Xenium Fails to read third party segmentation

[annasaax-rgb]

spatialdata_io.xenium() assumes the dataset contains both nucleus and cell segmentation masks as in the standard Xenium structure. E.g. Proseg-imported segmentation only contains a single segmentation mask and therefore cant be loaded properly. Also setting

cells_labels=False
nucleus_boundaries=False

does not solve the issue.

[AlirezaT99]

Confirming this on spatialdata-io==0.7.0, with a full traceback and root-cause pinpoint since the report above notes that cells_labels=False / nucleus_boundaries=False don't help.

A Proseg/Baysor bundle imported via xeniumranger import-segmentation writes only polygon_sets/0 into cells.zarr.zip (no polygon_sets/1 for cells). Loading it fails regardless of the label/boundary flags:

sio.xenium(
    xenium_dir,
    cells_labels=False,
    nucleus_labels=False,
    nucleus_boundaries=False,
    morphology_mip=False,
    morphology_focus=False,
)

FileNotFoundError: polygon_sets/1/cell_index
  ...
The above exception was the direct cause of the following exception:

  File ".../spatialdata_io/readers/xenium.py", line 383, in xenium
    cells_zarr_ctx = _XeniumCells.open(path, version)
  File ".../spatialdata_io/readers/xenium.py", line 165, in open
    return cls(...)
  File "<string>", line 8, in __init__
  File ".../spatialdata_io/readers/xenium.py", line 123, in __post_init__
    self.cell_indices_mapping = self.get_indices_mapping(mask_index=1)
  File ".../spatialdata_io/readers/xenium.py", line 192, in get_indices_mapping
    cell_index = self.group[f"polygon_sets/{mask_index}/cell_index"][...]
  ...
KeyError: 'polygon_sets/1/cell_index'

Root cause

This is a regression introduced by the 0.7.0 reader refactor (#386). In 0.6.0 the polygon_sets/{mask_index}/cell_index read lived inside _get_labels_and_indices_mapping, gated by if cells_labels: / if nucleus_labels: — so passing cells_labels=False, nucleus_labels=False was a working workaround for imported/single-mask segmentations.

In 0.7.0, xenium() opens _XeniumCells unconditionally, and __post_init__ eagerly computes both mappings before any flag is consulted:

def __post_init__(self) -> None:
    self.nucleus_indices_mapping = self.get_indices_mapping(mask_index=0)  # polygon_sets/0 -> OK
    self.cell_indices_mapping    = self.get_indices_mapping(mask_index=1)  # polygon_sets/1 -> KeyError

get_indices_mapping guards the access with if self.cell_id_str is not None and self.has_polygon_sets:, but has_polygon_sets is True here because the parent polygon_sets group exists (polygon_sets/0 is present) — the guard never checks that the specific polygon_sets/{mask_index} subgroup exists. So mask_index=0 succeeds and mask_index=1 raises. Because the read now happens in __post_init__, the cells_labels / nucleus_labels / *_boundaries flags can no longer prevent it.

Suggested fix

Make the has_polygon_sets guard (or get_indices_mapping) check the per-mask_index subgroup, e.g. f"polygon_sets/{mask_index}" in self.group, and return None when it's absent — rather than assuming both mask indices always exist. That both restores the single-mask (imported-segmentation) case and makes the mapping lazy w.r.t. the requested flags. This also addresses #308.

For anyone blocked today: pinning spatialdata-io>=0.6.0,<0.7 restores the cells_labels=False workaround.