Hello organizers,
Are there any clinical labels for phase 2 test dataset, similar to those from phase 1? It would be very helpful, considering that one of your recommendations was to take the multi-modal approach.
Posted by: JChojnacki @ Aug. 29, 2023, 12:13 p.m.The data for the phase 2 test set is derived from real world settings. The training data and test data for phase 1 were derived from the OLIVES dataset that sourced its images and labels from clinical trials.
The general settings of the new testset ensures that generalizability and personalizability is enhanced. In the real world, it is not always clear if the same clinical data will match up across various institutions which makes relying on clinical data for the testing phase sub-optimal. Since phase 2 is testing a much more realistic setting with the newly introduced test set, it doesn't have as wide of a variety of clinical labels. Our intention of using clinical labels during training was to use it as a self supervised pre trained strategy described in:
Kokilepersaud, K., Corona, S. T., Prabhushankar, M., AlRegib, G., & Wykoff, C. (2023). Clinically Labeled Contrastive Learning for OCT Biomarker Classification. IEEE Journal of Biomedical and Health Informatics.
If you wish to use clinical labels during testing for the competition, the one that is available with the test set is the patient identity. Every file has the format DME-patientid-number.jpeg. However, keep in mind about how inference using the clinical data is less generalizable from a real world practice perspective.
Posted by: OLIVES @ Aug. 29, 2023, 6:32 p.m.