Split by patient and not study?
Thank you for releasing this very important dataset!
I have a request if it's not already too late: we have been examining at least the PadChest portion of the data, and we noticed that many patients have studies appearing in both validation and train folds, and (inferring from which data is missing), also in the test split. Is it possible to update the splits so that patients don't appear in >1 split?
Hello,
Thanks for your suggestion! Unfortunately we did not decide to split per patient.
Also, all of the padchest data (109,792 reports) should be present in the train and validation split. None will be present in the test set.
While this version of the dataset will be maintained for the challenge, if you happen to sort out the samples so that its split per patient, I would gladly push the dataset on the hub and reference it in the paper /github for reference and future works :)
Best