search

ellisdg / 3DUnetCNN

Pytorch 3D U-Net Convolution Neural Network (CNN) designed for medical image segmentation
MIT License
1.89k stars 650 forks source link

Predicts only background for lung tumour segmentation #75

Closed trathpai closed 3 years ago

trathpai commented 6 years ago

First of all nice work implementing the U-net . I was trying to make the code work for lung tumour segmentation using isensee model. But it only predicts background in test set. I have 3 labels 0 as background 1 as lung and 2 as tumour in the lung. Any idea how to set the weights for dice coefficients and should i include 0 as label in the config dictionary?

Thanks!!

ellisdg commented 6 years ago

I can't tell you what is going wrong, but don't include 0 in the labels. I have it setup so that the default is 0. What does your training/validation loss look like?

trathpai commented 6 years ago

Thanks for replying , train validation curve look fine to me, seems to converge. But dice coefficient is very poor since its a CT volume and i need to preprocess the volumes.

Pmac23 commented 6 years ago

Hi I have kind of a similar issue with poor dice score with vessel segmentation in 3d volumes. @trathpai did you find a solution to improve the same?

stale[bot] commented 3 years ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. If you have questions, feel free to join the Slack group or email me at davidgellis2@gmail.com.