Model variation for brain tumor segmentation

[nshid0]

In the original nnUNet paper, brain tumor segmentation can benefit from using three convolutions per stage (Figure 6.) It does not say how much better the Dice score improves, but it looks like the nnUNet model with three convolutions beats the original model (with two convolutions) most of the cases.


I tried to train my network both using two convolutions and three convolutions while keeping every other parameters the same, but I was not able to see a meaningful improvement when using three convolutions. I tried to increase the number of epochs to 1500 but it didn't help me either.
I am just wondering how I can replicate the paper's finding. Any advice will be appreciated.