search

robert-graf / Readable-Conditional-Denoising-Diffusion

Readable Conditional Denoising Diffusion
MIT License
25 stars 1 forks source link

about unpaired MR-CT model #1

Open HBB0517 opened 4 months ago

HBB0517 commented 4 months ago

Dear author, I must commend you on your excellent work. However, I have a question about unpaired MR-CT images. Could you advise on the appropriate model for an MR to CT conversion? Would Syndiff be suitable for this purpose? I noticed that a pre-trained model（syndiff) for MR-CT hasn't been released. Do you have any other models to recommend, or perhaps do you possess a pre-trained model that could be shared? thank you!!!!

robert-graf commented 4 months ago

Syndiff crashed when I used a batch size greater 1. With batch_size = 1 it worked but only utilized 8 % of the GPU. So it took a whole month to train, but it worked out of the box. If you do not encounter this issue, it will be a good choice. It is a CycleGAN that feeds a paired Diffusion.

Alternative normal CycleGAN with a newer Backbone should also be fine. Replacing the generator usually involves workers being out of the box. Changing the Discriminator is very fiddly. There are many papers that improve CycleGAN, but none really stand out or have been established as a standard.

Be aware of this issue with CycleGAN: https://arxiv.org/abs/1712.02950
You will have to match the number/volumes of structures CT and MRI or other mitigating factors. CycleGAN is not reliable for medical tasks because of mismatching issues.

I can not tell you if Syndiff also suffers from stenographic issues.

HBB0517 commented 4 months ago

Thank you for your response! My aim is to use diffusion models as a comparative model, with the task of converting MR images into CT images. These images are in PNG format.