Low-Dose CT Denoising with Kernel Prediction Networks

Requirements

• numpy
• scipy
• Pillow
• matplotlib
• Pytorch (tested on 1.1.0)
• pydicom
• PyYAML
• astra-toolbox (if one needs to synthesize noisy inputs himself)

Run the code

We use a .yml file to specify various experiment settings. By default, all .yml file are stored in options folder. To run the code:

python main.py -opt AAPMKPN.yml

A sample_option_with_doc.yml is provided to illustrate various settings.

Prepare the data

Download the AAPM dataset into data folder.

The data used by the training and test script should be normalized and stored in .npz files.

The publicly available datasets include:

AAPM-NIH LDCT dataset

DOSE dataset (Notice: the low-dose and full-dose images are not perfectly aligned, only for qualitative comparison): store in bnm0xx-x-x.npz

CDE dataset (patient lung), store in CDE folder.

Train and Test

Config files like xxxx_EVAL.yml are used for evaluation.