search

xi-jia / LKU-Net

The official implementation of U-Net vs Transformer: Is U-Net Outdated in Medical Image Registration?
67 stars 9 forks source link

Bilinear interpolation when warping image labels. #1

Open junyuchen245 opened 2 years ago

junyuchen245 commented 2 years ago

Hi @xi-jia ,

This is Junyu, the author of TransMorph. Thank you for using our code, and I would like to share that TransMorph has been accepted for publication in Medical Image Analysis.

As pointed out by one of the reviewers from the journal, bilinear interpolation is preferable to nearest neighbor when warping labels. Accordingly, our paper has been revised, as well as the benchmark table on TransMorph_IXI. Using bilinear interpolation yields improved Dice scores. Therefore, I recommend you try it out as well. The modification is fairly simple, which you can find it here: https://github.com/junyuchen245/TransMorph_Transformer_for_Medical_Image_Registration/blob/ed7579aaf03d497edff9370163cc740c6656203f/IXI/TransMorph/infer_TransMorph.py#L64-L73 and https://github.com/junyuchen245/TransMorph_Transformer_for_Medical_Image_Registration/blob/ed7579aaf03d497edff9370163cc740c6656203f/IXI/TransMorph/infer_TransMorph.py#L42

Junyu

xi-jia commented 2 years ago

Hi Junyu

Thank you for the heads up and congrats on the acceptance of TransMorph! It is indeed a very great framework and I've learned and benefited much from it.

With your (as well as the reviewers') advice, I've updated the results, which can be found at https://github.com/xi-jia/LKU-Net#updates-2022-sep-08

If you have any further issues, please feel free to contact me.

Best, Xi