ECHOPulse

This repository contains the code for the paper ECHOPulse: ECG Controlled Echocardiograms Video Generation. Aiming to generate the ECHO video based on the ECG signal. Model weights will be released after acceptance.

🔍 Abstract

Echocardiography (ECHO) is essential for cardiac assessments, but its video quality and interpretation heavily relies on manual expertise, leading to inconsistent results from clinical and portable devices. ECHO video generation offers a solution by improving automated monitoring through synthetic data and generating high-quality videos from routine health data. However, existing models often face high computational costs, slow inference, and rely on complex conditional prompts that require expert annotations. To address these challenges, we propose ECHOPulse, an ECG-conditioned ECHO video generation model. ECHOPulse introduces two key advancements: (1) it accelerates ECHO video generation by leveraging VQ-VAE tokenization and masked visual token modeling for fast decoding, and (2) it conditions on readily accessible ECG signals, which are highly coherent with ECHO videos, bypassing complex conditional prompts. To the best of our knowledge, this is the first work to use time-series prompts like ECG signals for ECHO video generation. ECHOPulse not only enables controllable synthetic ECHO data generation but also provides updated cardiac function information for disease monitoring and prediction beyond ECG alone. Evaluations on three public and private datasets demonstrate state-of-the-art performance in ECHO video generation across both qualitative and quantitative measures. Additionally, ECHOPulse can be easily generalized to other modality generation tasks, such as cardiac MRI, fMRI, and 3D CT generation. We will make the synthetic ECHO dataset, along with the code and model, publicly available upon acceptance.

🎬 Visualization

We visualize each set of ECG-echodiagrams and compared them in the first row (Prediction) and the second row (Ground Truth).

Type	Set 1	Set 2	Set 3	Set 4	Set 5	Set 6
Prediction
Ground Truth

✨ We also introduce our powerful ECHOPulse to general IoT devices (i.e., ⌚ Apple Watch in this case) and obtain the following visualization.

ECG Signal from Apple Watch v9	Generated Echodiagram

📊 Results on Reconstruction Metrics

Methods	Dataset	MSE	MAE
EchoNet-Synthetic	Echo-Dynamic	$3.00 \times 10^{-3}$	$3.00 \times 10^{-2}$
ECHOPulse (Only natural videos)	Echo-Dynamic	$5.02 \times 10^{-3}$	$4.08 \times 10^{-2}$
🏆 ECHOPulse (Domain transfer)	Echo-Dynamic	$2.89 \times 10^{-3}$	$2.92 \times 10^{-2}$
ECHOPulse (Only natural videos)	Private data	$6.09 \times 10^{-3}$	$4.41 \times 10^{-2}$
🏆 ECHOPulse (Domain transfer)	Private data	$2.91 \times 10^{-3}$	$2.97 \times 10^{-2}$

📊 Results on Clinical and Time-inference Metrics

Methods	Condition	S. time ⬇️	R² ⬆️	MAE ⬇️	RMSE ⬇️	Parameters
EchoDiffusion	Image+EF	Gen. 146s	0.89	4.81	6.69	381M
ECHOPulse	Image+ECG	Gen. 6.4s	0.85	2.51	2.86	279M

📊 Results on Image Quality Metrics

Methods	Dataset	Condition	FID↓	A2C FVD↓	A2C SSIM↑	A4C FID↓	A4C FVD↓	A4C SSIM↑
MoonShot	CAMUS	Text	48.44	202.41	0.63	61.57	290.08	0.62
VideoComposer	CAMUS	Text	37.68	164.96	0.60	35.04	180.32	0.61
HeartBeat	CAMUS	Text	107.66	305.12	0.53	76.46	381.28	0.53
ECHOPLuse (Only natural videos)	CAMUS	Text	12.71	273.15	0.61	15.38	336.04	0.58
ECHOPLuse (Domain transfer)	CAMUS	Text	5.65	211.85	0.79	8.17	283.32	0.75
EchoDiffusion-4SCM	Echo-Dynamic	Text&LVEF	24.00	228.00	0.48	24.00	228.00	0.48
ECHOPLuse (Only natural videos)	Echo-Dynamic	Text	36.10	319.25	0.39	44.21	334.95	0.35
ECHOPLuse (Domain transfer)	Echo-Dynamic	Text	27.50	249.46	0.46	29.83	312.31	0.41
ECHOPLuse (Only natural videos)	Private data	Text	27.49	291.67	0.53	34.13	374.92	0.51
ECHOPLuse (Domain transfer)	Private data	Text	25.44	224.90	0.54	31.21	334.09	0.54
🥈 ECHOPLuse (Only natural videos)	Private data	ECG	18.73	200.45	0.56	27.37	302.89	0.55
🥇 ECHOPLuse (Domain transfer)	Private data	ECG	15.50	82.44	0.67	20.82	107.40	0.66
HeartBeat	CAMUS	Text&Sketch&Mask...	25.23	97.28	0.66	31.99	159.36	0.65
EchoNet-Synthetic (Video Editing)	Echo-Dynamic	Whole Video&LVEF	16.90	87.40	/	16.90	87.40	/

🛠️ Requirements

conda create -n ECHOPulse python==3.8
conda activate ECHOPulse
pip install -r requirements.txt

⚙️ Train & Test

Training video tokenization model

python step1_train.py

Training video generation model

python step1_train.py

Inference

echo_inference.ipynb

Citation

If you use the code, please cite the following paper:

@article{li2024echopulse,
  title={ECHOPulse: ECG controlled echocardio-grams video generation},
  author={Li, Yiwei and Kim, Sekeun and Wu, Zihao and Jiang, Hanqi and Pan, Yi and Jin, Pengfei and Song, Sifan and Shi, Yucheng and Yang, Tianze and Liu, Tianming and others},
  journal={arXiv preprint arXiv:2410.03143},
  year={2024}
}