Skip to content

Release notes

AUDIT v0.2.0 — MRI/US and format-aware I/O

📅 Release date: July 31, 2026

pyarrow is pinned to 19.0.0 because the Streamlit 1.41.1 dataframe serialization path can crash with newer incompatible releases in a fresh installation.

This release separates two concepts that were previously coupled:

  • modality identifies the clinical image type (MRI or US) and selects modality-specific preprocessing and feature extraction.
  • formats identifies how images, segmentations, and predictions are stored and selects the corresponding reader.

New functionality

  • A reader registry supports NIfTI (.nii, .nii.gz), NRRD (.nrrd), MetaImage (.mha, .mhd), raster images (.png, .jpg, .jpeg, .tif, .tiff, .bmp), NumPy arrays (.npy, .npz), DICOM series directories, and DICOM SEG files.
  • Image arrays and physical spacing now travel together through the feature and metric pipelines. SimpleITK spacing is converted to NumPy axis order.
  • MRI and US can independently use any compatible reader. A run still has one global modality, label map, and format configuration.
  • Feature extraction can pass an optional preprocessing mapping to the selected modality; US supports min-max normalization in this release.
  • US now computes the configured 2D spatial features instead of silently ignoring spatial: true.
  • The app's feature catalogue and segmentation error matrix are modality-aware; the error matrix no longer assumes NIfTI input.
  • AUDIT's native metric names are now explicit and community-readable: dice, jaccard, accuracy, precision, sensitivity, specificity, hausdorff_distance, and predicted_size.
  • pymia and Metrics Reloaded retain their backend-specific configuration keys and expose their segmentation-oriented catalogs: 30 metrics each, respectively.

Corrections

  • The custom Hausdorff metric now calculates the standard symmetric distance and applies physical spacing.
  • Empty-prediction sensitivity, precision, and predicted-size metrics use consistent zero semantics; undefined Hausdorff values are represented as NaN rather than backend-dependent infinity.
  • Metrics Reloaded now supports every binary-mask metric documented by AUDIT, including wck and youden_ind; its avdr calculation no longer suffers unsigned-integer underflow.
  • pymia 0.3.2 distance metrics work with empty masks under NumPy 2.x, and undefined non-finite results are exported as NaN.
  • Raster RGB input is converted to luminance rather than silently selecting the red channel.
  • Missing or shape-incompatible masks now raise explicit errors instead of producing an all-zero error matrix.
  • US tumor location is calculated from the 2D image centre and is no longer left undefined when spatial features are enabled.

Compatibility and known boundaries

  • Legacy YAML files without formats continue to load with a warning: MRI defaults to .nii.gz and US uses extension discovery.
  • Metric aliases used by earlier AUDIT configurations (jacc, accu, prec, sens, spec, haus, and size) remain readable for migration purposes. New AUDIT CSV files use the explicit canonical names; third-party backend CSV files retain their native keys.
  • DICOM series and DICOM SEG readers are available through SimpleITK/GDCM, but this release has no anonymized real DICOM fixtures. They require dataset-level validation before being considered production-verified.
  • Raster label masks must be grayscale or palette-indexed. Colour-to-label mapping for RGB masks is not included.
  • Ground truth and predictions must already be registered on the same grid. Shape and spacing are validated; automatic origin/direction validation and resampling are not included.
  • Mixed MRI and US datasets in the same run, dataset-specific label maps, and dataset-specific modalities remain intentionally out of scope.

AUDIT v0.1.2 — Multiple metric backends

📅 Release date: March 14, 2026

This release refactored metric extraction so the computation engine could be selected without changing the rest of the AUDIT workflow.

New functionality

  • Added audit, pymia, and metricsreloaded metric backends.
  • Added backend selection to metric_extraction.yml.
  • Extended the supported overlap, distance, classification, and volume metrics through pymia and Metrics Reloaded.
  • Added model-volume support to longitudinal and performance analyses.

Architecture and documentation

  • Replaced the previous monolithic metric pipeline with backend-specific modules and shared loading/processing utilities.
  • Added API documentation for the metric pipeline, pymia, and Metrics Reloaded.
  • Added the expanded getting-started workflow and updated documentation navigation.
  • Applied project-wide Black and isort formatting before tagging the release.

AUDIT v0.1.0

📅 Release date: July 10, 2025

We’re excited to announce the first public release of AUDIT, a lightweight and interactive Python tool for evaluating medical image segmentation models, especially on MRI datasets.
This version focuses on enabling fast exploration and quality control with minimal setup, bridging model developers and researchers.

📬 Get Involved

Thank you for trying AUDIT — let’s improve medical segmentation evaluation together!

🙌 Acknowledgements

Built on top of an amazing open-source ecosystem: