EyePACS — Diabetic Retinopathy Fundus Image Dataset

Image: EYEPACS Preprocess Samples.

---

📘 Overview

EyePACS (Eye Picture Archive Communication System) is a large-scale collection of retinal fundus images used for automated diabetic retinopathy (DR) detection.
It formed the basis of the Kaggle Diabetic Retinopathy Detection challenge, enabling research into DR classification and screening.

The dataset includes macula-centered color fundus images from real-world clinical screenings under varied imaging conditions.

---

📊 Dataset Summary

Field	Details
Task	Diabetic retinopathy classification (5 severity grades)
Description	High-resolution color fundus photographs from EyePACS DR screening program. Each image labeled by ophthalmologists using the ICDR scale.
Size	~88,702 images total (≈35k labeled for training, ≈53k unlabeled for testing)
Classes	0 = No DR, 1 = Mild, 2 = Moderate, 3 = Severe, 4 = Proliferative DR
Image Type	Macula-centered color fundus photos (640×480 to 5184×3456 px)
Source	EyePACS (USA) via Kaggle Diabetic Retinopathy Detection Challenge (2015)
Access	Kaggle Dataset
License	Usage restricted under Kaggle & EyePACS terms

---

🧱 Dataset Structure

eyepacs/
├── images/
│ ├── train/
│ │ ├── 00001_left.jpg
│ │ ├── 00001_right.jpg
│ └── test/
├── labels.csv # image_id, eye(L/R), dr_grade (0–4)
├── README.md
└── LICENSE.txt

---

🧩 Label Details

• Labels follow the International Clinical Diabetic Retinopathy (ICDR) grading system.
• Class imbalance is significant — most images show no DR.
• Labels were assigned by certified ophthalmologists; minor label noise may exist.

---

⚙️ Preprocessing Recommendations

• Crop to the circular fundus region and remove borders
• Resize to consistent resolution (e.g. 1024×1024)
• Normalize illumination and contrast
• Exclude blurred or ungradable images

---

💡 Research Applications

• DR detection and severity classification
• Automated retinal screening systems
• Transfer learning and robustness testing across imaging conditions
• Comparative studies with datasets like MESSIDOR, DDR, and APTOS

---

⚠️ Notes & Limitations

• Significant class imbalance (majority = No DR)
• Variations in camera type, exposure, and focus
• Some label noise and ungradable images present
• Redistribution may be restricted — verify Kaggle/EyePACS terms before publishing images

---

📄 Citation

If you use the dataset, cite:

Kaggle and EyePACS. “Diabetic Retinopathy Detection.” Kaggle Competition, 2015.
https://www.kaggle.com/c/diabetic-retinopathy-detection

---

🔗 References

• EyePACS Official Data Page — https://www.eyepacs.com/data-analysis
• Kaggle: Diabetic Retinopathy Detection — https://www.kaggle.com/c/diabetic-retinopathy-detection
• Research overview: Transfer Learning Based Classification of Diabetic Retinopathy on the Kaggle EyePACS Dataset, ResearchGate, 2021.

---