---
license: mit
task_categories:
- video-classification
tags:
- medical
- ultrasound
- eye
- ocular
- classification
size_categories:
- 1K<n<10K
language:
- en
pretty_name: erdes
dataset_info:
  features:
  - name: file_name
    dtype: string
  - name: label
    dtype:
      class_label:
        names:
          '0': Non_RD
          '1': Macula_Intact
          '2': Macula_Detached
          '3': PVD
  - name: duration (s)
    dtype: float32
  - name: fps
    dtype: float32
  splits:
  - name: train
    num_bytes: 220137
    num_examples: 5381
  download_size: 538819
  dataset_size: 220137
configs:
- config_name: default
  data_files:
  - split: train
    path: train/train.csv
---
# ERDES: Eye Retinal Detachment Ultrasound Dataset  

## 📌 Introduction  
**ERDES** is a large-scale, publicly available dataset of **3D ocular ultrasound videos** for retinal and macular detachment classification. It was introduced in our paper **[ERDES: A Benchmark Video Dataset for Retinal Detachment and Macular Status Classification in Ocular Ultrasound](https://huggingface.co/papers/2508.04735)**. The corpus consists of **5,381 expertly annotated video clips** totaling to **5 hours and 10 minutes**, providing a valuable resource for medical AI research in ophthalmology.  

### Key Features:  
✅ **5381 labeled ultrasound video clips**  
✅ **Expert annotations for retinal detachment (RD) and macular status**  
✅ **Structured classification (Normal, RD, PVD, macula-detached/intact)**  
✅ **Preprocessed for privacy and consistency**  

---

## 🎯 Motivation  
Medical video datasets for AI are scarce despite their clinical importance. ERDES bridges this gap by offering:  
- A **standardized benchmark** for retinal detachment classification in ultrasound videos.  
- Support for **spatiotemporal analysis** (e.g., 3D CNNs).  
- Open access to accelerate research in **ocular diagnostics**.  

---

## 📊 Dataset Overview  

### **1. Data Structure**  
Videos are categorized into:  
- **Normal**  
- **Retinal Detachment (RD)**  
  - *Macula-detached*: 
    - Bilateral
    - Temporal detachment  
  - *Macula-intact*: 
    - Nasal
    - Temporal detachment  
- **Posterior Vitreous Detachment (PVD)**  

Folder structure reflects labels: 
<div align="center"> 
<img src="media/folder_structure.PNG" alt="Data Statistics" width="150" />
</div>

### **2. Annotations**  
Each clip is labeled by sonologists for:  
- Presence/absence of retinal detachment.  
- Macular involvement (detached/intact).  

<div align="center">
<img src="media/labeling_rd.png" alt="Data Statistics" width="500" />
</div>

### **3. Preprocessing**  
- **Privacy**: PHI removed using YOLO-based globe detection.  
- **Consistency**: Cropped to the ocular ROI.  
- **Format**: MP4 videos.  

<div align="center">
<img src="media/data_stats.png" alt="Data Statistics" width="500" />
</div>

---

## 📥 Download  
Access the dataset via the HuggingFace API:

```py
from datasets import load_dataset

dataset = load_dataset("pnavard/erdes")
```

---

## 🛠️ Code & Baselines  
Repository: we open source our baseline experimetns on our **[GitHub](https://github.com/OSUPCVLab/ERDES)** repo. Which includes:  

- Baseline 3D CNN and ViT models for classification.
- End-to-end diagnostic pipeline for macular detachment.   

---

## 📜 Citation  
If you use ERDES, cite:  
```bibtex
@inproceedings{navardocular,
  title={A Benchmark Dataset for Retinal Detachment Classification in Spatiotemporal Ocular Ultrasound},
  author={Navard, Pouyan and Ozkut, Yasemin and Adhikari, Srikar and Yilmaz, Alper},
  booktitle={Nature Scientific Data (Under Review)},
  year={2025}
}
```