medio

Medical images I/O python package

This package unifies the io engines of itk, nibabel, and pydicom (and dicom-numpy) packages in a simple and comprehensive interface.

It includes conversion between the metadata conventions, reorientations, affine matrix computation for itk and pydicom and saving dicom series or file.

Installation

First, make sure you have the latest pip version (better to close PyCharm or any other program which uses the environments):

(<env-name>) >pip install -U pip

Install medio with:

(<env-name>) >pip install -U medio

This will install the medio python package and its dependencies in your environment.

Requirements

The dependencies are:

• numpy
• itk (itk-io, itk-filtering)
• nibabel
• pydicom
• dicom-numpy

A conda environment .yml file is in the project's root.

Usage

There are 3 main functions in medio: read_img, save_img and save_dir.

from medio import read_img, save_img
# read a dicom series from a folder
array, metadata = read_img('data/dicom-folder/', desired_ornt='IAR')
# do your stuff and save in any format
save_img('ct.nii.gz', array, metadata, backend='nib')

Documentation

Reading and Saving Images

read_img

medio.read_img(input_path, desired_ornt=None, backend=None, dtype=None, header=False, channels_axis=-1, **kwargs)

• input_path: path-like
  Path for the data to be read (str or pathlib.Path object for example). It can be a file or a folder (in the case of a dicom series). It is the only required parameter. If the input path is a folder, it should contain a single dicom series.
• Returns: array, metadata
  array of type numpy.ndarray and metadata of type medio.MetaData. The first is a numpy array of the image, and the second is a metadata object of the image (see MetaData class documentation).

Optional parameters:

• desired_ornt: orientation string or None
  The desired orientation of the returned image array, e.g. 'RAI'. If None, no reorientation is performed. The desired orientation is in itk standard, even when the IO engine ("backend") is nibabel which uses a different standard (see Orientation).
  If you use pydicom backend, it should be None.
  If desired_ornt is the same as the original image orientation, no reorientation is performed.
• backend: 'nib', 'itk', 'pydicom', 'pdcm', or None
  The backend IO engine to use: 'nib' (nibabel), 'itk' or 'pydicom' (also 'pdcm'). If None, the backend is chosen automatically: 'nib' for nifti files (e.g. '.nii' or '.nii.gz' suffix), otherwise 'itk'.
• dtype: numpy data-type or None
  If not None, equivalent to array.astype(dtype) on the returned image array.
• header: bool
  If True, the returned metadata includes also a metadata.header attribute which stores the raw metadata of the file as a dictionary.
  This is not implemented for series of files (folder input_path), and not used during saving.
• channels_axis: int or None
  If not None and the image has more than a single channel / component (e.g. RGB or RGBA), the channels axis are is channels_axis. If None, the backend's original convention is used.

**kwargs are additional per-backend optional parameters:

• 'itk' backend:

  • pixel_type=itk.SS: itk pixel-type or None
    Itk pixel type of the image file/folder. The default value is int16 (itk.SS - Signed Short). Other common pixel types are: itk.UC - uint8, itk.US - uint16.
    You can use the function itk.ctype in order to convert C-types to itk types. For example:
    itk.ctype('unsigned short') == itk.US
  • fallback_only=True: bool
    If True, the pixel type is automatically found and if failed then pixel_type is used (pixel_type must be not None in this case).
    Note: if itk.imread(input_path) fails, using fallback_only=True will result in a slightly inferior performance. If you know what is pixel-type of the image, you can set it with pixel_type and use fallback_only=False.

• 'pydicom' backend

  • globber='*': str
    Relevant for a directory - glob pattern for selecting the series files (all files by default).
  • allow_default_affine=False: bool
    Relevant for multiframe dicom file - if True and the dicom miss some physical tags for the affine calculation, use a default affine value.

save_img

medio.save_img(filename, np_image, metadata, use_original_ornt=True, backend=None, dtype=None, channels_axis=None, mkdir=False, parents=False, **kwargs)

• filename: path-like
  The file to be saved.
• np_image: numpy.ndarray
  The image array.
• metadata: medio.MetaData
  The corresponding metadata.

Optional parameters:

• use_original_ornt: bool
  Whether to save in the original orientation stored in metadata.orig_ornt or not.
• backend: 'nib', 'itk' or None
  The backend to use: 'nib' or 'itk'. If None, 'nib' is chosen for nifti files and 'itk' otherwise.
• dtype: numpy data-type or None
  If not None, equivalent to passing np_image.astype(dtype). Note that not every dtype is supported in saving, so make sure what is the dtype of the image array you want to save.
• channels_axis: int or None
  If not None, the image has channels (e.g. RGB) along the axis channels_axis of np_image.
• mkdir: bool
  If True, creates the directory of filename.
• parents: bool
  To be used with mkdir=True. If True, creates also the parent directories.

'itk' backend optional parameters (**kwargs):

• allow_dcm_reorient=False: bool
  When saving a dicom file ('.dcm' or '.dicom' suffix) the image orientation should be right-handed. If it is left-handed, the image can be reoriented to a right-handed orientation with setting this parameter to True, which flips the last axis direction.
• compression=False: bool
  Whether to use compression in itk writer. Using a '.nii.gz' suffix in filename also compresses the image.

save_dir

medio.save_dir(dirname, np_image, metadata, use_original_ornt=True, dtype=None, channels_axis=None, parents=False, allow_dcm_reorient=False, **kwargs)

Save a 3d numpy array np_image as a dicom series of 2d slices in the directory dirname (itk backend).

• dirname: path-like
  The directory to save the files in (str or pathlib.Path). If it exists - must be empty.

The other parameters: np_image, metadata, use_original_ornt, dtype, channels_axis, parents and allow_dcm_reorient are equivalent to those used in save_img.

Additional optional parameters (**kwargs):

• pattern='IM{}.dcm': str
  Pattern for the filenames to save, including a placeholder ('{}') for the slice number.
• metadata_dict=None: dict or None
  Dictionary of metadata for adding tags or overriding the default values. For example, metadata_dict={'0008|0060': 'US'} will override the default 'CT' modality and set it to 'US' (ultrasound).

Metadata Objects

Affine

medio.Affine

The affine of an image is a transformation between the index space of the array to the physical 3d space. The Affine class is a subclass of numpy.ndarray with some special properties (attributes): spacing, origin and direction which can be accessed and set. The method index2coord maps the indices to the physical space, clone clones the affine.

This class includes also some static methods for affine construction from its components (spacing, origin and direction) and also the inverse methods for getting the spacing, origin and direction matrix from a general affine matrix.

For a mathematical explanation about the affine matrix see NiBabel's affine documentation.

Some usage examples:

>>> import numpy as np
>>> from medio import Affine
>>> affine1 = Affine(np.eye(4))
>>> affine2 = Affine(direction=np.eye(3), spacing=[0.33, 1, 0.33], origin=[-90.3, 10, 1.44])
>>> index = [4, 0, 9]
>>> coord = affine2.index2coord(index)
>>> print(coord)
[-88.98  10.     4.41]

MetaData

medio.MetaData

Together with the image's numpy array, the MetaData object is a necessary component for the I/O functions.

A MetaData object 'metadata' is mainly comprised of:

• metadata.affine: the affine (of class Affine)
• metadata.coord_sys: coordinate system ('itk' or 'nib')
• metadata.orig_ornt: the original orientation of the image (used for saving)

Other properties of the metadata are derived from the affine:

• metadata.spacing: voxels spacing (a reference to metadata.affine.spacing)
• metadata.ornt: the current image orientation (also depends on the coordinate system)

All these properties can be viewed easily in the console:

>>> import medio
>>> array, metadata = medio.read_img('avg152T1_LR_nifti.nii.gz')
>>> print(metadata)
Affine:
[[  -2.    0.    0.   90.]
 [   0.    2.    0. -126.]
 [   0.    0.    2.  -72.]
 [   0.    0.    0.    1.]]
Spacing: [2. 2. 2.]
Coordinate system: nib
Orientation: LAS
Original orientation: LAS

The MetaData method metadata.is_right_handed_ornt() checks for a right handed orientation according to the determinant of the direction matrix (metadata.affine.direction). This method can be useful before saving a dicom file or series, which should have a right-handed orientation. The method clone clones the metadata object, convert converts the metadata inplace to the given coordinate system.

Orientation

The orientation of a 3d image is string of length 3 which is derived from its affine and coordinate system (the convention). It denotes along which physical axis we move when we increase a single index out of i, j, k in the expression np_image[i, j, k].

For example, 'RAS' orientation in itk:

• Right to left, Anterior to posterior, Superior to inferior

'RAS' in nib - also 'RAS+':

• left to Right, posterior to Anterior, inferior to Superior

Note that the conventions are opposite. Therefore, for stability reasons we use only itk convention in read_img's argument desired_ornt.

For further discussion see NiBabel's image orientation documentation.

Array and Metadata Operations

Some operations on an image affect also its metadata, for example resizing, rotations and cropping.

The class MedImg (medio.medimg.medimg.MedImg) holds an image array with its metadata, and supports some of these operations through the indexing syntax:

>>> from medio.medimg.medimg import MedImg
>>> mimg = MedImg(np_image, metadata)
>>> new_mimg = mimg[:, 4:-4, ::3]
>>> print(new_mimg.metadata)

Ellipsis ('...') syntax is also supported. This indexing allows cropping and basic down-sampling, along with correct metadata update.