v0.2.5

What's Changed

• Update README to remove Details link by @guillaumejaume in #190
• feat: h0-mini, openmidnight, gpfm supported by @guillaumejaume in #191
• Add GrandQC citation and license notice by @Superzlw in #192
• Add GenBio-PathFM patch encoder (ViT-Giant, 4608-dim, JEDI training)#196 by @guillaumejaume in #197
• Gja/speed/patching by @guillaumejaume in #199
• New OME ZARR Image Reader by @guillaumejaume in #200
• feat: option to dump patches when patching coords by @guillaumejaume in #201

New Contributors

• @Superzlw made their first contribution in #192

Full Changelog: v0.2.4...v0.2.5

v0.2.4

• Fixed compound-extension filename handling (.ome.tif / .ome.tiff) using splitext() and applied it across WSI, Processor, and Converter where artifact stems are derived.
• Added robust WSI lifecycle management with context-manager support (enter/exit), updated usage in run_single_slide.py, Processor (ExitStack), and related tests.
• Added a new third segmenter, otsu (image-processing only), implemented in trident/segmentation_models/otsu.py and integrated via segmentation_model_factory, CLI options, and module exports.
• Set Otsu segmentation default magnification to 1.25x and optimized its runtime path (lower allocation overhead + CPU execution path for Otsu in CLI flows).
• Added unit tests for Otsu (tests/test_otsu_segmenter.py) plus earlier lifecycle/extension tests (test_wsi_core_behaviors.py, test_processor_lifecycle.py).
• Improved dependency and install UX in pyproject.toml (including scikit-image for Otsu), plus expanded docs in README.md and DETAILS.md.
• Expanded trident-doctor (including convert profile checks and runtime library checks) and improved output formatting (FAIL/WARN/PASS, summary, JSON mode).
• Added/kept unified CLI wrapper behavior (trident batch/single/convert/doctor) while preserving legacy entrypoints.

v0.2.3

• Bug fix: When patching with overlap, the first patch was not used for overlapping. Now fixed using modified _colrow_to_xy function.

• Speed: GigaPath and Midnight12k now run by default on float16 for increased speed.

• Bug fix: Allow ResNet50-IN model weights to be restored from either the pytorch_model.bin or the model.safetensors

• Clear tissue mask (gdf_contours) after segmentation. Restored with lazy initialization. Avoids steady increase of CPU memory usage when processing 4k+ WSIs.

v0.2.2

• bug fix, default argument for segment_tissue throws an exception

• add num_workers arg to segment_tissue

• bug fix, display mpp for patcher vis if magnification is None

• don t install heavy libraries by default

• add ability to create WSIPatcher from CLAM type coordinates

• fix from_legacy_coords

• add coords_only arg to from_legacy

• add pil argument to from_legacy_coords

• pass slide_path as positional argument in CuCIMWSI, ImageWSI + fix src_pixel_size exception WSIPatcher

• add semantic segmentation inference

• add semantic segmentation inference

• test

• add forward_kwargs and create save_coords_h5

• cleanup before PR

• fix parameter propagation from ImageWSI to WSI

• cleanup before PR

• add README newline

• remove forward_kwargs

• remove get_channels

• fix: set correct wsi path in processor unit test

v0.2.1

• Add Feather slide encoder (ICLM'25)
• Set huggingface mirror
• Fix bug in tutorial 3
• Expose segmentation arguments in run single slide

v0.2.0

• Search nested to process WSIs located in nested dirs (--search_nested)

• New caching mechanism to process very large batches of WSIs. When processing 100k+ WSIs, they are often stored on slow HDDs. To address this, use local caching on a fast SSD (--wsi_cache ./cache --cache_batch_size 32). Data will be copied automatically in batches, and processed in parallel. Automatic cleanup.

• New release() function in the Processor and WSI to ensure GPU memory is released.

• Version automatically inferred from the the toml file.

v0.1.1

Add readthedoc documentation

v0.1.0

• Added native support for CuCIM and PIL.

• All segmentation models now use Hugging Face and support local checkpoints (no more torch.hub).

• All patch models migrated to Hugging Face with local support (except Hibou and MUSK).

• Improved error handling and enhanced docstrings.

v0.0.5

New AnyToTiffConverter. Handy to convert non openslide readable files into pyramidal tiff.

v0.0.4: Merge pull request #2 from mahmoodlab/more_models

Support for 7 additional patch encoders:

• hibou_l
• kaiko-vits8
• kaiko-vits16
• kaiko-vitb8
• kaiko-vitb16
• kaiko-vitl14
• lunit-vits8

Support for 1 additional slide encoder:

• madeleine (based on conch_v1 patch encoder with 256x256 patches at 10x)