Acquisition source confounds a widely used retinal fundus benchmark: audit, partitions and per-run predictions
Authors/Creators
- 1. Amity University, Noida, India
- 2. Cardiff Metropolitan University, Cardiff, United Kingdom
- 3. University of Technology Sydney, Sydney, Australia
-
4.
University of Kuala Lumpur
Description
Code, partition manifests and per-image predictions supporting the audit and leakage-controlled re-evaluation of the 4,217-image Eye Disease Retinal Images collection.
Contents: the provenance and integrity audit; the manifests defining every evaluation arm image by image (image-level, patient-grouped, size-matched, source-matched, the common test set and the six leave-one-source-out splits); the training and evaluation code; stored per-image class probabilities for every run; the shortcut probes; and the visual adjudication verdicts for all 335 cross-label near-duplicate candidates.
The image collection itself is not redistributed here. It is available at https://www.kaggle.com/datasets/python16/eye-disease-retinal-images and the manifests reference it by relative path.