CSDI: A Fine-Grained Fundus Image Dataset of Cataract Severity and Diagnostic Images

The CSDI Cataract Diagnosis Dataset is a curated collection of 187 fundus images with expert annotations, including cataract severity scores and bilingual diagnostic descriptions (English and Chinese). This dataset supports research in automated cataract screening, grading, and fundus image interpretation using both image and text modalities. This dataset is licensed under CC BY-NC 4.0. See LICENSE for details.

📁 Dataset Overview

• Total images: 187
• Image formats: .png and .jpg
• Annotation file: CSDI_annotations.csv
• Labels: Cataract severity score (0–10), expert-written diagnoses, optic disc localization, optic disc clarity

📂 File Structure

CSDI/
├── original_images/ # Folder containing 187 fundus images (.png and .jpg)
│ ├── cataract_001.png
│ ├── cataract_002.jpg
│ └── ...
├── CSDI_annotations.csv # Annotation CSV file (UTF-8 encoded)
├── README.md                  # Dataset description
├── LICENSE.md                 # Dataset license (CC BY-NC 4.0)
├── crop_fundus_images.py      # Script for cropping fundus images
└── augment_fundus_images.py   # Script for image augmentation

🗒️ CSV Annotation File (CSDI_annotations.csv)

• The CSV annotation file is encoded in UTF-8 to ensure proper display of English and Chinese characters.

• Column Name	Description
  id	Image file name (e.g., cataract_001.png)
  score	Cataract severity score (range 0–10)
  English_diagnosis	Diagnosis in English
  Chinese_diagnosis	Diagnosis in Chinese
  original_image_width	Original image width in pixels
  original_image_height	Original image height in pixels
  fundus_region_x1 / y1 / x2 / y2	Coordinates of the annotated fundus region in pixels; -1 if not visible
  optic_disc_clear	Optic disc visibility (visible / not visible)
  optic_disc_x1 / y1 / x2 / y2	Coordinates of the optic disc bounding box in pixels; -1 if not visible

The score field supports both regression (exact score prediction) and classification (e.g., mild/moderate/severe).

📊 Suggested Severity Levels

The cataract severity grading criteria are defined as follows, with decimal scores (to one decimal place) allowing for precise assessment within each range:

Severity Level	Score Range	Quantity	Percentage (%)
Normal	[0, 1)	9	4.81
Acceptable	[1, 3)	30	16.04
Mild	[3, 5)	39	20.86
Moderate	[5, 7)	48	25.67
Severe	[7, 10]	61	32.62

• 0–1 (Normal): No cataract; fundus images are clear with no lens opacity affecting image quality.
• 1–3 (Acceptable): Acceptable cataract; images remain clear with subtle blurring, low likelihood of visual impairment.
• 3–5 (Mild): Mild cataract; image clarity decreases, regular visual monitoring recommended, surgery unlikely immediately necessary.
• 5–7 (Moderate): Moderate cataract; noticeable reduction in image clarity, higher probability of visual impairment, elective surgery may be considered.
• 7–10 (Severe): Severe cataract; marked reduction in image clarity, significant impact on visual function, prompt surgical intervention advised.

---

🔍 Diagnostic Content

Each record contains detailed diagnostic descriptions in both Chinese and English, following a fixed sentence structure evaluating:

• Overall Color: Typical orange-red fundus background; deviation toward yellow-white indicates increased cataract severity.
• Optic Disc and Vessel Clarity: Blurring of optic disc margins and reduced vessel clarity indicate lens opacity affecting image quality.
• Macular Area: Accurate localization and assessment of the macula is crucial; inability to do so suggests advanced cataract.
• Retinal Vessel Clarity and Branching: Severity increases as visualization decreases from fine branch vessels, to major vessels and second-order branches, to only major vessels visible, to complete inability to distinguish vascular structures.

This standardized scoring system and diagnostic protocol were strictly adhered to by annotators to ensure high-quality, reliable annotations.

---

English and Chinese versions of the diagnostic descriptions are included to support multilingual and cross-lingual research applications.

💻 Code Scripts

• crop_fundus_images.py: Crop fundus images to remove black borders and save cropping info in CSV.

  python crop_fundus_images.py -i csdi_datasets/original_images -o csdi_datasets/cropped_images -p 0 -c crop_info.csv
  

• augment_fundus_images.py: Apply random rotations, zoom-ins, and rotation+zoom augmentations to cropped images.

  python augment_fundus_images.py
  

🎯 Applications

• Automated cataract screening and grading
• Ophthalmic report generation (image-to-text)
• Fundus image quality analysis
• Cross-modal learning and medical vision-language modeling
• Optic disc detection and segmentation

👨‍🔬 Authors

Zixun Xie^1,2,*, Mingxin Ao^3,*, Haiming Tang^1,4,*, Xuemin Li³, Xiang Bai^1,2, Shanghang Zhang^1,†, Dawei Li^5,†

¹State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, China
²School of Software and Microelectronics, Peking University, China
³Department of Ophthalmology, Peking University Third Hospital, China
⁴School of Computing, National University of Singapore, Singapore
⁵School of Future Technology, Peking University, China

* Equal contribution and co-first authors. † Corresponding authors.

📬 Contact

• Zixun Xie: [email protected]
• Mingxin Ao: [email protected]
• Haiming Tang: [email protected]
• Xuemin Li: [email protected]
• Xiang Bai: [email protected]
• Shanghang Zhang: [email protected]
• Dawei Li: [email protected]

📜 Citation

@misc{csdi2025cataract,
  title        = {CSDI: A Fine-Grained Fundus Image Dataset of Cataract Severity and Diagnostic Images},
  author       = {Xie, Zixun and Ao, Mingxin and Tang, Haiming and Li, Xuemin and Bai, Xiang and Zhang, Shanghang and Li, Dawei},
  year         = {2025},
  note         = {Under review at Scientific Data}
}