This code is deprecated and paper + code should be taken with a big grain of salt
This is the codebase for diffusion models for MRI rec The code is very much based on Improved Denoising Diffusion Probabilistic Models. So, if you have any problems check out their repo first...
The training data is stored on the cluster at /itet-stor/username/bmicdatasets_bmicnas01/Processed/fastMRI_with_espiritcoils/fastMRI_train_mid
script/image_train.py
script/image_sample.py This script is a bit adhoc. Works but needs improvements.
Complex values is stored as channels.
Since fastMRI is different sizes, batchsize of 1 is implemented for now.
If you have any questions, feel free to reach out to me: Jonatan Kronander jonatan@kronander.se
First sota paper: Denoising Diffusion Probabilistic Models https://proceedings.neurips.cc/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf
Improving DDPMS:
Improved Denoising Diffusion Probabilistic Models
https://arxiv.org/abs/2102.09672
Latent spaced DDPMs: High-Resolution Image Synthesis with Latent Diffusion Models https://arxiv.org/abs/2112.10752
Diffusion for MRI: Towards performant and reliable undersampled MR reconstruction via diffusion model sampling https://arxiv.org/pdf/2203.04292.pdf High-Frequency Space Diffusion Models for Accelerated MRI https://arxiv.org/pdf/2208.05481.pdf Measurement-conditioned Denoising Diffusion Probabilistic Model for Under-sampled Medical Image Reconstruction https://arxiv.org/pdf/2203.03623.pdf Adaptive Diffusion Priors for Accelerated MRI Reconstruction https://arxiv.org/pdf/2207.05876.pdf
"paper.pdf": I wrote a short summary of our plan that didn't work due to sensitivity maps (Ender have more infos on why it didn't work). Note that this paper is not read by anyone but me so there is no proof reading done and should be read with a grain of salt. You can find it paper.pdf
I found the the model to work ok for single coil images. Needs to be check more though... However, I never got the multicoil setting to work. This is probably because the sensitivity maps, talk to Ender for more information. One way forward would be to fix this. Another way forward would be to do the diffusion process in latent space (ie. latent varibles of AE. Read more here https://arxiv.org/abs/2112.10752)