3. Data Preparation

Both scripts read a folder of files and process everything they find. This chapter explains the accepted formats and how each script scans the folder.

3.1. Accepted input formats

Format

Extension

Segmentation

EF prediction

Video

.avi, .mp4

βœ…

βœ…

NumPy array (single frame)

.npy

βœ…

❌

Key difference

PredictSegmentation.py auto-detects both videos and .npy files in the same folder and handles each with a dedicated module. PredictEF.py only looks for .avi/.mp4 videos β€” EF is a temporal measurement and needs a full clip.

3.2. How each script scans the folder

3.2.1. Segmentation

all_files   = glob.glob(os.path.join(input_dir, "*"))
video_files = [f for f in all_files if f.lower().endswith(('.avi', '.mp4'))]
npy_files   = [f for f in all_files if f.lower().endswith('.npy')]

So a single INPUT_DIR may mix videos and .npy frames; each group is processed separately. (There are two modules in the code. Module A for .npy, Module B for videos.)

3.2.2. EF prediction

self.video_paths = glob.glob(os.path.join(video_dir, "*.avi")) + \
                   glob.glob(os.path.join(video_dir, "*.mp4"))

Only top-level .avi/.mp4 files are picked up. Sub-folders are not searched recursively β€” put your videos directly inside the input folder.

3.3. Image / video assumptions

The scripts were built around EchoNet-Dynamic data, so they expect:

  • Apical-4-chamber (A4C) echocardiogram views for meaningful EF.

  • Frames are internally resized to 112 Γ— 112 pixels, so the source resolution does not need to match β€” but very low quality inputs degrade accuracy.

  • Videos are read with OpenCV; for EF, each frame is converted to grayscale and back to RGB (the model expects 3 identical channels).

3.3.1. .npy frame handling (segmentation only)

Module A is deliberately tolerant of .npy layouts. It normalises any array to a (H, W, 3) uint8 image:

  • 2-D (H, W) grayscale β†’ stacked into 3 channels

  • channel-first (C, H, W) β†’ transposed to (H, W, C)

  • single-channel (H, W, 1) β†’ repeated to 3 channels

  • pixel values are min–max scaled to 0–255

Anything with an unsupported number of dimensions is skipped with a warning.

3.4. Normalisation constants

Every frame is scaled to [0, 1] and then normalised with a per-channel mean and standard deviation, defined in a single MEAN/STD block at the top of each script:

MEAN = [0.128, 0.129, 0.130]
STD  = [0.196, 0.196, 0.197]

These same statistics are used consistently across all paths:

Script / module

Mean (RGB)

Std (RGB)

Scale

Segmentation β€” .npy

MEAN

STD

0–1

Segmentation β€” video

MEAN

STD

0–1

EF β€” video

MEAN

STD

0–1

Recompute for your own data

The defaults come from the EchoNet-Dynamic training set. If your data looks different, recompute MEAN/STD with CalculateStats.py and paste the result into both scripts β€” see Normalization Statistics.

3.5. Project folder layout

If you keep this layout, the paths in the scripts will match your files as-is:

EchoNet-dynamic/
β”œβ”€β”€ scripts/                     # the three .py scripts
β”œβ”€β”€ stats/                       # model weights
β”‚   β”œβ”€β”€ deeplabv3_resnet50_random.pt
β”‚   └── r2plus1d_18_32_2_pretrained.pt
└── demo/                        # input data (replace with your own)
    β”œβ”€β”€ video/   *.avi / *.mp4
    └── npy/     *.npy            # only used by the segmentation script

To run on your own study, you can easily drop your videos into demo/video/ (and any .npy frames into demo/npy/), or just point INPUT_DIR at a folder of your choice.

With the data in place, continue to LV Segmentation or EF Prediction.