Mirp

Major changes

- Added support for SEG DICOM files for segmentation.

- Added support for processing RTDOSE files.

- It is now possible to combine and split masks, and to select the largest mask or mask slice, as part of the image
processing workflow. Masks can be combines by setting `mask_merge = True`, which merges all available masks for an
image into a single mask. This can be useful when, e.g., multiple regions of interest should be assessed as a single
(possibly internally disconnected) mask. Masks are split using `mask_split = True`, which separates every disconnected
region into its own mask that is assessed separately. This is used for splitting multiple lesions inside a single mask
into multiple separate masks. The largest region of interest in each mask is selected by
`mask_select_largest_region = True`. This can be used when, e.g., only the largest lesion of multiple lesions should be
assessed. Sometimes, only the largest slice (i.e. the slice containing most of the voxels in a mask) should be
assessed. This is done using `mask_select_largest_slice = True`. This also forces `by_slice = True`.

These mask operations are implemented in the following order: combination -> splitting -> largest region ->
largest slice.

- Masks from an RT-structure file that shares a frame of reference with an image but does not have a one-to-one
mapping to its voxel space can now be processed. This facilitates processing of masks from RT structure sets that
are, e.g., defined on CT images but applied to co-registered PET imaging, or from one MR sequence to another.

Fixes

- Providing a mask consisting of boolean values in a numpy array no longer incorrectly throws an error.
- Configuration parameters from `xml` files are now processed in the same manner as parameters defined as function
arguments. The same default values are now used, independent of the parameter source. This fixes a known issue where
outlier-based resegmentation would occur by default using `xml` files, whereas the intended default is that no
resegmentation takes place.
- Masks can now be exported to the file system without throwing an error.
- DICOM files from frontal or sagittal view data are now correctly processed.

Minor changes

- Randomisation in MIRP now uses the generator-based methods in `numpy.random`, replacing the legacy functions.
The generator is seeded so that results are reproducible. The seed depends on input image, mask and configuration
parameters, if applicable.

Fixes

- Numpy arrays can now be used as direct input without throwing a `FileNotFoundError`.
- Relaxed check on orientation matrix when importing images, preventing errors when the l2-norm is around 1.000 but not
to high precision.
- To prevent high loads through internal multithreading in `numpy` and other libraries when using `ray` for parallel
processing, each ray thread is now initialised with environment parameters that prevent multi-threading.

Major changes

- MIRP was previously configured using two `xml` files: [`config_data.xml`](mirp/config_data.xml) for configuring
directories, data to be read, etc., and [`config_settings.xml`](mirp/config_settings.xml) for configuring experiments.
While these two files can still be used, MIRP can now be configured directly, without using these files.

- The main functions of MIRP (`mainFunctions.py`) have all been re-implemented.
- `mainFunctions.extract_features` is now `extractFeaturesAndImages.extract_features` (functional form) or
`extractFeaturesAndImages.extract_features_generator` (generator). The replacements allow for both writing
feature values to a directory and returning them as function output.
- `mainFunctions.extract_images_to_nifti` is now `extractFeaturesAndImages.extract_images` (functional form) or
`extractFeaturesAndImages.extract_images_generator` (generator). The replacements allow for both writing
images to a directory (e.g., in NIfTI or numpy format) and returning them as function output.
- `mainFunctions.extract_images_for_deep_learning` has been replaced by
`deepLearningPreprocessing.deep_learning_preprocessing` (functional form) and
`deepLearningPreprocessing.deep_learning_preprocessing_generator` (generator).
- `mainFunctions.get_file_structure_parameters` and `mainFunctions.parse_file_structure` are deprecated, as the
the file import system used in version 2 no longer requires a rigid directory structure.
- `mainFunctions.get_roi_labels` is now `extractMaskLabels.extract_mask_labels`.
- `mainFunctions.get_image_acquisition_parameters` is now `extractImageParameters.extract_image_parameters`.

- MIRP previously relied on `ImageClass` and `RoiClass` objects. These have been completely replaced by `GenericImage`
(and its subclasses, e.g. `CTImage`) and `BaseMask` objects, respectively. New image modalities can be added as
subclass of `GenericImage` in the `mirp.images` submodule.

- File import, e.g. from DICOM or NIfTI files, in was previously implemented in an ad-hoc manner, and required a rigid
directory structure. Now, file import is implemented using an object-oriented approach, and directory structures
are more flexible. File import of new modalities can be implemented as a relevant subclass of `ImageFile`.

- MIRP uses type hinting, and makes use of the `Self` type hint introduced in Python 3.11. MIRP therefore requires
Python 3.11 or later.

Minor changes
- MIRP now uses the `ray` package for parallel processing.

Minor changes
- `SimpleITK` has been removed as a dependency. Handling of non-DICOM imaging is now done through `itk` itself.
- Rotation - as a perturbation or augmentation operation - is now performed as part of the interpolation process.
Previously, rotation was implemented using `scipy.ndimage.rotate`. This, combined with any translation or
interpolation operation would involve two interpolation steps. Aside from removing a computationally intensive
step, this also prevents unnecessary image degradation through the interpolation process. The new implementation
operates using affine matrix transformations.
- Discretisation of intensities after filtering (i.e. intensities of response maps) now uses a *fixed bin number*
method with 16 bins by default. Previously, no default was set, which could lead to unintended results. These
parameters can be manually specified using the `response_map_discretisation_method`,
`response_map_discretisation_bin_width`, and `response_map_discretisation_n_bins` arguments; or alternatively
using the `discretisation_method`, `discretisation_bin_width` and `discretisation_n_bins` parameters of the
`img_transform` section of the settings configuration file.

Fixes

- Fixed a deprecation warning caused by `slic` of the `scikit-image` module.
- Fixed incorrect merging of contours of the same region of interest (ROI) in the same slice. Previously, each contour
was converted to a mask individually, and merged with the segmentation mask using `OR` operations. This functions
perfectly for contours that represent separate objects spatially. However, holes in RTSTRUCT objects are not
always represented by a single contour. They can also be represented by a separate contour (of the same region of
interest) that is contained within a larger contour. For those RTSTRUCT objects, holes would disappear. This has
now been fixed by first collecting all contours of a ROI for each slice, prior to converted them to a segmentation
mask.

Major changes
- Updated filter implementations to the current (August 2022) IBSI 2 guidelines.
- Settings read from the configuration files are now parsed and checked prior to starting computations. This is a
preliminary to command-line configuration of experiments in future versions. Several `xml` tags were renamed or
deprecated. Most renamed tags are soft-deprecated, and support backward compatibility. The following tags will now
throw deprecation warnings:
- `new_non_iso_spacing` has been deprecated. Non-isotropic spacing can be set using the existing `new_spacing`
argument.
- `glcm_merge_method` has been deprecated and merged into `glcm_spatial_method`.
- `glrlm_merge_method` has likewise been deprecated and merged into `glrlm_spatial_method`.
- `log_average` has been deprecated. The same effect can be achieved by giving the
`laplacian_of_gaussian_pooling_method` the value `mean`.

Minor changes
- It is now possible to compute features for multiple images for the same subject and modality.

Fixes
- White-space is properly stripped from the names of regions of interest.
- Several issues related to one-voxel ROI were resolved.
- Computing no features or features that do not require discretisation do no longer prompt providing for a
discretisation method.
- Computing no features from, e.g., the base image no longer generate errors.
- Fixed an issue where rotated masks were not returned correctly.
- A number of other fixes were made to improve stability.

Major changes
- The `extract_images_for_deep_learning` and underlying functions have been reworked.
- The `deep_learning` section of the settings configuration xml file have been deprecated in favour of function
arguments.