Experiment discussion

[mathpluscode]

Method

We agree to benchmark with https://github.com/voxelmorph/voxelmorph first.
But which paper/method should be tested first?

Comparison

We can compare

• performances on different metrics (Dice, Hausdorff distance, SSD, etc.)
• training time vs performance (using the same machine and train for the same amount of epochs)
• results visualization

[YipengHu]

ok - i think if we can do these two above, it should be a good start. Three experiments, but with a bit more implementation for the probabilistic loss.

I do not think we need exactly the same dataset as discussed.

the unsupervised version applies mainly on atlas-based registration, so should be reasonably insensitive to pre-processing. #1