Broaden use of the reader slice/region fast path across reshape and xarray_dask_data call sites

[BrianWhitneyAI]

Summary

BioImage.get_image_data recently gained a slice fast path that defers a sub-region read directly to reader.get_image_data(...) instead of materializing the full array and slicing in memory (see bioio/bio_image.py ~L913–916, from #195). Right now that fast path only fires under a narrow condition. There are several other places in the repo where we either read the full image and then reshape/slice, or interact with xarray_dask_data directly, that could instead fall back to the reader — which often has a faster, format-native way to derive the result.

This issue tracks broadening use of the reader slice/region fast path (and reader fallbacks generally) across these call sites.

Existing fast path

# bioio/bio_image.py — get_image_data
if dimension_order_out is not None and kwargs and not stitching_mosaic:
    return self.reader.get_image_data(dimension_order_out, **kwargs)

# Otherwise read the full image: return it as-is, or reshape/slice it.
if dimension_order_out is None and not kwargs:
    return self.data
return biob.transforms.reshape_data(
    data=self.data,            # <-- full in-memory read, then slice
    given_dims=self.dims.order,
    return_dims=dimension_order_out,
    **kwargs,
)

Target areas

1. get_image_data with slice kwargs but no dimension_order_out (bio_image.py ~L915–926). When a caller passes only slice kwargs (e.g. get_image_data(C=0)), the fast-path condition (dimension_order_out is not None) is not met, so we fall through to a full self.data read + reshape_data. We could default dimension_order_out to the natural order and still route through the reader slice fast path.

2. get_image_dask_data (bio_image.py ~L834). Mirrors get_image_data but builds from xarray_dask_data. Evaluate an analogous reader fallback (reader.get_image_dask_data) so delayed slicing can use a reader-native region path where available.

3. reshape_data on full arr.data in _transform_data_array_to_bioio_image_standard (bio_image.py ~L519–524) and the get_image_data fallback (~L921). Where the reshape is purely a slice/reorder, prefer pushing the selection down to the reader rather than reshaping a fully-materialized array.

4. channel_names (bio_image.py ~L1056). Currently pulls self.xarray_dask_data[Channel].values, which forces task-graph construction just to read channel coordinate labels. Add a reader fallback that derives channel names from metadata.

5. get_xarray_stack / generate_stack (bio_image.py ~L1016; bioio_base/transforms.py generate_stack). Stacking iterates scenes and concatenates .data / xarray_dask_data; check whether per-scene region reads can flow through the reader fast path.

Proposed approach

• Add a reader fallback hook for these operations: when the reader exposes a faster region/metadata path, prefer it; otherwise fall back to the current xarray_dask_data / reshape_data behavior.
• Preserve current behavior for mosaic stitching (_reconstruct_mosaic + MosaicTile), where results must reflect the stitched array, not raw reader output.
• Keep changes behavior-preserving: same returned dims/shape/dtype/values, just cheaper derivation.

Related

• Metadata fast path (dims/shape/dtype): Add fast-path access for metadata properties (dims, shape, dtype) without building full dask task graph #197