By: huanmiaozhao on May 11, 2023, 5:37 a.m.
Hello organizer, we downloaded the data of the competition and found that the Dense label of some data is incomplete, a normal mandibular canal should not be broken. For example, P92, and of course some others. Looking forward to your reply.
By: llumetti on May 12, 2023, 7:26 a.m.
Hi huanmiaozhao and thank you for getting interested in the challenge! You are right, some dense labels are incomplete while a normal mandibular canal should not. This happened due to a lack of density difference in such areas of the jawbone, this can be due to CBCT acquisition noise or patient specific conditions. Sometimes (in a few cases) this makes it hard (extremely so) or impossible to provide the 3D annotation and this is why you'll find holes. All the sparse labels, as they are annotated with a different approach (unfortunately not suitable for the dense labels) are complete and do not have any broken parts. Moreover, in the split we provided by default in the dataset, the test set is a small set of 15 patients such that all their annotations are complete. We ensure that all GT samples in the test set will exactly have two connected components, one for each canal.
If you have any further questions or concerns, please let me know.
Regards, Luca
