Image formats and readers
AUDIT configures the clinical modality and the storage format separately. They answer different questions:
modality: what kind of image is this? It selects domain behaviour such as preprocessing, default sequences, 2D/3D feature semantics, and viewers.formats: how is each input stored? It selects the reader and preserves available metadata such as physical spacing.
For example, an ultrasound image can be stored as PNG, NRRD, DICOM, or NIfTI. Conversely, a NIfTI file is not necessarily an MRI.
Supported readers
| Configuration value | Reader | Input |
|---|---|---|
.nii, .nii.gz, nifti |
NiftiReader |
NIfTI file |
.nrrd, nrrd |
NrrdReader |
NRRD file |
.mha, .mhd, mha |
MhaReader |
MetaImage file |
.png, .jpg, .jpeg, .tif, .tiff, .bmp, raster |
RasterImageReader |
2D raster image |
.npy, .npz, array |
ArrayReader |
NumPy array |
dicom_series |
DicomSeriesReader |
Directory containing one DICOM series |
dicom_seg, .dcm |
DicomSegReader |
DICOM SEG file |
auto |
inferred from path | Supported file extension or DICOM directory |
Configuration
The normal form is a mapping so that raw images, ground truth, and predictions may use different storage formats:
For .npz, select the array explicitly when the archive contains more than
one:
Array files do not contain spatial metadata. Physical spacing can be supplied in NumPy axis order:
If a DICOM directory contains multiple series, select one with
reader_options.<kind>.series_id.
Raster segmentations should be grayscale or palette-indexed masks whose pixel values are the configured labels. RGB colour-coded masks are converted to luminance; v0.2.0 does not include a colour-to-label mapping.
Dataset layout
File readers use the existing AUDIT subject layout:
For a single-image US case with no configured sequences:
A DICOM series can be placed in either
case_001/case_001_<sequence>/ or the subject directory itself.