RED-CNN: Implementing a Residual Encoder-Decoder Convolutional Neural Network Using PyTorch for Low-Dose Medical CT

RED-CNN is the pytorch implementation of the paper: Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN). It amis to eliminate image noise when reconstructing ct images.

The structure of RED-CNN

The database used in RED-CNN

The 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge by Mayo Clinic

The path of this dataset looks like:

data_path
├── L067
│   ├── quarter_3mm
│   │       ├── L067_QD_3_1.CT.0004.0001 ~ .IMA
│   │       ├── L067_QD_3_1.CT.0004.0002 ~ .IMA
│   │       └── ...
│   └── full_3mm
│           ├── L067_FD_3_1.CT.0004.0001 ~ .IMA
│           ├── L067_FD_3_1.CT.0004.0002 ~ .IMA
│           └── ...
├── L096
│   ├── quarter_3mm
│   │       └── ...
│   └── full_3mm
│           └── ...      
...
│
└── L506
    ├── quarter_3mm
    │       └── ...
    └── full_3mm
            └── ...

The result of RED-CNN

There are several results of RED-CNN in the papaer, we will give a comparison one.

How to use RED-CNN model?

1.run python prep.py to convert ‘dicom file’ to ‘numpy array’

2. run python main.py –load_mode=0 to training.

If the available memory(RAM) is more than 10GB, it is faster to run –load_mode=1.

3.run python main.py –mode=’test’ –test_iters=100000 to test.

The source code is here: RED-CNN Source Code