At testing phase, for the dice score of Kidney, whether the tumor is considered as a part of kidney?
Thanks for the good question and sorry for the delayed response.
The tumor will be considered part of the kidney when computing the kidney Dice. Otherwise the kidney Dice will have too high a dependence on the kidney-tumor delineation, which is not our intention.
It occurs to me that this has not been made clear in the rules. I will amend them.
Thanks for organizing the competition.
I am confused by your statement about tumor being part of kidney. If a method predicts a tumor voxel as tumor but not kidney would it be penalized in the metric. In other words, should I consider tumor+kidney as one meta class (i.e. kidney) and the tumor as a subclass of the meta class. This way this problem become a multi-label problem rather than a multi-class problem.
Yes, your understanding it correct. Tumor+kidney is one meta class, and tumor is a subset of it. I’ve added a
starter_code/evaluation.py file to the GitHub repo which I hope makes this clear.
Hello, I encountered “RuntimeWarning: invalid value encountered in long_scalars” when computing the dice coefficient, and the result was Not A Number(NaN). It seems that the numbers are very big. Can you improve this algorithm?
I’m not sure why this is happening, but overflow seems unlikely. I’ve created a new issue on GitHub for this, let’s move the discussion there. Could you provide a full stack trace?