EchoNet-Dynamic

# EchoNet-Dynamic Official EchoNet-Dynamic A4C echocardiography videos and labels, plus EchoFlow `avae-4f4` latent archives used by this project. Official website: https://echonet.github.io/dynamic/ ## Contents | Path | Description | |---|---| | `FileList.csv` | 10030 cases with EF/ESV/EDV and TRAIN/VAL/TEST split | | `VolumeTracings.csv` | ED/ES contour tracings | | `Videos.tar.gz` | All 10030 source `.avi` videos | | `EchoFlow_20260605_latent.tar.gz` | EchoFlow `avae-4f4` latents (2026-06-05 encoding) | | `EchoFlow_20260615_echosyn_avae-4f4.tar.gz` | EchoFlow `avae-4f4` latents (2026-06-15 `echoflow-standard` encoding) | | `EchoVDiff_CCs_13f/` | EchoVDiff-style 13-frame cardiac-cycle release (videos, EF/phase labels, latents, val real_reference) | Split sizes from `FileList.csv`: | Split | Cases | |---|---:| | TRAIN | 7465 | | VAL | 1288 | | TEST | 1277 | | Total | 10030 | ## Video archive layout After extracting `Videos.tar.gz`, videos live under `Videos/` in a mixed layout that preserves the local working copy: - `Videos/*.avi` — 2410 files at the top level - `Videos/batch_01/` — 3000 files - `Videos/batch_02/` — 3000 files - `Videos/batch_03/` — 1620 files Together these directories contain exactly the 10030 filenames listed in `FileList.csv`, with no duplicates. ```bash tar -xzf Videos.tar.gz find Videos -name '*.avi' | wc -l # expect 10030 ``` ## EchoFlow AVAE latent archives Both latent archives are produced with EchoFlow VAE `avae-4f4`. Latent tensors are `torch.Tensor` with shape `[T, 4, 28, 28]` (`float32`). The two archives cover the same 10030 cases but come from different encoding runs and are **not** numerically identical. - Prefer `EchoFlow_20260615_echosyn_avae-4f4.tar.gz` for new EchoFlow-standard experiments (`encoding_mode: echoflow-standard`). - Keep `EchoFlow_20260605_latent.tar.gz` for reproducing earlier 2026-06-05 results. ```bash tar -xzf EchoFlow_20260615_echosyn_avae-4f4.tar.gz tar -xzf EchoFlow_20260605_latent.tar.gz ``` ## Download Pin a full 40-hex ModelScope dataset revision when deploying: ```bash modelscope download \ --dataset LuminousXHJ/EchoNet-Dynamic \ --revision <40-hex-commit> \ --local_dir /path/to/EchoNet-Dynamic ``` ## EchoVDiff cardiac-cycle release `EchoVDiff_CCs_13f/` packages **30,252** accepted ED→ES→ED cycles as 13-frame clips with case-level EF/EDV/ESV, phase labels, EchoFlow `[13,4,28,28]` latents, and a val `real_reference/` cache. See `EchoVDiff_CCs_13f/README.md`. ## Not included EchoNet-Synthetic reproduction artifacts, working MedSAM2 staging/masks, and local caches are intentionally omitted from this dataset repository. The raw working tree `EchoVDiff_CCs/` (variable clips / masks) is not published; use `EchoVDiff_CCs_13f/` instead.