network topology rule

[Joeycho]

Hi all,

I checked 'experiment_planner_baseline_3DUNet_v21.py' and 'common_utils.py' code to understand how network topology is chosen. Of course, I read nnUNet paper as well. In my conclusion is 'target_spacing' is a decisive factor in the end.

In the paper, patch size also affects network topology, but I think, since initial patch size is also from spacing, in the end, spacing of dataset is crucial for network topology unless GPU status is changed.

Am I right? I wanted to figure it out what are the factors which decide the network topology. (Spacing of dataset?, other factors? maybe size too) I have the same network topologies in all experiment setups. I changed the number of data, add more modalities... but the network topology looks the same. And I realized maybe it is because of the same spacing. I already resampled all images to have the same spacing beforehand. Ah, I also resampled all images to have the same size beforehand.

So, if the given dataset has images which have the same size and spacing, then the network topology will be the same?

nnUNet/nnunet/experiment_planning/experiment_planner_baseline_3DUNet_v21.py

Line 83 in 6d02b5a

def get_properties_for_stage(self, current_spacing, original_spacing, original_shape, num_cases,

nnUNet/nnunet/experiment_planning/experiment_planner_baseline_3DUNet.py

Line 284 in 6d02b5a

self.plans_per_stage.append(self.get_properties_for_stage(target_spacing_transposed, target_spacing_transposed,