points between expert readers, while the Stanford EchoNet-Dynamic model achieved a mean absolute error of just 4.1 percentage points against expert consensus (Ouyang et al., Nature, 2020) — comparable to, or better than, typical human-to-human variability.
A number is only clinically useful once it reaches a report, a trend line, and a decision. The final stage folds the AI-derived EF — along with EDV, ESV, and confidence metadata — into the structured echocardiography report, automatically comparing it against normal ranges and prior studies, and flagging results that warrant clinical attention.
Modern echo reporting platforms use structured templates (aligned with ASE reporting standards) where each measurement — including AI-derived EF, EDV, and ESV — populates a discrete, searchable field rather than free text. This allows automated logic to run on every study: an EF at or below 40% is automatically flagged as HFrEF, prompting a suggested referral for guideline-directed medical therapy; values of 41–49% are flagged as HFmrEF; and any significant drop from a patient's prior study is highlighted regardless of the absolute value, since a falling trend can matter as much as the number itself.
Because the AI pipeline also reports a frame confidence score, low-confidence studies (noisy image, foreshortened view, arrhythmia) are flagged separately for a human sonographer or cardiologist to confirm before the number is trusted clinically.
Certain cancer therapies — most notably anthracyclines (e.g., doxorubicin) and HER2-targeted agents (e.g., trastuzumab) — carry a well-documented risk of cardiotoxicity, making serial EF surveillance a routine part of cancer treatment protocols. The 2022 ESC Cardio-Oncology Guidelines define cancer therapy-related cardiac dysfunction (CTRCD) as a drop in LVEF of 10 percentage points or more to a value below 53%.
Because these patients may undergo echocardiograms every treatment cycle or roughly every three months for months to years, the reproducibility of automated EF measurement is especially valuable: small, spurious point-to-point variation from manual tracing could otherwise trigger unnecessary treatment interruptions, while genuine early declines could be missed among measurement noise.
A drop in LVEF of 10 percentage points to below 53% during cancer therapy defines cancer therapy-related cardiac dysfunction (CTRCD) under the 2022 ESC Cardio-Oncology guidelines — a threshold that reproducible, AI-automated serial echo monitoring can detect earlier and more consistently than manual measurement alone.
Beyond the dedicated echo lab, handheld point-of-care ultrasound (POCUS) devices paired with on-device AI segmentation now allow emergency physicians and internists to obtain a reasonable EF estimate at the bedside within minutes — useful for rapidly distinguishing cardiogenic from non-cardiogenic causes of dyspnea in the emergency department or ICU. These estimates are not a replacement for a comprehensive echo study, but they can meaningfully accelerate triage.
In dedicated heart failure clinics, automated EF tracking over years of serial studies produces a longitudinal trend line that supports titration of guideline-directed medical therapy, detection of recovery (in which EF normalizes and therapy may eventually be reconsidered), and early identification of decompensation — turning a single snapshot measurement into a continuous, trackable signal of cardiac health.