The mean dice after training is less than the initial mean dice

[idhamari]

@mattiaspaul @han-sie

Many thanks for sharing, the idea is interesting, is there any pre-print version of your paper available?

I run your code, as it is, for 900 epochs but I don't see improvement, is there some missing pre-processing or augmentation steps?

Here is part of the log:

Initial Dice (all): 0.35, 0.30, 0.25, 0.25 (mean: 0.29)
epoch (train): 00 -- mean dice case 0: 0.05
epoch (train): 00 -- mean dice case 1: 0.00
epoch (train): 00 -- mean dice case 7: 0.60
epoch (train): 00 -- mean dice case 5: 0.31
epoch (train): 00 -- mean dice case 3: 0.31
epoch (train): 00 -- mean dice case 4: 0.12
epoch (train): 00 -- mean dice case 6: 0.03
epoch (train): 00 -- mean dice case 2: 0.32
epoch (train): 00 -- loss: 0.0112 -- time(s): 92.1

100 epoch:

epoch (train): 100 -- mean dice case 5: 0.62
epoch (train): 100 -- mean dice case 1: 0.00
epoch (train): 100 -- mean dice case 2: 0.50
epoch (train): 100 -- mean dice case 3: 0.48
epoch (train): 100 -- mean dice case 4: 0.35
epoch (train): 100 -- mean dice case 6: 0.29
epoch (train): 100 -- mean dice case 0: 0.23
epoch (train): 100 -- mean dice case 7: 0.44
epoch (train): 100 -- loss: 0.0063 -- time(s): 92.2

900 epoch:

epoch (train): 900 -- mean dice case 6: 0.00
epoch (train): 900 -- mean dice case 3: 0.44
epoch (train): 900 -- mean dice case 1: 0.00
epoch (train): 900 -- mean dice case 4: 0.51
epoch (train): 900 -- mean dice case 2: 0.36
epoch (train): 900 -- mean dice case 5: 0.19
epoch (train): 900 -- mean dice case 0: 0.09
epoch (train): 900 -- mean dice case 7: 0.64
epoch (train): 900 -- loss: 0.0055 -- time(s): 92.1

The mean dice is 0.278 less than the initial mean dice.