Discrepancy in KiTS19 spacing for reproducing 3D full resolution U-Net results

[w6335089]

Dear nnU-Net authors, I am currently working on reproducing the results from your paper "Automated Design of Deep Learning Methods for Biomedical Image Segmentation" (https://arxiv.org/abs/1904.08128) for the KiTS19 dataset (D17). Specifically, I am focusing on the 3D full resolution U-Net results: 3D_fullres: Kidney: 0.9702 Tumor: 0.8367 Mean: 0.9035 In the paper, you mention using a target spacing of 0.78 x 0.78 x 0.78 mm for this dataset. However, when I use this spacing, my results differ significantly from those reported. After some experimentation, I found that using a spacing of 3.22 x 1.62 x 1.62 mm produces results much closer to those reported in the paper. I noticed that you mentioned removing six cases in your paper: "Cases 15 and 37 were confirmed to be faulty by the challenge organizers (https://github.com/neheller/kits19/issues/21) which is why we replaced their respective segmentation masks with predictions of one of our networks. We furthermore excluded cases 23, 68, 125 and 133 because we suspected labeling errors in these cases as well." My question is: Is there any preprocessing step or parameter adjustment not mentioned in the paper that might explain this discrepancy in spacing? I would greatly appreciate any insights you could provide to help me understand and resolve this discrepancy. Your guidance would be invaluable in ensuring the reproducibility of your excellent work. Thank you for your time and assistance.