This repository contains the code for the main models described in “Back-Projection Diffusion: Solving the Wideband Inverse Scattering Problem with Diffusion Models”, which is available on ArXiv. The paper is authored by Borong Zhang, Martín Guerra, Qin Li, and Leonardo Zepeda-Núñez. This repository is written by Borong Zhang.
We present Wideband Back-Projection Diffusion, an end-to-end probabilistic framework for approximating the posterior distribution of the refractive index using the wideband scattering data through the inverse scattering map.
This framework produces highly accurate reconstructions, leveraging conditional diffusion models to draw samples, and also honors the symmetries of the underlying physics of wave-propagation. The procedure is factored into two steps, with the first, inspired by the filtered back-propagation formula, transforms data into a physics-based latent representation, while the second learns a conditional score function conditioned on this latent representation.
These two steps individually obey their associated symmetries and are amenable to compression by imposing the rank structure found in the filtered back-projection formula. Empirically, our framework has both low sample and computational complexity, with its number of parameters scaling only sub-linearly with the target resolution, and has stable training dynamics. It provides sharp reconstructions effortlessly and is capable of recovering even sub-Nyquist features in the multiple-scattering regime.
Project Environment can be installed by
conda create -n back_projection_diffusion python=3.11
conda activate back_projection_diffusion
pip install git+https://github.com/borongzhang/back_projection_diffusion.git@main
pip install --upgrade "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
Demos for EquiNet-CNN, B-EquiNet-CNN, Analytical-CNN and SwitchNet-CNN trained on the 10h Overlapping Squares dataset can be found in the examples folder.
We make a sample dataset and corresponding trained model parameters publicly available via Zenodo. To download the dataset and trained model parameters, simply run:
zenodo_get 14911327
After downloading the data, by default, please place the datasets in a data folder. Optionally, the tmp folder—which contains the trained model parameters—can be downloaded and placed in the examples folder. If the trained model parameters in tmp are loaded, the training scripts will be automatically skipped.
To change the data directory, modify training_data_path and test_data_path. To change the directory for the trained model parameters, modify cond_workdir.
The following is the default directory structure for the back_projection_diffusion project:
back_projection_diffusion/
├── examples/
│ └── tmp/ # Contains trained model parameters
│ ├── equinet_cnn_10hsquares/
│ ├── b_equinet_cnn_10hsquares/
│ ├── analytical_cnn_10hsquares/
│ └── switchnet_cnn_10hsquares/
├── data/
│ ├── 10hsquares_trainingdata/ # Contains training data for 10h squares
│ └── 10hsquares_testdata/ # Contains test data for 10h squares
└── src/
Datasets can be generated using the MATLAB code in the data_generation folder.
Perturbations Shepp-Logan, 3-5-10h Triangles, and 10h Overlapping Squares can be generated using the corresponding eta_generation_?.m scripts.
The resolution and number of perturbations should be specified in the setup scaling parameters section. The generated perturbations will be stored as an HDF5 file with the following structure:
eta.h5/
├── /eta
The Brain MRI images used as our perturbations were obtained from the NYU fastMRI Initiative database, as described in the works of Florian Knoll et al., fastmri: A publicly available raw k-space and dicom dataset of knee images for accelerated mr image reconstruction using machine learning, 2020, and Jure Zbontar et al., fastmri: An open dataset and benchmarks for accelerated mri, 2019. As such, NYU fastMRI investigators provided the data but did not participate in the analysis or writing of this report. A listing of NYU fastMRI investigators, subject to updates, can be found at fastmri.med.nyu.edu. The primary goal of fastMRI is to test whether machine learning can aid in the reconstruction of medical images.
We padded, resized, and normalized the perturbations to a native resolution of
We make the processed MRI brain perturbations publicly available via Zenodo. The perturbations are stored as HDF5 files, with filenames in the format eta-n.h5, where n corresponds to the resolution. To download the dataset, please run:
zenodo_get 14760123
Scattering data can be generated using the data_generation.m script.
The dimension, size, and frequencies of the data should be specified in the setup scaling parameters section. The generated scattering data will be stored as an HDF5 file with the following structure:
scatter.h5/
├── /scatter_imag_freq_1
├── /scatter_real_freq_1
├── /scatter_imag_freq_2
├── /scatter_real_freq_2
├── /scatter_imag_freq_3
├── /scatter_real_freq_3
To use, place the eta.h5 and scatter.h5 files into a folder, then move that folder to the data directory.
We have made the baseline deterministic models, and the U-ViT diffusion model publicly available in the repository.
This repository makes use of code from the following sources:
- The code in this repository is largely based on Swirl-Dynamics.
- Random Shepp-Logan Phantom by Matthias Chung, which was used for generating the
Shepp-Logandataset. - Original data generation code was provided by the authors of Wide-Band Butterfly Network.
- The structure of the repository is inspired by Owen Melia and Olivia Tsang‘s repository.
If this code is useful to your research, please cite our preprint:
@misc{zhang2024backprojectiondiffusionsolvingwideband,
title={Back-Projection Diffusion: Solving the Wideband Inverse Scattering Problem with Diffusion Models},
author={Borong Zhang and Martín Guerra and Qin Li and Leonardo Zepeda-Núñez},
year={2024},
eprint={2408.02866},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2408.02866},
}