Is there clDice loss function for 3D image?

dmitrysarov / clDice

pyTorch implementation of clDice

28 stars 5 forks source link

Is there clDice loss function for 3D image? #2

Closed artzers closed 3 years ago

artzers commented 3 years ago

Hello, I want to find 3D clDice loss function for 3D tublar segmentation. As I know in some cldice-relate articles, 3D cldice loss has been applied but the corresponding codes are not published. Do you have any idea ? Thank you.

dmitrysarov commented 3 years ago

Good question. Sure you can by substituting pooling with 3d version and tweak minor parts. Contribution welcome!

AlphaJia commented 3 years ago

I tried the 3D clDicel loss, compared to other loss trained model, this clDice loss model achieved the highest recall and lowest precision. Interesting! it is most of true positive and false positive has been segmented out!

AlphaJia commented 3 years ago

I think the pred contour can be get using sobel operator

jocpae commented 3 years ago

Dear community, after acceptance in CVPR 2021 we now offer an official 2D and 3D implementation of cldice at https://github.com/jocpae/clDice . Thanks a lot to @dmitrysarov for implementing cldice :)

dmitrysarov commented 3 years ago

glad to help