Hospital bias

Hello,

I performed a umap projection of the 2,000 CT scans as shown below. Some clusters clearly appear and when looking at the CT scans, it seems that each cluster is associated with a different scanner. Was the train / test split of the 10,000 patients performed randomly, or did you separate patients from an hospital / scanner to evaluate the robustness of the proposed models ?

Hi Simon,

The full STOIC dataset consists of CT volumes from 20 French hospitals. The full cohort was divided randomly into the public train, private train, and test sets. We hope that training on data from a large number of hospitals already makes the trained algorithms somewhat robust to different scanners/hospitals.

Best regards, Luuk