Glossary

Echocardiogram

An ultrasound scan of the heart. The input to both scripts.

A4C (Apical 4-Chamber)

A standard echocardiographic view showing all four heart chambers. The view the EF model was trained on.

Left Ventricle (LV)

The heart’s main pumping chamber. Its cavity is what the segmentation model outlines.

Segmentation

Pixel-wise labelling of an image — here, marking which pixels belong to the LV cavity.

Mask

The binary image produced by segmentation: 1 inside the LV, 0 elsewhere.

LV Area

The number of mask pixels in a frame (reported at the 112×112 scale). Used as a proxy for cavity size over time.

Systole

The contraction phase of the heartbeat, when the LV is smallest. Detected as the troughs of the area curve.

Diastole

The filling phase, when the LV is largest.

Ejection Fraction (EF)

The percentage of blood ejected from the LV per beat. Normal is roughly 50–70 %.

DeepLabV3-ResNet50

The convolutional segmentation network used by PredictSegmentation.py.

R(2+1)D-18

The spatio-temporal video network used by PredictEF.py to regress EF.

Clip

A short sequence of sampled frames fed to the EF model. Many overlapping clips per video are averaged into one EF value.

Checkpoint (.pt)

A saved file of trained model weights, loaded at runtime (here r2plus1d_18_32_2_pretrained.pt and deeplabv3_resnet50_random.pt).

EchoNet-Dynamic

The Stanford project providing the models and dataset the scripts build on: https://echonet.github.io/dynamic/.