|
--- |
|
license: mit |
|
tags: |
|
- medical |
|
pretty_name: d |
|
--- |
|
# Paired Retina Dataset |
|
|
|
## Dataset Summary |
|
|
|
This dataset contains paired retinal images captured from the same patients using both tabletop and portable devices, designed to support research in cross-device retinal image analysis and diabetic retinopathy screening. |
|
|
|
## Dataset Details |
|
|
|
### Dataset Description |
|
|
|
**Curated by:** Research group developing RetSyn |
|
|
|
**Modality:** CFP |
|
|
|
|
|
The Paired Retina Dataset comprises retinal fundus photographs from 327 patients, with each patient contributing paired images - one captured using a tabletop device and another using a portable device, taken simultaneously under identical conditions. Each record contains matched images from both devices, enabling direct comparison of imaging characteristics between traditional tabletop cameras and portable smartphone-based devices. |
|
|
|
### Dataset Sources |
|
* **Paper:** RetSyn: Improving Synthetic Samples for Ophthalmology with DPO Using Paired Data |
|
|
|
## Uses |
|
|
|
### Direct Use |
|
|
|
The dataset is designed for: |
|
* Training and evaluating cross-device generalization of retinal image analysis models |
|
* Studying the differences between tabletop and portable device retinal imaging |
|
* Developing and validating domain adaptation techniques for retinal image analysis |
|
* Supporting research in making retinal diagnosis AI systems more robust across different imaging devices |
|
|
|
## Dataset Structure |
|
|
|
The dataset contains: |
|
|
|
**Overview:** |
|
* Total number of patients: 327 |
|
|
|
**Record Structure:** |
|
* Each record contains: |
|
* One image captured using a tabletop device |
|
* One image captured using a portable device |
|
* DR status labels |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
The dataset was created to address the challenge of developing robust retinal diagnosis AI systems that can perform consistently across different imaging devices, particularly between traditional tabletop cameras and portable devices. It specifically aims to support research in improving AI model performance on portable device images, which are crucial for expanding screening access in resource-limited settings. |
|
|
|
### Source Data |
|
|
|
#### Data Collection and Processing |
|
|
|
* Images were collected from patients with both devices simultaneously under identical conditions |
|
* Images were labeled with DR status (normal, non-proliferative DR, or proliferative DR) |
|
* Quality labels were provided for the images |
|
|
|
#### Source Data Producers |
|
|
|
The data was collected by a research group during clinical examinations. Images were captured using: |
|
* **Tabletop devices:** Standard retinal cameras |
|
* **Portable devices:** Samsung Galaxy S10 smartphones running Android 11 with integrated handheld retinal cameras |
|
|
|
### Personal and Sensitive Information |
|
|
|
The dataset contains medical images that have been anonymized to protect patient privacy. All personal identifiers have been removed from the images and associated metadata. |
|
|
|
## Additional Information |
|
|
|
For additional details about the dataset and its applications, please refer to the RetSyn paper which introduces this dataset as part of a larger study on improving synthetic data generation for ophthalmology. |
|
|
|
## Dataset Card Contact |
|
|
|
Contact information can be found in the associated paper. |
