no-contrast CTs get bad performance on KiTS model about kits21 HOT 1 OPEN

It's always difficult to say what features, exactly, the segmentation model is using the make its predictions, but your results seem to suggest that enhancement is an important feature. This doesn't surprise me too much since the kidneys enhance so conspicuously, especially in the arterial contrast phase.

Of course, in the context of the KiTS21 training set, non-contrast CT scans are "out of distribution" and so it's difficult to reason about expected performance on that population. But one thing that might help is to train on other contrast phases as well, if not non-contrast. The KiTS23 dataset includes cases in the nephrogenic phase, which might force the model to learn through some of the aterial-dependent features its using.

from kits21.