NEWS.md

Version 2.3.0

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.

Version 2.2.4

Fixes

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

Version 2.2.3

Minor changes

• Tables with feature values now contain extra columns. These columns specify the file name (for non-DICOM input), the directory path of the image and masks and several DICOM tags for identifying the input.

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

Fixes

• Fixed an error that occurs when attempting to create a deep copy ImageITKFile objects.

Version 2.2.2

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.

Version 2.2.1

Minor changes

• If mask-related parameters are not provided for computing features or processing of images for deep learning, a mask is generated that covers the entire image.

• Add fall-back methods for missing installation of the ray package for parallel processing. This can happen when a python version is not supported by the ray package. ray is now a conditional dependency, until that package is released for python 3.12.

• The default export format for deep_learning_processing and deep_learning_processing_generator is now dict, because the sample name is important for matching against observed outcomes.

• write_file arguments of extract_mask_labels and extract_image_parameters were deprecated as these were redundant.

Fixes

• Streamlined importing and reading DICOM files results in faster processing of DICOM-based imaging.

• Fixed an indexing issue when attempting to split masks into bulk and rim sections in a slice-wise fashion.

• Fixed an indexing issue in Rank's method for noise estimation.

• Fixed incorrectly named image parameters file export. Instead of mask_labels.csv, image parameters are now correctly exported to image_metadata.csv.

Version 2.2.0

Major changes

• Added support for intensity scaling using the intensity_scaling parameter. Intensity scaling multiplies intensities by a scalar value. Intensity scaling occurs after intensity normalisation (if any) and prior to adding noise (if any). For example, intensity scaling can be used after intensity normalisation to scale intensities to a different range. intensity_normalisation = "range" with intensity_scaling = 1000.0 maps image intensities to [1000.0, 0.0] instead of [1.0, 0.0].

• Added support for intensity transformation filters: square root ("pyradiomics_square_root"), square ("pyradiomics_square"), logarithm ("pyradiomics_logarithm") and exponential ("pyradiomics_exponential"). These implementations are based on the definitions in the pyradiomics documentation. Since these filters do not currently have an IBSI reference standard, these are mostly intended for reproducing and validating radiomics models based on features extracted from pyradiomics.

• Modules were renamed according to the PEP8 standard. This does not affect the documented public interface, but may affect external extensions. Public and private parts of the API are now indicated.

Minor changes

• Added support for Python version 3.10 using typing-extensions.
• Several changes were made to ensure proper functioning of MIRP with future versions of pandas.
• Some changes were made prevent deprecation warnings in future version of numpy.

Version 2.1.1

Fixes

• Fixed missing merge changes from version 2.1.0 to the main branch.
• Fixed reading of mask_name from data xml files.
• image_name and mask_name configuration parameters are now parsed as single strings if only one value is specified to match argument-based configuration.
• Fixed and updated several exception messages.
• Filter kernel names, specified using filter_kernels in xml files, are now correctly parsed as strings instead of floats.

Version 2.1.0

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.

Version 2.0.1

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.

Version 2.0.0

Major changes

• MIRP was previously configured using two xml files: config_data.xml for configuring directories, data to be read, etc., and 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.

Version 1.3.0

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.

Version 1.2.0

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.

Version 1.1

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.