Why the Dice of nnU-Net on KiTS21 in your paper is so much lower than which reported in KiTS21's official repository?

[zhaoshishuang]

In your paper, the dice of nnU-Net on KiTS21 is 73.07

However, the dice reported in KiTS21's official repository is higher than 0.84

In your code, there is nothing about nnU-Net.

[zhaoshishuang]

Can you explain how you get the dice 73.07 with nnU-Net?

[peterant330]

Can you explain how you get the dice 73.07 with nnU-Net?

Hi,

As we discussed in the Discussion and Conclusion Section, we are not using the organ segmentation label. The original nnU-NET treat the task as three-class segmentation and the organ label can help the tumor segmentation task, while our experiments treat it as binary segmentation problem without using any information from organ labels.

The reason that there is no nnU-Net code in our github is that nnU-Net is highly integrated. Which is hard to be integrated into our framework. You can reproduce this 73.07 result using the official code provided by nnU-Net by removing the organ labels.

[zhaoshishuang]

Thank you!