datasets:
- EchoDynamic
- RVENet
- EchoNet-Pediatric-LVH
language: en
library_name: pytorch
license: mit
tags:
- self-supervised-learning
- echocardiography
- medical-imaging
- video-representation
model_index: deep-learning
paper: https://arxiv.org/pdf/2506.11777
pipeline_tag: video-feature-extraction
🫀 DISCOVR — Self-Supervised Echocardiography Representations
Paper: Self-Supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation — NeurIPS 2025
📄 arXiv:2506.11777
Code: https://github.com/mdivyanshu97/DISCOVR
📦 Available Checkpoints
| Epochs | Filename | Description |
|---|---|---|
| 200 | checkpoint-199.pth |
Model trained for ~200 epochs |
| 300 | checkpoint-299.pth |
Model trained for ~300 epochs |
| 400 | checkpoint-399.pth |
Model trained for ~400 epochs |
| 600 | checkpoint-599.pth |
Model trained for ~600 epochs |
| 800 | checkpoint-799.pth |
Model trained for ~800 epochs |
Each checkpoint corresponds to a model trained for the indicated number of epochs on adult and pediatric echocardiography datasets (EchoDynamic, RVENet, EchoNet-Pediatric LVH).
🧠 Model Overview
DISCOVR is a self-supervised framework for learning spatio-temporal echocardiographic video representations via online cluster distillation.
It learns both fine-grained anatomical semantics and global temporal dynamics, supporting downstream tasks such as:
- Cardiac view classification
- Functional abnormality detection
- Video segmentation
- Representation learning for medical imaging
Not for clinical or diagnostic use.
Sample Usage
To pretrain the model on echocardiographic videos:
python -m torch.distributed.launch --nproc_per_node=NUM_GPUS \
scripts/run_mae_pretraining.py \
--data_path /path/to/echo_videos \
--data_path_csv /path/to/train.csv \
--data_path_val /path/to/val.csv \
--data_path_test /path/to/test.csv \
--mask_type multi_local \
--loss_func SIGMA \
--model pretrain_videomae_base_patch16_224 \
--batch_size 48 \
--num_frames 64 \
--opt adamw \
--opt_betas 0.9 0.95 \
--warmup_epochs 40 \
--epochs 400
🔖 Quick Facts
- Repo:
Div97/DISCOVR_ADULT_PEDIATRIC_MODEL - Model family: DISCOVR checkpoints (199 → 799)
- Architecture: ViT-Base backbone, 64-frame clips (stride 3)
- Datasets used: EchoDynamic, RVENet, EchoNet-Pediatric LVH
- Training objective: Self-supervised online cluster distillation
- Intended use: Research & education
- Not intended for: Clinical decision-making or real-world patient care
🧩 Citation
If you use DISCOVR in your work, please cite:
@article{mishra2025self,
title={Self-supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation},
author={Mishra, Divyanshu and Salehi, Mohammadreza and Saha, Pramit Saha and Patey, Olga and Papageorghiou, Aris T and Asano, Yuki M and Noble, J Alison},
journal={arXiv preprint arXiv:2506.11777},
year={2025}
}
License
This project is licensed under the MIT License.