Six Retinal Fundus Image Collections for Disease Detection

Six distinct retinal fundus image datasets, including EyePACS, MESSIDOR, DeepDRID, IDRID, DDR, and APTOS 2019, are aggregated. The collection facilitates comparative analysis and model training across multiple sources. The datasets were compiled by Mian Hamza, though specific creation dates and original sources are not detailed.

Use Cases

• Classify diabetic retinopathy severity levels using images from the APTOS 2019 and EyePACS datasets.
• Train segmentation models to identify lesions like hemorrhages and exudates on images from the IDRID and DeepDRID collections.
• Benchmark model generalization across different imaging protocols using the combined MESSIDOR, DDR, and other dataset images.
• Develop multi-dataset pretraining strategies for fundus image analysis to mitigate source-specific bias.

Strengths

• Aggregates six established public datasets, enabling cross-dataset validation.
• Includes datasets like IDRID and DeepDRID which provide pixel-level lesion annotations.

Limitations

• Individual dataset sizes, row counts, and specific label distributions are not provided.
• Potential inconsistencies in image resolution, quality, and annotation standards across the six sources.
• Lack of metadata on patient demographics, geographic origin, and data collection timelines for the aggregated set.

Provenance

Source

Compilation of public datasets: EyePACS, MESSIDOR, DeepDRID, IDRID, DDR, APTOS 2019.

Collection Method

Aggregated from existing public challenges and repositories.