Modularized Adaptive Processing Neural Network (MAP-NN)

This repository contains the code for our neural network “MAP-NN” described in the following paper:

Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction Nature Machine Intelligence 1, 269–276 (2019)

Prerequisites

The dependent packages for MAP-NN are as follows:

• Python v2.7
• TensorFlow v1.8.0
• Numpy v.1.2.2
• Scikit-learn v0.19.1
• Matplotlib v2.1.2
• CUDA v9.2

Usage

Prepare the training data

In order to start the training process, please prepare your training data in the following form:

• data: N x 1 x W x H
• label: N x 1 x W x H

where data and label represent the low-dose and normal-dose CT patches respectively, N, W, and H are the number, width, and height of the training data respectively. In this study, N=128,000, W=H=64, and data and label are stored in your hdh5 file.

Note that the real training and validation sets are around 2GB and 1GB respectively. Here we only randomly sampled our original data into reduced training and validation sets of 1,280 and 1,280 patches of size 64x64 respectively to run the training script. However, here we have provided the fully trained model for you to reproduce the denoising results reported in our paper.

Training network

Once the training data are ready, you can directly change the filename in the training script MAP_NN_training.py.

python MAP_NN_Training.py

Testing network

After specifying the testing file name in the testing script MAP_NN_Demo.py, please run the following command to obtain the denoised images.

python MAP_NN_Demo.py

For example, we take a LDCT image ('L506_QD_1_1.CT.0004.0087.2015.12.22.20.46.00.71702.358798544.IMA'). The model will load this LDCT image and produce the denoised images at different denoising depths, which are then saved into the file test_results.pdf.

Citation

If you find this code or our work useful, please cite the following paper:

@article{shan2019competitive,
  title={Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction},
  author={Shan, Hongming and Padole, Atul and Homayounieh, Fatemeh and Kruger, Uwe and Khera, Ruhani Doda and Nitiwarangkul, Chayanin and Kalra, Mannudeep K and Wang, Ge},
  journal={Nature Machine Intelligence},
  volume={1},
  number={6},
  pages={269--276},
  year={2019},
  publisher={Nature Publishing Group}
}

Acknowledgement

The code was modified from https://github.com/igul222/improved_wgan_training.

Contact

shanh at rpi dot edu

Any discussions, suggestions and questions are welcome!