Changelog#

Current Snapshot#

This documentation reflects the current codebase state (commit f996af4, 1 October 2026) rather than a published release tag.

Dataset Creation Tools (f996af4)#

  • Documented six-step dataset creation pipeline, with hoefflin2026spatial as a worked example (examples/create_hoefflin2026spatial.py)

  • scripts.convert_dataset_to_omezarr for adding OME-Zarr metadata and channel labels to converted .zarr images

  • Conversion helpers in scripts/dataset_creation_utils/

  • scripts.compute_tissue_masks no longer hard-codes a local checkpoint path

  • --dataset_name accepted by the tiling and statistics scripts, and --tile_size by the tiling script

  • Cell task label encoders may use an index column instead of id

  • Stats-backed standardization no longer returns cached statistics

Features#

  • Dataset classes for H&E, IHC, IMC, CODEX, CyCIF, and MIBI modalities

  • ComposedImagingDataset for multi-modal loading

  • SporaDataset for multi-cohort tissue and tile sampling

  • Optional split filtering in SporaDataset and composed/unimodal loaders

  • Stats-backed multiplex standardization framework

  • Greedy tiling algorithm with adaptive stopping

  • Padded fixed-grid tiling via scripts.compute_tiling --grid

  • Shared tissue masks and cell masks under segmentations/<resolution>/

  • Shared parquet-backed tiling under tiling/<resolution>/<strategy>/

  • Label filtering by metadata columns