search

YipengHu / label-reg

(This repo is no longer up-to-date. Any updates will be at https://github.com/DeepRegNet/DeepReg/) A demo of the re-factored label-driven registration code, based on "Weakly-supervised convolutional neural networks for multimodal image registration"
Apache License 2.0
117 stars 32 forks source link

Dice loss problems #16

Closed Lijiaying201812 closed 5 years ago

Lijiaying201812 commented 5 years ago

When I run the program with example downloaded data,I found there was a problem with label_indices =1 and label_indices =2 ,because the dice were almost zero for these condition,What's the reason for the case? Thanks!

YipengHu commented 5 years ago

Hi thanks for the question. It is expected in particular during the initial stages of training. As most of the labels are small (except for Labels #0 which are always the gland segmentation), so they would not have any overlap before more iterations in. Even when it closes to convergence, some of them still can be quite small even zeros.

Just FYI: these non-overlap landmarks still generate useful gradients, due to the multiscale loss, as explained in the paper.

On Wed, 20 Mar 2019 at 02:41, Lijiaying201812 notifications@github.com wrote:

When I run the program with example downloaded data,I found there was a problem with label_indices =1 and label_indices =2 ,because the dice were almost zero for these condition,What's the reason for the case? Thanks!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/YipengHu/label-reg/issues/16, or mute the thread https://github.com/notifications/unsubscribe-auth/ACd9_YR0JLQSpFJ4kpcSTXbrr9GW8eCsks5vYZ_kgaJpZM4b-B4k .