The repository aims to collect different tomographic reconstruction methods based on deep learning. For the forward operator, we need another package TomoForward (Check the installation below).
For the UNet architecture, I modify the code from [3]. Many thanks to the author of [3]! For the bilinear upsampling layer, I use the codes from the branch: FluxML/Flux.jl#1180 . If it is merged, it will be removed.
Install Julia and in Julia REPL type:
julia> ]
pkg> add CUDA, Flux, Zygote
pkg> add https://github.com/JuliaTomo/TomoForward.jl
pkg> add https://github.com/JuliaTomo/DeepReconstruction.jl
- Gadelha et al. [2] based on Deep Image Prior [1]: apply DIP to the tomographic reconstruction task. This code implements [2] with some differences. We use a sparse array as a forward projection operator and use the automatic differentiation from Flux. This makes it possible to use more flexible geometry, whereas [2] only supports the limited geometry.
Please see codes in examples folder.
- Check the paper: M. Z. Darestani and R. Heckel, “Can Un-trained Neural Networks Compete with Trained Neural Networks at Image Reconstruction?,” arXiv:2007.02471 http://arxiv.org/abs/2007.02471.
Loss per iteration
Reconstruction error per iteration
Best reconstruction result (from clean and noisy)
Final reconstruction result
- [1] Ulyanov, D., Vedaldi, A., Lempitsky, V., 2018. Deep image prior, in: CVPR.
- [2] Gadelha, M., Wang, R., Maji, S., 2019. Shape Reconstruction Using Differentiable Projections and Deep Priors, in: ICCV. https://doi.org/10.1109/ICCV.2019.00011
- [3] https://github.com/chinglamchoi/julia_unet/