Some data questions

Hi, I have some questions about the data, would be curious to hear what other people think:

1. Images 106 and 107 are clearly of the same eye but have different DR labels: Do these look like they are the same person at different visits and the DR has progressed in mean time?
2. For the segmentation task, can we assume that any image in the training data that doesn't have an annotation is a negative, i.e. no pixel should be annotated in those images?
3. Does the testing data follow exactly the same distribution as the training data (i.e. was it created with a random split stratified on class labels) or might there be some distribution shift?

Cheers!

Hi,

Here are answers to your questions.

1. After confirming with the doctor, images 106 and 107 are both grade 0 in task 3 (DR Grading). You can correct that by yourself, and we will inform other participants.
2. Yes. For some images, no annotations (masks) mean there are no lesions in these images.
3. When dividing the test set, we consider the problem of sample imbalance for each class. So I guess for your answer is our testing data follows the same distribution as the training data.

Thanks.

DRAC2022 Organizers

Hi! Thanks a lot for the answers!

Can I just double check if I understood the answer for 2. correctly: For some images no annotations for task 1 means no lesions, but not for all? So there might be some images that do show lesions but don't have the annotations?

Thanks!

Hi, It is possible that some images have lesions but don't have the annotations, it depends on the doctor's experience and annotation quality. Our labels have been reviewed by doctors many times. And we will appreciate it if you found some mistakes in the annotations and give the feedback to us. We will also ask the doctors to review the labels again, and if there is an update, we will send it to you as soon as possible. Thanks.

DRAC2022 Organizers