cilix-ai / on-the-fly-guidance

[MICCAI 2024] On-the-Fly Guidance Training for Medical Image Registration. Pre-print available in link below.
https://arxiv.org/abs/2308.15216
Creative Commons Attribution 4.0 International
29 stars 3 forks source link

DSC low #1

Closed shy922 closed 1 year ago

shy922 commented 1 year ago

I am training voxelmorph with IXI dataset when 200 epochs. DSC only 0.41. Why?

yuelinxin commented 1 year ago

Hi @1135451439 , Are you training with or without the optimizer? If your DSC is only 0.4 when 200 epochs, there could be a couple of reasons for this:

  1. Dataset problem: you could be using the dataset incorrectly, or the dataset itself might not be what you expected it to be;
  2. Hyper-parameter problem: check if your HP setting aligns with VoxelMorph's ideal setting (which should be the ones we provided given that you haven't changed ours);
  3. Or something else;

In any case, please let us know the problem and we will try to help you solve it.

Here's our training validation DSC for your reference (with VoxelMorph on IXI): image

shy922 commented 1 year ago

I have not changed your code, I downloaded the officially processed ixi data set, I do not know why only 0.4.

yuelinxin commented 1 year ago

If you downloaded the dataset from the official IXI website, that dataset is NOT properly preprocessed yet, and that's most likely the source of your problem, there are 2 solutions to this:

  1. You can preprocess the dataset yourself;
  2. Or you can go to TransMorph's repository to download their already preprocessed dataset;
shy922 commented 1 year ago

I downloaded the dataset from transmorph.

yuelinxin commented 1 year ago

Hi @1135451439 , I hope you have resolved your issue, if you still have problems please let us know :)

shy922 commented 1 year ago

Hi @1135451439 , I hope you have resolved your issue, if you still have problems please let us know :)

Thank you evry much!The problem has been solved because of my pytorch environment. Do you know the LKU paper?Could you try the LKU paper also using this module you proposed?

jsxzs commented 1 year ago

Maybe you can try it yourself. It's easy and convenient to apply our method on such deep-learning-based models. It should have improvement as well.