Two ways to use EchoForge

Both run on the same underlying models, so a result in Studio can be reproduced in Python.

No coding needed Launching soon

EchoForge Studio

Select an echocardiographic video, choose an analysis, and get the result. Studio runs the models needed behind the scenes, so you pick the measurement, not the algorithm.

Example: A4C video → LV volumes and ejection fraction

Results let you inspect the steps behind them, such as ED/ES frames and LV contours.

Open EchoForge Studio

For researchers and developers

Python library

Download pretrained models, load them by name, and build them into your own research pipelines. Documentation, weights and evaluation for every model are on GitHub.

Example: run LV segmentation from Python

Python library on GitHub

Models in the library

Each card is a task. The model behind it is named above the title, which is the name to use in code and in citations.

predicted viewA4C

Model: EchoFine

Echo view classification

Identifies the echocardiographic view from a 2D image, and doubles as a feature extractor for other echo tasks.

Input
Echo image
Output
One of 47 views
Data
TTE47

94.1% accuracy on the TTE47 test split

EchoFine documentation

Model: EchoLVSNet

Left ventricular segmentation

Delineates the left ventricular cavity, the starting point for quantitative measures of cardiac structure and function.

Input
A4C echo frame
Output
LV mask
Tested
Clinical and public datasets

Dice 0.91 on EchoNet-Dynamic, 0.94 on CAMUS

EchoLVSNet documentation
EDEDES

Model: EchoPDNet

Cardiac phase detection

Finds end-diastolic and end-systolic frames and tracks cardiac phase across multibeat sequences.

Input
Echo video
Output
ED/ES frames, phase curve

About 1.5 s per clip, with a mean error under 1 frame against cardiologists on the PACS set

EchoPDNet documentation
EF = (EDV − ESV) / EDVper beat, from one cine

Pipeline: EchoPDNet + EchoLVSNet New

Ejection fraction estimation

Chains phase detection and LV segmentation with a disk-summation volume estimate to give EF from a single clip.

Input
One A4C cine, DICOM or video
Output
EF per beat, mean and median
Pipeline documentation

Also in the library: EchoSDNet (single-heart versus dual-heart frames), further EchoFine backbones, and eight TTE47 baseline classifiers for benchmarking. The full list is in the repository.

One interface for every model

Change the model name to switch task. Weights download when first needed and are cached locally for reuse in your own pipelines.

Installation guide All models Weights on Hugging Face 
# pick any registered model by name
from echoforge import load_model

model = load_model("EchoFine_classifier")

From single models to complete pipelines

EchoForge is built so that models compose. In Studio, choosing ejection fraction runs this chain for you. In Python, you can call the pipeline or run each model on its own. It is the first pipeline to ship.

  1. Echo cineDICOM or video, A4C view
  2. EchoPDNetED and ES frames for every beat
  3. EchoLVSNetLV mask on each ED and ES frame
  4. EFPer beat, with mean and median

View identification with EchoFine will sit in front of this chain as more of the pipeline is automated.

Models and data together

EchoForge models are developed and validated with datasets that are available to the community. This closes the loop: models improve with more benchmarks, and data becomes more useful with better models.

Data→Annotation→Model development→Benchmarking→Reuse
Explore THRIVE datasetsTTE47

Growing EchoForge

  • Image quality assessment, starting with an A4C rotation model, released with its paper
  • Anatomical landmark detection
  • Further quantitative measurements built on the existing pipeline

New models are added as THRIVE research projects mature and are validated.

Work with us

We want to hear from clinical, academic and industry groups who can evaluate models on new populations, benchmark new methods, contribute models or data, or take EchoForge into clinical and educational tools.

Contribute on GitHub

Collaboration enquiries: [email protected]

EchoForge, including Studio and the Python library, is provided for research and development. The models are released under a CC BY-NC-SA 4.0 licence. Results must not be used for clinical diagnosis or patient-management decisions unless the relevant model and application have received the necessary clinical validation and regulatory approval.