Dataset

The image data was provided by EyePACS LLC, Santa Cruz, CA, USA. Labels were provided by the Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, The Netherlands.

The JustRAIGS dataset is divided into a training subset with 101,442 gradable fundus images, spanning both referable and no referable glaucomatous cases, and a test subset comprising 9,741 fundus images.  For more information about the dataset please see this paper.

The training set, named JustRAIGS Dataset, is publicly available under a CC BY-NC-SA license at the Zenodo here.

The test set will not be publicly released and will be kept private even after the end of the challenge.

Labels of Dataset

• No referable glaucoma (NRG)
• Referable glaucoma (RG)
  

Additional features for referable glaucoma:

• Appearance neuroretinal rim superiorly (ANRS)
  
• Appearance neuroretinal rim inferiorly (ANRI)
• Retinal nerve fiber layer defect superiorly (RNFLDS)
• Retinal nerve fiber layer defect inferiorly (RNFLDI)
• Baring circumlinear vessel superiorly (BCLVS)
• Baring circumlinear vessel inferiorly (BCLVI)
• Nasalisation of vessel trunk (NVT)
• Disc hemorrhages (DH)
• Laminar dots (LD)
• Large cup (LC)

How the annotation of the dataset has been done:

Each image was annotated by 2 graders (G1 and G2), randomly selected from a pool of qualified graders. If they agreed on their main classification (referable glaucoma (RG), no referable glaucoma (NRG)), this classification was considered as final; in case of disagreement, however, the image was subsequently graded by grader 3 which is a glaucoma specialist (G3); the classification was considered as final. 

When graders selected "referable glaucoma" as their main classification, they were asked to check boxes corresponding to the reasons why they felt the patient should be referred. There were 10 options available (additional features for referable glaucoma); graders could select as many as they thought to be applicable to the image shown. 

Disagreement on the additional labels was not resolved. Instead, the ground truth of the additional labels is based on both the initial graders (if they agreed on the main classification); or on one grader and the glaucoma specialist (if one of the graders agreed on the main classification provided by the specialist); or on the glaucoma specialist (if none of the graders agreed on the main classification provided by the specialist). If for any of the additional labels there was disagreement, that label will be ignored for that image during the evaluation of the algorithms.