v0.7.0

Added

• New algorithm for NCV estimation based on Random Indexing. It is used by default now.
• Support of Parquet format for input molecules (can directly use transcripts.parquet from Xenium)
• Now, in case of continuous z-stack (like Xenium), 3D polygons are estimated by binning the z-stack into 20 slices.
• Added support for string IDs for prior segmentation. To specify the label for non-assigned molecules, use --config.segmentation.unassigned_prior_label.
• New plotting configuration options (see example_config.toml)

Removed

• no-ncv-estimation was removed, as the NCV algorithm is fast and memory-efficient now
• The Dirichlet sampling was removed, as it slowed-down the algorithm significantly, but didn't improve the segmentation quality
  • Correspondingly, parameters new_component_weight and new_component_fraction were removed

Fixed

• Improved multithreading
• Optimized the algorithm performance
• Improved algorithm for polygon estimation: fixed bugs and reduced overlaps
• Fixed a bug in the split step, improved cell continuity

Changed

• Cell IDs in polygon GeoJSON are now strings to match segmentation.csv
• The whole polygon.json format was changed from GeometryColection to FeatureCollection to match 10x format.
• Polygons are now saved by default using the 10x FeatureCollection format. Parameter save-polygons is replaced with polygon-format=FeatureCollection. Set it to GeometryCollection to save polygons in the format from Baysor v0.6 or to none to prevent saving polygons.
• 2D and 3D polygons are now stored in separate files (polygons_2d/3d.json)