I'm really sorry for bothering you.

ellisdg / 3DUnetCNN

Pytorch 3D U-Net Convolution Neural Network (CNN) designed for medical image segmentation

MIT License

1.9k stars 653 forks source link

I'm really sorry for bothering you. #331

Open 111678900 opened 9 months ago

111678900 commented 9 months ago

I'm really sorry for bothering you. Your code is designed for four channels, but I only need one channel for my application. I have already changed the number of channels. Could you please let me know what other parts I need to modify?

ellisdg commented 9 months ago

If you haven't already, checkout BraTS example notebook for an explanation of the different arguments: https://github.com/ellisdg/3DUnetCNN/blob/master/examples/brats2020/create_config_lowmem.ipynb

Some arguments you may consider changing:

"desired_shape"
"setup_label_hierarchy" should be false since you only have a mask as output based on your previous comment
"resample" and "crop_foreground" may not be necessary depending on the images and image sizes in your dataset.

111678900 commented 9 months ago

718b1af33889dbe20441f44aa34227f I'm really sorry for bothering you，I'm using it to train in my own dataset.Discovering these images in the example folder after training, does it mean that the training was successful?"If it is successful, how should I run the predict.py file?

ellisdg commented 9 months ago

Those are 3D renderings that I made as an example of skull completion using the AutoImplant dataset.

You should have a training log file that you can look at to see how many epochs were run and if the training and validation losses decreased like you would expect.