search

LeeJunHyun / Image_Segmentation

Pytorch implementation of U-Net, R2U-Net, Attention U-Net, and Attention R2U-Net.
2.66k stars 594 forks source link

Testing these models on Retinal Vessel Segmentation #61

Closed DeVriesMatt closed 1 year ago

DeVriesMatt commented 4 years ago

Hi, thanks for the brilliant repo.

I was wondering if you may have trained and tested these models on retinal segmentation tasks. I am currently doing so and the R2U-Net produces great training results however for validation and test sets, it returns all-black images.

Do you have any suggestions to what may fix this?

Thanks again!

sdc17 commented 4 years ago

@DeVriesMatt Hey, I have the same problem with you when I was training R2U-Net and R2AttU-Net on my own dataset. Have you solved this problem yet? These two models produced good training scores but really bad test scores while AttU-Net worked well on both training and testing. Thanks!

DeVriesMatt commented 4 years ago

@sdc17 Yeah, I still haven't been able to figure out the problem. I have tried adding dropout, using data augmentation, and tried many different loss functions however nothing has worked. Please let me know if you figure anything out.

xlh1135355585 commented 3 years ago

I also conducted experiments on this data set and found that the final experimental effects of these networks are almost the same. Have you encountered this problem?