Release Notes

1.0.2 [2026-08-21]

Fixes:

Fixes:

Fixed SSRF vulnerability in download_model by restricting URL schemes to http and https only.

Maintenance:

Enhancements:

Upgraded from nnunetv1 to nnunetv2 for inference, enabling the use of the latest nnU-Net architecture (nnUNetResEncUNetMPlans).
Bundled ten pre-trained weights (Medical Segmentation Decathlon datasets Dataset001_BrainTumour through Dataset010_Colon) directly into the gear, eliminating the need to download models at runtime.
Added pretrained_models/hierarchy_curator_training.py script for training nnunetv2 models on Flywheel using the Hierarchy Curator.
Added get_pretrained_model() utility to install bundled pre-trained model zip archives at inference time.

Fixes:

Fixed process_file_name() to cast input_path to str before applying the regex matcher, preventing a TypeError on Path objects.
Fixed nested else/if in split_input_as_needed() to use elif, removing unreachable code path.
Copied input files to the work directory before processing to work around Flywheel's restriction on modifying the input directory.

Maintenance:

Migrated from flywheel-gear-toolkit / GearToolkitContext to fw-gear / GearContext.
Migrated from poetry to uv for dependency management and build system (hatchling).
Upgraded base image from flywheel/nnunetv1_base:0.2.0 to flywheel/python:3.12-wolfi-build with a multi-stage Dockerfile.
Upgraded python runtime from 3.11 to 3.12.
Renamed package from fw_gear_nnunetv1 to fw_gear_nnunet throughout.
Updated CI configuration to use a larger runner with a 4h timeout for gear tests.
Updated tests to use fw-gear mocks and added @pytest.mark.skipif guard for nnunetv2-dependent tests.
Added LICENSE and NOTICES files.

Documentation:

Rewrote README.md to document nnunetv2 models, updated dataset list, workflow steps, and license references.
Updated CONTRIBUTING.md to replace poetry instructions with uv equivalents and expanded pre-commit hook descriptions.

Breaking Changes:

Renamed gear from nnunetv1 to nnunet; any workflows referencing the gear by name will need to be updated.
Renamed Python package from fw_gear_nnunetv1 to fw_gear_nnunet.

Enhancements:

Added support for up to 4 input modalities (modality_0 through modality_3), enabling multi-modal inference workflows.
Expanded pretrained_model options 1 to 10, covering Tasks 001–010 (brain tumour, heart, liver, hippocampus, prostate, lung, pancreas, hepatic vessel, spleen, and colon).
Added automatic 4D NIfTI input splitting into separate 3D modality volumes via new split_input_as_needed() utility.
Added automatic input filename correction to comply with .*_000[0-9].nii.gz format required by nnUNet_predict.
Added model download functionality via new download_model() utility, fetching pre-trained models from Zenodo at runtime.
Added input validation to ensure the number of provided modalities matches the selected model's expected count before inference.
Added dynamic model_url to QC metadata output, replacing a hardcoded Task004_Hippocampus URL.

Fixes:

Fixed nnUNet_predict subprocess call to use subprocess.check_output() with error handling, replacing silent subprocess.run() that did not detect failures.
Fixed parse_config to skip modality inputs that are not provided (i.e., when context.get_input_path() returns None), preventing None values from entering the modality inputs dictionary.

Maintenance:

Updated base Docker image from flywheel/nnunetv1_base:0.1.0 to flywheel/nnunetv1_base:0.2.0.
Extracted input preparation and filename correction logic from main.py into new fw_gear_nnunetv1/utils.py module.
Refactored run() in main.py to accept a pre-prepared input_folder instead of raw modality_inputs and work_folder arguments.
Updated run.py to orchestrate input preparation and model download before calling run().
Updated and expanded test coverage in tests/test_main.py and tests/test_parser.py; added new tests/test_utils.py with comprehensive tests for utility functions.
Updated pyproject.toml license field to reflect dual-license (Apache 2.0 for nnUNet source, MIT for the remainder of the gear).

Documentation:

Updated README.md to document all 4 input modalities, 10 pretrained model options, debug config option, and improved workflow description.
Updated README.md citations to include two additional references for pretrained models and the Medical Segmentation Decathlon dataset.
Added license section to README.md clarifying dual-license terms.
Updated manifest.json citation field to include all three references.

Enhancements:

Added nnUNet v1 inference gear supporting Task004_Hippocampus pretrained model on NIfTI input files.
Added automatic filename normalization to comply with nnUNet's .*_0000.nii.gz input format requirement.
Added QC metadata attachment to output files with pass/fail state based on inference exit code. Maintenance:
Added Dockerfile using flywheel/nnunetv1_base:0.1.0 as base image with python 3.11 runtime.
Added CI/CD pipeline configuration via .gitlab-ci.yml.
Added pre-commit hooks for linting and testing with ruff, pytest, hadolint, markdownlint, yamllint, and jsonlint.
Added pyproject.toml with poetry build system, depending on flywheel-gear-toolkit ^0.6 and flywheel-sdk ^18.
Added unit tests for run() and parse_config() functions.

Documentation: