Output file formats

Reference for the files written by baysor run. For choosing a bundle, see Outputs.

Conventions

• N — retained molecules after filtering and cropping
• C — final cells; G — retained genes
• Cell names are cell_<1-based-id>; noise / unassigned molecules have cell name 0.
• Bracketed columns below are conditional: z for 3D, cluster fields when clustering runs, NCV colors unless skipped, and confidence / history fields when available.

Legacy bundle

segmentation.csv

One row per retained molecule (N rows). Column order:

[transcript_id,] cell, gene, x, y [,z] [,confidence] [,cluster] [,ncv_color] [,assignment_confidence], is_noise

Column	Meaning
transcript_id	Numeric source transcript ID, present when retained from the input.
cell	Assigned cell name, or 0 for noise.
gene	Gene name.
x, y, z	Molecule coordinates.
confidence	Molecule confidence from the noise model.
cluster	Integer molecule-cluster label.
ncv_color	NCV color as hex RGB, e.g. #4F9BD5.
assignment_confidence	Assignment stability estimated from recent iteration history.
is_noise	Always last; true / false when source transcript IDs are retained, 1 / 0 otherwise.

segmentation_cell_stats.csv

One row per final cell (C rows). Column order:

cell, x, y [,z] [,cluster], n_transcripts, density, elongation, area, avg_confidence [,avg_assignment_confidence] [,max_cluster_frac] [,lifespan]

Column	Meaning
cell	Cell name.
x, y, z	Centroid of assigned molecules.
cluster	Cell-level cluster label.
n_transcripts	Number of assigned molecules.
density	Molecule count / area when area is available.
elongation	Ratio of the two x/y covariance eigenvalues.
area	2D convex-hull area of the cell's molecules.
avg_confidence	Mean molecule confidence.
avg_assignment_confidence	Mean assignment stability.
max_cluster_frac	Fraction of molecules in the most frequent molecule cluster.
lifespan	Consecutive stored iterations in which the component exists; limited by the retained history window.

All columns except cell are numeric; labels and counts are stored as floating-point values. Low avg_assignment_confidence can flag unstable assignments, while low max_cluster_frac can flag mixtures of molecule clusters. These are diagnostics, not automatic cell-quality filters.

segmentation_polygons_2d.json

One joined polygon per assigned cell. For 3D data, molecules are pooled across z. A failed boundary estimate gets a fallback rectangle so that the cell set matches segmentation.csv.

--polygon-format controls the GeoJSON schema:

Value	Schema
FeatureCollection (default)	One feature per cell, with id and properties.cell set to the cell name; geometry.type is Polygon.
GeometryCollection	One polygon geometry per cell with a string cell field.
GeometryCollectionLegacy	Same collection, but cell is an integer (cell_17 → 17) for Xenium Ranger 3.x.
none	No polygon files written.

Coordinates use data units and one closed outer ring. Format names are case-insensitive; unknown names are rejected.

segmentation_polygons_3d.json

Written for 3D runs only: a JSON object keyed by layer (e.g. z_003), with a 2D polygon collection per layer using the schema above. [plotting] max_z_slices (default 10) limits the layer count; stacks with more distinct z values are binned first.

segmentation_counts.loom

Counts in Loom 3.0.0 format:

HDF5 path	Type and shape	Contents
/matrix	float32, (G, C)	Gene × cell counts.
/attrs/LOOM_SPEC_VERSION	UTF-8 string array, length 1	3.0.0.
/row_attrs/Name	UTF-8 string array, length G	Gene names.
/col_attrs/Name	UTF-8 string array, length C	Cell names.
/col_attrs/CellID	float64 array, length C	IDs 1 … C.

segmentation_counts.tsv

Dense tab-separated counts: the first column is gene, followed by one column per cell name, with one row per gene. Shape: G × (1 + C).

segmentation_params.dump.toml

Config values used by the pipeline, with the CLI invocation in a comment on the first line. Keep the original config and command for reproducibility: not every setting is serialized (for example, tol is omitted).

segmentation_log.log

Text log with stage progress, iteration summaries, convergence and save messages.

diagnostic_report.html

Self-contained HTML diagnostics, written only with --plot.

segmentation_plot.html

Self-contained molecule / cell visualization, written only with --plot.

Parquet bundle

molecules.parquet

N rows, with the same column order as segmentation.csv except that transcript_id is not written:

Columns	Type
cell, gene, ncv_color	UTF-8 string
x, y, z, confidence, assignment_confidence	float64
cluster	int32
is_noise	boolean

The same conditional-column rules apply. Use legacy output for Xenium Ranger compatibility.

cells.parquet

C rows, with the same columns and order as segmentation_cell_stats.csv: cell is UTF-8 string; all statistics are float64.

cell_boundaries.parquet

One row per joined 2D cell polygon (pooled across z for 3D data):

Column	Type	Contents
cell	UTF-8 string	Cell name.
n_vertices	int32	Vertices, excluding the repeated closing vertex.
geometry	binary	WKB polygon.

GeoParquet metadata sets the primary geometry column to geometry, encoding WKB, geometry type Polygon, and no CRS.

cell_boundaries_3d.parquet

3D runs only: one row per (cell, layer) polygon. Column order is cell, layer, n_vertices, geometry; layer is UTF-8 string. Other types and GeoParquet metadata match cell_boundaries.parquet.

feature_matrix.h5

10x-style sparse CSC matrix, genes × cells (nnz = nonzero entries):

HDF5 path	Type and shape	Contents
/matrix/barcodes	UTF-8 strings, length C	Cell names.
/matrix/data	int32, length nnz	Counts.
/matrix/indices	int32, length nnz	Gene-row indices.
/matrix/indptr	int64, length C + 1	Column pointers.
/matrix/shape	int64, length 2	[G, C].
/matrix/features/id, /matrix/features/name	UTF-8 strings, length G	Gene names in both arrays.
/matrix/features/feature_type	UTF-8 strings, length G	Gene Expression.
/matrix/features/genome	UTF-8 strings, length G	Empty strings.

run_params.toml

Same config dump and limitations as segmentation_params.dump.toml.

run.log

Text log for Parquet output.

diagnostic_report.html / segmentation_plot.html

Same reports as in legacy output; written only with --plot.