yingsong1 / dosepredcition Goto Github PK

Dear Yingsong,
Congrats for the publication. I read your paper thoroughly. I'm a little bit confused about the "without perceptual loss, there was a more uniform dose fall-off around the PTV with no obvious modulated dose distribution features " in the paper.

I saw you used sigmoid at the last layer of the deeplap3+, could you explain a bit why do you use it? Could it be the problem that caused the problem you've mentioned in the paper? From my understanding, the dose prediction is a regression task, while sigmoid's nonlinearity destroys the linear requirements from dose values. Adding a trained VGG16 perceptual loss acts as an encoder to map [0,1] post-sigmoid values back to ordinary dose values, through back-propagation it alleviates the impact from sigmoid. This might be a redundant solution. Have you tried to remove the sigmoid layer in the experiment?

Maybe I'm wrong, could you explain more?

Best regards,
Jintao

Could you please provide several training samples ? Thanks! (Since I cannot run the code successfully for lacking the training data)

could you give a data example for display?