CAMUS

# CAMUS Official CAMUS 2D echocardiography dataset (NIfTI) plus the EchoLVFM reproduction assets used by this project: the full MP4/mask export and the `20260612` AVAE latent set. Paper: Leclerc et al., "Deep Learning for Segmentation using an Open Large-Scale Dataset in 2D Echocardiography", IEEE TMI, 2019. ## Contents | Path | Description | |---|---| | `database_nifti.tar.gz` | Official 500-patient NIfTI release (2CH + 4CH) | | `database_split/` | Official TRAIN/VAL/TEST patient lists (400/50/50) | | `jupyter/` | Official EF notebook and license terms | | `EchoLVFM_video_dataset_20260612_full.tar.gz` | 1000 MP4 sequences with ED/ES/frame masks | | `EchoLVFM_20260612_latent.tar.gz` | EchoFlow `avae-4f4` latents used by EchoLVFM reproduction runs | | `README.md` | This file | ## Official split | Split | Patients | |---|---:| | TRAIN | 400 | | VAL | 50 | | TEST | 50 | | Total | 500 | Each patient has both `2CH` and `4CH` views (1000 sequences after export). ## Extract archives ```bash tar -xzf database_nifti.tar.gz tar -xzf EchoLVFM_video_dataset_20260612_full.tar.gz tar -xzf EchoLVFM_20260612_latent.tar.gz ``` Expected counts after extraction: ```bash ls -d database_nifti/patient* | wc -l # 500 ls -d EchoLVFM_video_dataset_20260612_full/patient*_2CH \ EchoLVFM_video_dataset_20260612_full/patient*_4CH | wc -l # 1000 ls EchoLVFM_20260612_latent/patient*.pt | wc -l # 1000 ``` ## EchoLVFM_20260612_latent - VAE: EchoFlow `avae-4f4` - Source videos: `EchoLVFM_video_dataset_20260612_full` - Per-case file: `patientXXXX_{2CH,4CH}.pt` containing `mu` / `var` - Latent shape: `[T, 4, 28, 28]`, `float32` - Also includes `CAMUS_Scaling_4f4.pt` and `metadata.csv` This is the latent directory used by later EchoLVFM CAMUS reproduction experiments (for example `camus_video_quality_20260625_direct_gray` and subsequent 20260703/20260704 reruns). ## Not included Historical latent encodings (`EchoLVFM_20260605_latent`, `EchoLVFM_20260608_latent`) and debug video subsets (`sample10`, `subset10_fast`) are intentionally omitted. ## Download Pin a full 40-hex ModelScope dataset revision when deploying: ```bash modelscope download \ --dataset LuminousXHJ/CAMUS \ --revision <40-hex-commit> \ --local_dir /path/to/CAMUS ```