Data loaders: prefer nuclei_seg, fall back to nuclear_seg

[mark-a-potts]

Summary

Updates the two cell-extraction data paths to try the new native-20x nuclei_seg label first (produced by submit_nuclei_segmentation_jobs in ops_process PR #113), with fall-through to the legacy 5x-upscaled nuclear_seg from segment_and_stitch_pheno.

Files touched (each is a 9-line addition, identical pattern):

• src/ops_model/data/data_loader.py:466 — CellProfileDataset.__getitem__
• src/ops_model/features/cp_extraction.py:1019 — bulk CP feature read

Both labels are 20x-shaped at level 0 in phenotyping_v3.zarr (the legacy step segmented at 5x and then 4× nearest-neighbor upscaled to 20x for storage), so bbox slicing is unchanged. The existing min_h = min(...) clip block at data_loader.py:498 handles the residual ~3 px shape diff between native-20x and 5x-upscaled-to-20x masks.

Test plan / measured impact

Per-cell A/B comparison over 500 sampled cells from A1_linked_pheno_iss_cp.csv (ops0094_20251217):

metric	legacy mean	new mean	Δ
nuc area (px)	2179.6	2140.4	−1.80%
nuc/cell ratio	0.200	0.197	−1.78%

Per-cell nuclear-mask IoU (NEW vs LEGACY, within the cell mask):

• mean: 0.845, median: 0.856
• IoU < 0.5: 0.8% of cells (large disagreement — over/under-seg cases)
• IoU < 0.7: 2.6% of cells
• IoU ≥ 0.9: 16.6% of cells (near-identical)

Migration semantics

• Per-experiment migration: experiments that have only run the legacy step continue to work; experiments that have run the new step transparently pick up the better masks.
• No re-run of legacy experiments needed.
• Re-training of models that consume these masks is optional but recommended once a critical mass of experiments has been re-segmented (a model trained on legacy masks will see a small distribution shift inferring on new masks: ~1.8% mean feature shift, ~0.8% of cells with meaningfully different features).

Cross-PR

Pairs with: royerlab/ops_process#113

🤖 Generated with Claude Code