search

bigmb / Unet-Segmentation-Pytorch-Nest-of-Unets

Implementation of different kinds of Unet Models for Image Segmentation - Unet , RCNN-Unet, Attention Unet, RCNN-Attention Unet, Nested Unet
MIT License
1.87k stars 345 forks source link

About the performance on ISIC2017 dataset #34

Closed wangyxxjtu closed 4 years ago

wangyxxjtu commented 4 years ago

Hello, Great Work! I have some questions about the results on ISIC2017(skin cancer segmentation dataset). I trained the model using your code, and got satisfied results on DRIVE (eye ) and Lung segmentation datasets, but I can't get normal results on ISIC 2017, the performance is very poor, the dc is only 0.34. So, Do you have any idea about the problem?

bigmb commented 4 years ago

Thanks. Did you check the intermediate training images and see if the segmentation is detecting the location of Gt correctly?

wangyxxjtu commented 4 years ago

Thank you for your reply. Following your suggestion, I randomly visualized several images and the corresponding masks during tranining, all these samples are exactly matched.

bigmb commented 4 years ago

Ok. And all the Unets are showing the same results? Did you try with Nested-Unet?

wangyxxjtu commented 4 years ago

Thank you. I found the problem, I made some changes on data transforms in training, learning to the inconsistent between the training data and the testing data. Unfortunately, after I fixed the data transform problem, the performance is still not reasonable, for unet on ISIC 2017, the DC is only 0.69 (0.86 reported in r2u-net paper), and the nestedu- net is only 0.73, which are much worse.

bigmb commented 4 years ago

Try to increase the depth of the model or decrease it. Maybe that can provide a better dice score.

wangyxxjtu commented 4 years ago

Ok, thank you.