Unity Imaging Cardiac Segmentation

This repository holds the training and inference code for our Unity Imaging cardiac keypoint and segmentation networks.

These networks are used to derive

1. Cardiac keypoints used for measuremet (such as the internal LV diameter) or for tracking.
2. Boundries (such as LV endo and epicardium)

Data

The training, test, and validation data consists of anonymised echocardiographic images and associated label data. The labels are supplied as a JSON file, with a set of labeles for each file. The labels can either be a single point, or a series of points that define the control points of a cubic spline. For each label for each file, it may be

1. Present as a point
2. Present as a curve (cubic spline)
3. Absent because either the point/curve is not within the image, or would not be expected to be within the image.
4. Absent because the point/curve is not clear within the image

The loss for (3) and (4) handeled differently.

The labels for the data comes from our Expert Cardiologists and Sonographers using the Unity Imaging web app.

The data required to train and validate each model will be released with each paper.

The code and trained models are released before publication to allow review during peer-review.

Dependencies

Outside of the standard anaconda distribution the only additional package needed is pytorch

Training

Run train-unity-pytorch-ddp.py There are some example helper files for e.g. running as multiple systemd jobs (one per GPU), or for running as a PBS job on a university cluster.

Inference

Run predictions-unity-pytorch.py

Contact

Dr. Matthew Shun-Shin, Imperial College London