Learning Difference-of-Convex Regularizers for Inverse Problems: A Flexible Framework with Theoretical Guarantees

Environment Installation

conda env create -f environment.yml

If the installation fails using the environment.yml file, you can install the required packages manually.

Prepare Data(Mayo Grand Challenge)

Please go to the link and download the data. Please put the data in the following path:

./data/[test/valid/train]

In all of our experiments, we use 1mm B30 data of the highest precision.

Run Our Algorithm(ADCR)

example call:

CUDA_VISIBLE_DEVICES=0 python train.py --setup=5 --dataperc=100 --epochs=1 --lr=1e-4 --eps=1e-6 --alg=ADCR --iterates=200 --valid=128 --test=True --batch-size=10 --noise=3.2 --seed=10  --setting=limited --load=./data/nets_new/ADCR/limited/limited.pt --test_mode=GD 

• epochs: number of epochs
• lr: learning rate (used to train the neural network)
• eps: step size for the optimization
• alg: algorithm to be used (ADR, TV, FBP, ADCR etc.), please take a look at files names in the Algorithms folder
• iterates: maximum number of iterations for the optimization
• valid: number of validation images
• batch-size: batch size
• detectors: number of detectors
• noise: noise level
• load: path to the checkpoint
• seed: random seed
• setting: limited or sparse
• test_mode: GD or CCP or PSM
• K: number of innner loop iterations for the CCP/PSM algorithm

Train a regularizer

example call:

CUDA_VISIBLE_DEVICES=0 python train.py --setup=5 --dataperc=100 --epochs=20 --lr=1e-4 --eps=1e-5 --alg=ADCR --iterates=100 --valid=10 --batch-size=10  --gpu=1 --noise=3.2 --load=False --seed=10 --wclip=True --setting=sparse --mu=10

Credits

This code is based on the CT_framework.

Description of other methods

• FBP (Filtered back-projection)
• TV (Total Variation)
• ADR (Adversarial Regularizer): https://arxiv.org/abs/1805.11572
• LG (Learned gradient descent): https://arxiv.org/abs/1704.04058
• LPD (Learned primal dual): https://arxiv.org/abs/1707.06474
• FL (Fully learned): https://nature.com/articles/nature25988.pdf
• FBP+U (FBP with a U-Net denoiser): https://arxiv.org/abs/1505.04597
• ACR: https://arxiv.org/abs/2008.02839
• ACNCR: https://openreview.net/forum?id=yavtWi6ew9
• AWCR: https://arxiv.org/abs/2402.01052
• ADCR(ours)

In order to add your own algorithms to the list, create a new file in the Algorithms folder in the form name.py and use BaseAlg.py as the template.