Skip to content

Latest commit

 

History

History
60 lines (47 loc) · 5.54 KB

dataset_format.md

File metadata and controls

60 lines (47 loc) · 5.54 KB
Raw

Dataset format

In this work, the dataset consist of three components: CT and PET images and the corresponding lesion segmentation mask, all in NIFTI file format. If your dataset is in DICOM format, you can convert them to NIFTI using the method described in dicom_to_nifti_conversion.md. After converting DICOM images to NIFTI format, you may have to resample you CT (and/or GT) images to PET geometry (if your CT or GT images are not in PET geometry). If this is the case, use the functions resample_ct_to_pt_geometry() and resample_gt_to_pt_geometry() in data_conversion/resample_ct2pt.py.

Training cases filenaming convention

We follow a similar filenaming convention as used by nnUNet. Each training case is associated with a unique identifier, which is a unique name for that case. This identifier is used by our code to connect images (PET/CT) with the correct segmentation mask. We suggest using the unique identifier as {PatientID}_{StudyDate}.

A training case consists of images and their corresponding segmentation masks.

Images: Our networks utilize two channel 3D images, the first channel being the CT images and the second channel being the PET image. Both CT and PET MUST have the same geometry (same size, spacing, origin, direction) and must be (approximately) coregistered (if applicable). To resample CT images to PET resolution, use the function resample_ct_to_pt_geometry() in data_conversion/resample_ct2pt.py. Within a training case, all image geometries (input channels, corresponding segmentation) must match. Between training cases, they can of course differ.

Segmentations must share the same geometry as their corresponding images (same size, spacing, origin, direction). Segmentations are integer maps with each value representing a semantic class; the background is represented by 0. In our work, we used segmentation masks with two classes: 0 for background and 1 for lesions. All masks in the training set MUST have 0s and 1s; the current version of code cannot handle negative images (images with no lesions) without changing some of the preprocessing transforms (like RandCropByPosNegLabeld, etc.) applied to the images before giving them as inputs to the network.

Given a unique identifier for a case, {PatientID}_{StudyDate}, the CT, PET and GT image filenames should be:
CT image: {PatientID}_{StudyDate}_0000.nii.gz,
PET image: {PatientID}_{StudyDate}_0001.nii.gz,
GT image: {PatientID}_{StudyDate}.nii.gz,

Important: The input channels must be consistent! Concretely, all images need the same input channels in the same order and all input channels have to be present every time. This is also true for inference!

Dataset folder structure

Create a folder named lymphoma.segmentation in the location of your choice. The is the master folder that stored all your datasets, the trained models and training/validation logs, predictions or any other results based on predictions. Go to the file config.py and update the variable LYMPHOMA_SEGMENTATION_FOLDER as the absolute path to the folder lymphoma.segmentation. Within lymphoma.segmentation, create a folder named data, which should be the location of your training and test datasets. After these steps, your directory structure is expected to look like this:

└───lymphoma.segmentation/data
    ├── imagesTr
    ├── imagesTs  # optional
    ├── labelsTr  
    └── labelsTs # optional
  • imagesTr contains the images (CT and PET) belonging to the training cases. Each corresponding CT and PET images should be in the same geometry (same size, spacing, origin, direction) in this folder.
  • imagesTs (optional) contains the images that belong to the test cases. Each corresponding CT and PET images should be in the same geometry (same size, spacing, origin, direction) in this folder.
  • labelsTr contains the images with the ground truth segmentation maps for the training cases. These should be in the same geometry (same size, spacing, origin, direction) as their corresponding PET/CT images in imagesTr.
  • labelsTs (optional) contains the images with the ground truth segmentation maps for the test cases. These should be in the same geometry (same size, spacing, origin, direction) as their corresponding PET/CT images in imagesTs.

After moving all the training and test images and masks in the respective folders, the directory structure should look like this:

└───lymphoma.segmentation/data/
    ├── imagesTr
    │   ├── Patient0001_20110502_0000.nii.gz
    │   ├── Patient0001_20110502_0001.nii.gz
    │   ├── Patient0002_20150514_0000.nii.gz
    │   ├── Patient0002_20150514_0001.nii.gz
    │   ├── ...
    ├── imagesTs  # optional
    │   ├── Patient0003_20190402_0000.nii.gz
    │   ├── Patient0003_20190402_0001.nii.gz
    │   ├── Patient0004_20150514_0000.nii.gz
    │   ├── Patient0004_20150514_0001.nii.gz
    │   ├── ...
    ├── labelsTr 
    │   ├── Patient0001_20110502.nii.gz
    │   ├── Patient0002_20110502.nii.gz 
    │   ├── ...
    └── labelsTs # optional
        ├── Patient0003_20190402.nii.gz
        ├── Patient0004_20160204.nii.gz 
        ├── ...