Minor changes

• Added kurtosis method for local binary pattern filter.

Major changes

• It is now possible to use and process (in-memory) images and masks in a native mirp format. It was already
  possible to export imaging and masks, e.g. using extract_images(..., image_export_format="native") or
  extract_features_and_images(..., image_export_format="native"). Now the resulting images and masks can be used
  as input, e.g. extract_features(image=native_images, masks=native_masks, ...), with native_images and
  native_masks being the resulting images and masks, respectively.

  This allows for external processing of the contents of images and masks, such as performing gamma corrections. The
  image and mask contents are retrieved using the get_voxel_grid method, and set using the set_voxel_grid method.
  set_voxel_grid expects a numpy.ndarray of the same shape and type (float for images, bool for masks) as the
  original.

• Parallel processing is now possible using the joblib backend in addition to ray. This can be specified using the
  parallel_backend argument. Both libraries are now optional, and not installed automatically using pip.

Fixes

• Setting file types is now case-insensitive.
• The co-occurrence matrix-based maximum correlation coefficient no longer has complex values. This was already the
  case, but the return value could still be of a complex type.
• Sample names are now more effectively determined based on file name and folder structure.
• Computing semi-axes length for flat geometries no longer produces occasional warnings due to machine precision.
• Computing morphological features for line-like structures no longer results in divisions by zero.
• Computing morphological features with an empty intensity-mask no longer results in illegal divisions.
• Computing aggregated texture feature values from underlying NaN values no longer generates warnings.
• Features that are not computed because they are not IBSI-compliant are now no longer exported together with valid
  features.
• Fixed a warning caused by a division by 0 when computing the coefficient of dispersion.

Minor changes

• It is now possible to compute local binary patterns as a filter.
• Computation of the co-occurrence matrix-based maximum correlation coefficient now relies less on look-up, at the cost
  of a larger memory footprint. Computation should be more efficient.

Fixes

• It is now possible to merge labelled segmentation masks (e.g. 1, 2) using the settings xml file. Previously,
  this would result in an error when attempting to merge the names of the regions of interest.
• The co-occurrence matrix-based maximum correlation coefficient no longer has complex values.
• Directories with sample names without any further underlying directories (no image_sub_folder or
  mask_sub_folder) are now correctly filtered using sample_name.

Minor changes

• Added the co-occurrence matrix-based maximum correlation coefficient feature. Since no reference standards for this
  feature exist, you need to specify ibsi_compliant = False to compute it.

Fixes

• numpy.trapz was deprecated and is now replaced by numpy.trapezoid.
• pyproject.toml was updated to conform with PEP 639.

Documentation

• The documentation erroneously mentioned that the by_slice parameter could take "largest" as an option. This
  was no longer possible since the introduction of the mask_select_largest_slice parameter in version 2.1.0.

Minor changes

• Improved checks on absent (missing) image transformation parameters where the user is expected to specify one or
  more values. This now provides clear errors.
• The default value of bias_field_correction_n_fitting_levels was changed from 1 to 3, as
  bias_field_correction_n_fitting_levels=1 yielded only very minor improvements to image quality,
  i.e. did not noticeably reduce bias fields in MR.

Fixes

• Background label is now specified as an integer for scikit-image's measure.label function.

Documentation

• Several fixes to the documentation were made.

Fixes

• SUV values with decay correction START are now computed correctly. The previous code, based on the QIBA
  vendor-neutral code, computed the SUV values as if decay correction NONE was used.
• The error message for invalid feature families for filtered images (response maps) now correctly mentions
  response_map_feature_families.
• Setting response_map_feature_families = "all now correctly excludes morphological features.
• Trying to compute local intensity features from imaging with high resolution and low image dimensions no
  longer causes a hard crash due to memory errors produced by scipy.ndimage.convolve. Instead, these errors are
  handled gracefully -- local intensity features are not computed, however.

Major changes

• The proper ancient feature computation code running in the background of MIRP has been completely refactored. We
  moved from a functional backend where all features were computed per feature family to a more flexible
  object-oriented approach. Although this change is not visible at the user-end, it offers several new possibilities:

  • Single features can now be computed. In addition, for some features (e.g. percentile statistics), a flexible
    percentile value could be passed.
  • Creation of feature maps.
  • Output of features and their metadata to machine-readable formats, instead of just tabular data.

  Important: Though the name of features in the tabular exports has not changed, their ordering may have.
  Avoid using column position when processing or analysing feature data.

• Apparent diffusion coefficient (ADC) maps, diffusion contrast-enhanced (DCE) MRI and multi-frame DICOM objects in
  general are now supported.

• Planar imaging, i.e. computed radiography, digital X-ray and digital mammography DICOM files are now supported.

• Python version 3.12 is now supported thanks to recent updates by the maintainers of ray and itk.
  This means that mirp now supports Python version 3.10 and later.

Fixes

• Internal use of numpy.cross no longer produces deprecation warnings.

Fixes

• Masks can now be plotted in images without causing an error when using matplotlib version 3.9.0 or later.

Minor changes

• Tables with feature values now contain extra columns to better identify the input data. For example, the new columns specify the file name (for non-DICOM input), the directory path of the image and masks and several DICOM tags, if available.

• MIRP now checks whether there are potential problems between the frames of reference of image and mask files.

Minor changes

• show method of GenericImage and subclasses now indicate if a user-provided slice_id is out-of-volume and
  select the nearest slice instead.

• Naming of branches in the settings xml file now matches that of their respective settings classes. xml files
  with the previous branch names still function.

• Errors encountered during file import and handling are now more descriptive.

• extract_mask_labels and extract_image_parameters now export extra information from DICOM metadata, e.g. series
  UID.

Documentation

• Added a new tutorial on applying image filters to images.
• Added documentation on the feature naming system.
• Added documentation on the design of MIRP.

Fixes

• Computing features related to the minimum volume enclosing ellipsoid no longer produces warnings due to the use of
  deprecated numpy.matrix class.