To ensure long-term sustainability (keep complexity in check) we need to keep auto3dseg_segresnet_inference.py inference script model-independent.
Instead of if 'whole-head' in model_file, we should have an option called something like KeepLargestConnectedComponent in the model. It should be also nicely documented (it could be just 1-2 sentences as comments at the top of auto3dseg_segresnet_inference.py file) to make it clear for model developers that they are allowed to use it.
I did not emphasize this with BRATS, as I thought that it was an exception - that nobody should follow that example, so no need to generalize. But at least we should document it in auto3dseg_segresnet_inference.py file. If we think that this kind of multichannel input is acceptable (not just a bad workaround) then we should add an option that describes the concept (MultiChannelInput?) and use that option name instead of referring to a specific model name.
To ensure long-term sustainability (keep complexity in check) we need to keep auto3dseg_segresnet_inference.py inference script model-independent.
Instead of
if 'whole-head' in model_file, we should have an option called something likeKeepLargestConnectedComponentin the model. It should be also nicely documented (it could be just 1-2 sentences as comments at the top of auto3dseg_segresnet_inference.py file) to make it clear for model developers that they are allowed to use it.I did not emphasize this with BRATS, as I thought that it was an exception - that nobody should follow that example, so no need to generalize. But at least we should document it in auto3dseg_segresnet_inference.py file. If we think that this kind of multichannel input is acceptable (not just a bad workaround) then we should add an option that describes the concept (MultiChannelInput?) and use that option name instead of referring to a specific model name.