search

Amshaker / unetr_plus_plus

[IEEE TMI-2024] UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
Apache License 2.0
340 stars 32 forks source link

SA and CA concatenated or summed? #40

Closed Leonngm closed 1 year ago

Leonngm commented 1 year ago

Hi all,

in the paper your state that you perform a sum fusion of the spatial and channel attention features. In your code however you use a linear layer to half the sizes and then simply concatenate them.

Which of the two methods did you use to actually produce the results that you mentioned?

Thank you.

Best regards, Leon

Amshaker commented 1 year ago

Hi @Leonngm ,

Thank you for your question.

The results are with some fusion as stated in the paper. However, both ways work well with minor difference in accuracy ~0.1.

Best regards, Abdelrahman.