Hi! Thank you for your excellent and inspiring work — I really appreciate the clarity and quality of your implementation.
I have one confusion and would love your insight.
In your Dice evaluation, you compute Dice between:
the ground-truth segmentation one-hot maps, and
the atlas one-hot segmentation after warping & interpolation.
When I reproduce the evaluation, I notice that if I argmax the warped one-hot segmentation back to discrete labels and then compute Dice with the original segmentation, the Dice score becomes significantly higher.
Do you have any thoughts on why evaluation chooses to use the soft (interpolated) one-hot rather than the discretized labels?
Thanks a lot in advance!
Hi! Thank you for your excellent and inspiring work — I really appreciate the clarity and quality of your implementation.
I have one confusion and would love your insight.
In your Dice evaluation, you compute Dice between:
the ground-truth segmentation one-hot maps, and
the atlas one-hot segmentation after warping & interpolation.
When I reproduce the evaluation, I notice that if I argmax the warped one-hot segmentation back to discrete labels and then compute Dice with the original segmentation, the Dice score becomes significantly higher.
Do you have any thoughts on why evaluation chooses to use the soft (interpolated) one-hot rather than the discretized labels?
Thanks a lot in advance!