EchoGen: Controllable Echocardiogram Generation

• EchoGen is the synthesis of echocardiographic images and videos driven by anatomical, physiological, or pathology cues to achieve realistic and clinically useful outputs.
• Leading methods use conditional GANs, diffusion models, and latent generative pipelines to translate segmentation masks, ECG signals, and view data into synthetic echo frames or videos.
• Emerging systems demonstrate improved diagnostic utility and privacy by validating synthetic data with clinical benchmarks like ejection fraction regression and pathology classification.

In the cited literature summaries, EchoGen is used as a shorthand for echocardiogram generation: the synthesis of ultrasound frames or videos under explicit anatomical, physiological, or multimodal control. Within cardiac imaging, this usage spans segmentation-mask-conditioned frame synthesis, view-conditioned video synthesis, ECG-conditioned video generation, pathology-conditioned generation, and privacy-aware latent generative pipelines (Abdi et al., 2019, Kondori et al., 25 Aug 2025, Li et al., 2024, Reynaud et al., 28 Mar 2025, Muhammad et al., 21 Sep 2025). The term is also ambiguous, because an unrelated vision-language framework for layout-to-image generation and image grounding was later published under the same name (Zou et al., 18 Mar 2026). In the medical-imaging sense, EchoGen denotes a research direction whose central objective is not only visual realism, but also controllability, downstream clinical utility, and, increasingly, privacy-preserving data release.

1. Research problem and clinical motivation

EchoGen-style work is motivated by the practical constraints of cardiac ultrasound. Echocardiography is clinically central, but acquisition quality and interpretation depend heavily on trained operators, expert annotations are expensive, and large-scale sharing is constrained by privacy. These constraints are especially visible in semi-supervised segmentation, ejection fraction estimation, rare-pathology modeling, and point-of-care ultrasound settings, where one or more standard views may be missing or of low quality (Abdi et al., 2019, Kondori et al., 25 Aug 2025, Li et al., 2024, Reynaud et al., 28 Mar 2025).

Across the literature, synthetic echocardiography serves several distinct but related purposes. One is paired synthesis, in which a known structural condition such as a segmentation mask is mapped to a plausible image. A second is cross-view synthesis, where one real view, such as apical 4-chamber (A4C), is used to generate a complementary view, such as apical 2-chamber (A2C), for downstream ejection fraction estimation. A third is multimodal conditioning, where ECG signals or pathology labels drive the generator. A fourth is privacy-preserving release, where generated images and videos are filtered to reduce anatomical re-identification risk. This suggests that EchoGen has evolved from proof-of-concept image synthesis toward clinically oriented controllable generation.

A recurrent theme is that visual plausibility alone is insufficient. Later systems explicitly test whether synthetic echocardiograms improve segmentation, diagnosis, or regression models, and some papers frame downstream performance as the most meaningful test of generative quality (Kondori et al., 25 Aug 2025, Reynaud et al., 28 Mar 2025, Muhammad et al., 21 Sep 2025).

2. Early anatomy-conditioned frame synthesis

A clear early formulation appears in “GAN-enhanced Conditional Echocardiogram Generation” (Abdi et al., 2019). That work studies conditional echocardiogram synthesis from segmentation masks of the left ventricle, ventricular myocardium, and atrium / left atrium, and explicitly positions generation as a feasibility study for semi-supervised segmentation and anatomical regularization.

The method is a conditional GAN with a patch-based discriminator trained on paired data in the spirit of pix2pix, but with two important design choices. First, the mapping from mask to image is treated as deterministic, with no random latent vector, because each segmentation mask in CAMUS corresponds to a single target frame. Second, the adversarial objective is a least-squares GAN loss combined with a pixel-wise mean absolute error reconstruction term. The generator is described as a UNet-like network with 7 convolutional layers, 7 deconvolutional layers, and no skip connections; the discriminator has 5 convolutional layers and operates on 16 × 16 patches. Training uses Adam, batch size 8, generator learning rate 0.00013, discriminator learning rate 0.00015, adversarial weight $\lambda = 0.01$, and 100k iterations on CAMUS, using ED frames from the apical four-chamber (4CH) view after resizing to 256 × 256 and normalizing intensities to [0, 1] (Abdi et al., 2019).

Five conditioning configurations are reported: left ventricle only, left atrium only, left ventricle + ventricular myocardium, left ventricle + atrium, and left ventricle + ventricular myocardium + atrium. The principal finding is qualitative: the generator can produce plausible echo frames whose visible cardiac structures match the supplied masks, and richer conditioning increases anatomical specificity. At the same time, the study is explicit about its limitations: only ED frames and one view are modeled, the formulation is deterministic despite the one-to-many nature of image generation from masks, and no quantitative evaluation is reported. Historically, this work established the basic anatomy-conditioned formulation that later EchoGen-style systems generalized from single frames to clinically relevant video synthesis.

3. Expansion to controllable video generation

By 2024–2025, EchoGen-style work had shifted decisively from single-frame synthesis toward video generation, broader conditioning, and explicit clinical tasks.

System	Conditioning	Core mechanism
GAN-enhanced conditional echocardiogram generation	Segmentation masks	Conditional GAN with patch-based discriminator
ControlEchoSynth	Real A4C video + motion mask	Controlled video diffusion for A2C synthesis
ECHOPulse	ECG signal, optionally prior image	Tokenized video generation with masked visual token modeling
EchoFlow	Segmentation mask + view class for images; anatomy frame + EF for videos	Latent flow matching with privacy filtering
Echo-Path	ASD/Non-ASD or PAH/Non-PAH labels	Pathology-conditioned latent diffusion

ControlEchoSynth addresses a concrete acquisition bottleneck in biplane EF estimation: A2C is harder to obtain than A4C, especially in point-of-care settings. Its generator synthesizes A2C echo videos conditioned on real A4C videos, using the A4C clip together with a motion mask derived by frame differencing and Gaussian smoothing. The architecture is trained in two stages: an unconditional A2C video diffusion prior is first learned, and a ControlNet-style branch is then added for conditional generation. The downstream purpose is explicit: synthetic A2C videos are added to EF-regression training (Kondori et al., 25 Aug 2025).

ECHOPulse replaces expert-curated prompts with ECG time-series signals. The paper states that this is the first work to use a time-series prompt like ECG for ECHO video generation. ECG is treated as a natural, temporally aligned condition, available in routine clinical practice and even from wearables. The model can also be co-conditioned on a prior image plus ECG to update cardiac function information such as left ventricular ejection fraction (LVEF/EF) (Li et al., 2024).

EchoFlow recasts synthetic echocardiography as a foundation-model pipeline with four components: an adversarial variational autoencoder for the latent space, a latent image flow matching model for generating anatomies, a latent re-identification model for privacy filtering, and a latent video flow matching model for animating anatomies into videos conditioned on ejection fraction. The pipeline explicitly separates anatomy generation, privacy control, and functional video synthesis (Reynaud et al., 28 Mar 2025).

Echo-Path extends conditioning from anatomy or function to disease labels. It is a pathology-conditioned echo video generator focused on atrial septal defect (ASD) and pulmonary arterial hypertension (PAH). The system uses a latent image diffusion model to establish pathology-consistent anatomy, a privacy-preserving re-identification filter, and a latent video diffusion model to generate 64-frame disease-specific motion sequences, with autoregressive chaining for longer videos (Muhammad et al., 21 Sep 2025).

4. Architectural patterns and training formulations

Despite their diversity, EchoGen-style systems share a small set of recurring architectural ideas.

The first is paired conditional translation. In the 2019 GAN formulation, the input is a segmentation mask and the output is a single echo frame. Training combines a least-squares adversarial objective with pixel-wise MAE, and the patch-based discriminator enforces local texture realism rather than only whole-image realism. This is the simplest EchoGen pattern: deterministic, paired, and anatomically conditioned (Abdi et al., 2019).

The second is two-stage video diffusion. ControlEchoSynth first learns an unconditional prior over A2C videos, then attaches a ControlNet-style conditional branch so that A4C-derived structure and motion can steer generation. The paper reports that freezing the U-Net in the second stage slightly hurts quality, and that jointly fine-tuning the U-Net and control branch is best. This suggests that conditioning is not merely an add-on; the entire denoiser must adapt to the echo domain (Kondori et al., 25 Aug 2025).

The third is discrete-token generation. ECHOPulse tokenizes echocardiogram videos with VQ-VAE / LFQ-style discrete tokens, patchifies ECG signals, and uses a bidirectional transformer trained with masked token prediction in the style of MaskGIT. The practical consequence is parallel decoding, avoiding the slow inference associated with diffusion. The model also supports progressive extrapolation, where the last generated frames are re-encoded and generation continues, enabling length-unconstrained synthesis (Li et al., 2024).

The fourth is latent generative modeling with explicit privacy control. EchoFlow learns a cardiology-specific 2D latent space using an A-VAE, normalizes latent channels with training-set statistics, and then performs both image and video synthesis in latent space via flow matching. The privacy module is trained in latent space using positive pairs from the same video and negative pairs from different videos, and generated anatomies are rejected when their maximum Pearson correlation with training anatomies exceeds a threshold set from training-vs-validation similarities (Reynaud et al., 28 Mar 2025).

The fifth is class-conditioned latent diffusion. Echo-Path injects pathology labels into the latent image diffusion model through cross-attention, uses $v$-prediction, and enables classifier-free guidance by randomly dropping class conditioning during training. In the video model, the generated first-frame latent is concatenated with each frame so that the whole sequence remains anchored to the same disease-specific anatomy (Muhammad et al., 21 Sep 2025).

Across these systems, conditioning variables move from segmentation masks to views, motion masks, ECG traces, ejection fraction, and pathology labels. A plausible implication is that EchoGen has progressively shifted from static structural priors toward richer combinations of anatomy, motion, physiology, and disease state.

5. Evaluation and downstream clinical utility

Evaluation practice has changed substantially over time. The 2019 mask-to-frame GAN study is explicitly qualitative and reports no quantitative evaluation (Abdi et al., 2019). Later work uses perceptual and video-distribution metrics such as FID, FVD, IS, SSIM, and LPIPS, but also tests whether synthetic data improves a clinically relevant downstream model.

For ControlEchoSynth, the reported video-synthesis results are FVD 69.58, FID 26.64, SSIM 0.57, and LPIPS 0.16. Its stronger claim is downstream: for ResNet2+1D, performance improves from $R^2 = 0.623$, MAE = 4.58, RMSE = 6.25 with A4C only to $R^2 = 0.713$, MAE = 3.85, RMSE = 5.46 with A4C + synthetic A2C. For EchoCoTr-S, the corresponding change is from $R^2 = 0.231$, MAE = 6.90, RMSE = 8.93 to $R^2 = 0.642$, MAE = 4.75, RMSE = 6.09 (Kondori et al., 25 Aug 2025).

For ECHOPulse, the strongest reported setting is ECG-conditioned generation on a large private dataset. Reported results are A2C: FID 15.50, FVD 82.44, SSIM 0.67 and A4C: FID 20.82, FVD 107.40, SSIM 0.66. For EF consistency, the paper reports $R^2 = 0.85$, MAE = 2.51, RMSE = 2.86), together with sampling time = 6.4 s for 64 frames and 279M parameters. The qualitative analysis links ECG phase to generated motion, noting that the R wave corresponds to end-diastole (ED) and the T wave corresponds to end-systole (ES) (Li et al., 2024).

For Echo-Path, generative fidelity is reported with FID, FVD$_{16}$, and IS. On ASD, Ours, $w=5$ gives FID 18.35, FVD$_{16}$ 127.49, and IS $v$0; on PAH, FID 22.23, FVD$v$1 143.42, and IS $v$2. More importantly, when synthetic data is added to real training data, ASD diagnosis improves from ACC 0.848, AUC 0.846 to ACC 0.918, AUC 0.905, and PAH diagnosis improves from ACC 0.783, AUC 0.729 to ACC 0.863, AUC 0.830 (Muhammad et al., 21 Sep 2025).

For EchoFlow, the headline result is that models trained exclusively on EchoFlow-generated synthetic datasets achieve performance parity with models trained on real datasets for EF regression. The paper reports examples in which Dynamic A4C has real $v$3 and EchoFlow PC also 0.81, Pediatric A4C has real $v$4 and EchoFlow PC 0.72, Pediatric PSAX has real $v$5 and EchoFlow PC 0.72, and LVH PLAX has real $v$6 and EchoFlow PC 0.55 (Reynaud et al., 28 Mar 2025).

A common misconception is that echocardiogram generation is evaluated only by appearance. Later EchoGen-style work explicitly rejects that narrow criterion by measuring EF regression, diagnosis augmentation, or synthetic-only generalization.

6. Limitations, privacy, and adjacent multimodal directions

The literature is explicit about unresolved limitations. The 2019 mask-conditioned GAN is a preliminary feasibility study with no quantitative evaluation, restricted to ED frames and the apical four-chamber view (Abdi et al., 2019). ControlEchoSynth states that it does not aim to produce a perfect A2C solely from A4C, and notes that control could be improved using additional views, patient metadata, and other controllable factors such as brightness or depth (Kondori et al., 25 Aug 2025). Echo-Path reports that the domain gap persists in fully synthetic validation and suggests that PAH remains more difficult because of greater complexity and heterogeneity (Muhammad et al., 21 Sep 2025). ECHOPulse presents broad generalization claims to other modalities, but the experimental validation is on echocardiography (Li et al., 2024).

Privacy becomes a first-class concern in EchoFlow. That paper argues that generation alone does not guarantee privacy because models can memorize training anatomies, and therefore inserts a latent re-identification filter at inference time. The threshold is chosen as the 95th percentile of maximum training-vs-validation similarity, and the rejection rate is reported as around one-third of generated samples. This is a pragmatic privacy formulation: the released synthetic dataset, rather than every internal model component, is the primary object of protection (Reynaud et al., 28 Mar 2025).

Adjacent work has also moved in the reverse direction, using echocardiography to improve ECG representations rather than generating echo directly. EchoingECG is a probabilistic student-teacher model that combines PCME++ with a frozen ECHO-CLIP teacher to improve ECG-based prediction of ECHO-derived findings, and Echo2ECG aligns ECGs to multi-view Echo studies rather than single views, using a CLIP-style multimodal contrastive framework and a frozen EchoPrime encoder (Gao et al., 30 Sep 2025, Liman et al., 9 Mar 2026). These systems are not Echo generators, but they indicate a broader convergence between echocardiographic generation, cross-modal supervision, and morphology-aware cardiac representation learning. This suggests that future EchoGen research may be increasingly entangled with multimodal cardiac foundation models.

7. Terminological ambiguity and the unrelated namesake

The term EchoGen is not unique to cardiac imaging. “EchoGen: Cycle-Consistent Learning for Unified Layout-Image Generation and Understanding” (Zou et al., 18 Mar 2026) is an unrelated vision-LLM for layout-to-image generation and image grounding. It uses a pretrained multimodal backbone, specifically Janus-Pro 1.5B, and introduces a three-stage progressive pipeline: Parallel Multi-Task Pre-training (PMTP), Dual Joint Optimization (DJO), and Cycle RL.

In that work, generation and grounding are treated as dual tasks linked by a layout $v$7 image $v$8 layout cycle. DJO introduces a loop loss over sampled generated images, while Cycle RL replaces explicit visual supervision with a reward based on grounding-box discrepancy and optimizes the model with GRPO (Group Relative Policy Optimization). The reported results are state of the art on both MS-COCO layout-to-image generation and Ref-L4 grounding, including 54.61 AP / 68.85 AP$v$9 / 65.01 AP$R^2 = 0.623$0 / 25.18 CLIP / 20.12 FID for generation and 83.20 / 75.42 / 47.32 / 68.46 on the grounding metrics $R^2 = 0.623$1, $R^2 = 0.623$2, $R^2 = 0.623$3, and mAcc (Zou et al., 18 Mar 2026).

For encyclopedia use, the ambiguity matters. In cardiac-imaging contexts, EchoGen refers to echocardiogram generation and its extensions; in general computer vision, EchoGen can refer to a cycle-consistent unified model for layout-image generation and grounding. The two uses share only the name.