search

HuCaoFighting / Swin-Unet

[ECCVW 2022] The codes for the work "Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation"
1.66k stars 302 forks source link

Further experiments advise #1

Open parap1uie-s opened 3 years ago

parap1uie-s commented 3 years ago

Nice work!

The listed experiments shown the U-shape Swin Transformer could capture global and local information. However, its ability to capture small target information has not yet been shown, i.e, further experiment on lung nodule segmentation task using LIDC dataset would be helpful.

Siyuan89 commented 3 years ago

I agree with this comment. If you only use transformers for segmentation, it will not be able to capture the small details because transformers are not made for such a task. Convolution is needed to retain the lost local information.

parap1uie-s commented 3 years ago

I agree with this comment. If you only use transformers for segmentation, it will not be able to capture the small details because transformers are not made for such a task. Convolution is needed to retain the lost local information.

Even if the outperformance is not achieved, such an experiment is valuable: Point out the limitations of the proposed method and provide the possibility for improvement.

Acmenwangtuo commented 3 years ago

I also use transformer to segment the nucleis ,but i can't get a better perfomance. Pure transformer suit this small objects too?

Acmenwangtuo commented 3 years ago

I agree with this comment. If you only use transformers for segmentation, it will not be able to capture the small details because transformers are not made for such a task. Convolution is needed to retain the lost local information.

i meet the same question.

swjtulinxi commented 3 years ago

the transformer can't capture small details ?I have read a lot of segmentation papers based on it, and the results are OK