NIfTI
Last updated
Was this helpful?
Last updated
Was this helpful?
This converter allows you to import NIfTI files into a Supervisely project. It also supports annotations in the NIfTI format (.nii and .nii.gz).
The converter supports both semantic and instance segmentation annotations, as well as import of volumes with no annotations. We will provide an examples of the input structure below.
The converter is backwards compatible with the application.
All volumes from the input directory and its subdirectories will be uploaded to a single dataset
Supported image formats: .nii, .nii.gz.
With annotations: Yes (semantic and instance segmentation).
Supported annotation format: .nii, .nii.gz.
Data structure: Information is provided below.
The NIfTI file should be structured as follows:
If the volume has annotations, they should be in the corresponding directory with the same name as the volume (e.g. CTChest, without extension).
Annotation files should be named according to the following pattern:
Name of the class (e.g. lung, tumor) + .nii or .nii.gz.
The class name should be unique for the current volume (e.g. tumor.nii.gz, lung.nii.gz).
Annotation files can contain multiple objects of the same class (each object should be represented by a different value in the mask).
The NIfTI file should be structured as follows:
For semantic segmentation:
The filename must contain one of the required plane identifiers: axl, cor, or sag, anywhere in the name.
The file representing the anatomical volume should have anatomical string in it's filename representing the volume type.
The annotation file for all classes should also include the prefix, volume type label (inference, mask, ann etc.), ending with .nii or .nii.gz.
For instance segmentation:
The anatomical volume file must include the plane identifier (axl, cor, or sag), anatomic type label and end with .nii or .nii.gz.
Each annotation file must also include the plane identifier and type label (inference, mask, ann etc.), ending with .nii or .nii.gz.
Multiple annotation files per plane are supported, each representing a separate class (and may contain multiple objects).
Note: Filenames can include other descriptive parts such as patient or case UIDs, body parts, arbituary strings or other identifiers, as long as the required plane and type identifiers are present and the file extension is .nii or .nii.gz.
The plane identifier must be one of: cor, sag, or axl. The converter uses these prefixes to group volumes and their annotation files, requiring exactly three volumes — one for each prefix per folder.
Structure example for semantic segmentation:
Structure example for instance segmentation:
If you need to import multiple items at once, place each item in a separate folder. The converter supports any folder structure. Folders may be at different levels, and files will be matched by directory (annotation files must be in the same folder as their corresponding volume). All files will be imported into the same dataset.
Structure example for multiple items directory:
The converter will look for an optional TXT file in the input directory. If present, it will be used to create the classes with names and colors corresponding to the pixel values in the NIfTI files.
The TXT file should be structured as follows:
where:
1, 2, ... are the pixel values in the NIfTI files
Femur, Femoral cartilage, ... are the names of the classes
255, 0, 0, ... are the RGB colors of the classes
Plane-structured converter supports uploading annotations separately (uploading annotations to existing volumes). This functionality supports both dataset-scope and project-wide annotation imports.
By default, annotations are matched with their corresponding volumes based on filenames. However, a custom mapping can be provided via a .json file to explicitly define the mapping.
Input structure example for dataset scope:
Input structure example for project-wide import:
Mapping structure should be as follows:
Where key should be annotation filename, and volume ID as value
If you want to import annotations for the entire project via a JSON mapping:
Pack annotations inside folders with corresponding dataset name as in an example above
Specify the dataset name in a .json file in a path-like manner (dataset_name/annotation_filename)
Example JSON structure with dataset specification: