Data preparation

Data Downloading

The dataset can be downloaded via our project on Physionet. The dataset should be downloaded into this folder, data folder structure after downloading:

"project-home"/data
├── annotation
│   ├── train.json
|   └── test.json
├── train_images
│   ├── 00073745e02e69432c002b527c565151.dicom
│   ├── ...
|   └── fffa8adcc5e692cdb816051b6202870d.dicom
└── test_images
    ├── 004004095d8a302b1c0815ccb044c018.dicom
    ├── ...
    └── ff6a81f9fa386401ce11a0eb74e1f661.dicom

Converting DICOM images to PNG format

Run the bash script. Note that you might need to change the number of CPU workers.

bash convert_dicom.sh

Data folder structure after converting:

"project-home"/data
├── annotation
├── train_images
├── test_images
├── train_pngs
│   ├── 00073745e02e69432c002b527c565151.png
│   ├── ...
|   └── fffa8adcc5e692cdb816051b6202870d.png
└── test_pngs
    ├── 004004095d8a302b1c0815ccb044c018.png
    ├── ...
    └── ff6a81f9fa386401ce11a0eb74e1f661.png

Parse DICOM metadata

To read metadata of DICOM images, run the parse_dicom.sh. Since Detectron2 requires image size for data augmentation, this step is essential.