A characterised image-analysis workflow for bright-field Gram-stained smears: code, measurements and manual reference counts

Deposit accompanying a manuscript submitted to Microscopy Research and Technique. It contains an image-analysis pipeline for bright-field Gram-stained smears — Omnipose bact_phase_affinity segmentation, 90 quantitative descriptors per image, and a five-criterion per-cell Gram-phenotype classifier — together with the measurements, the manual reference counts used to validate it, and the supplementary tables. The pipeline is characterised rather than only demonstrated. Segmentation parameters were selected by two criteria independent of any biological hypothesis, which reach their optimum at the same value. A trained observer, blind to the automated output, counted and classified 13,642 cells on 24 images; automated and manual counts agree with r = 0.996 for S. aureus and r = 0.903 for E. coli, and Gram-phenotype proportions to within 1.8 percentage points across a manual range of 0–98%. Measurement precision is quantified for every descriptor from the five fields imaged per slide. The dataset comprises 363 micrographs across 72 slides: two species × three storage temperatures × four durations × three biological replicates × five fields. A second experiment of 291 images with optical density and viable counts is included. Analyses should use slide means: the five fields of a slide are the same sample on the same preparation, and treating them as independent observations changes the conclusions materially. No effect of storage temperature survives analysis at the level of the biological replicate with control of the false discovery rate; effects reported in earlier work on this dataset are withdrawn, and a power analysis shows four of five would require 8 to 54 biological replicates against the three used. Configuration 2026-09e, parameter fingerprint 312d54e8aab9, written into every output row. See README.md for the file list, the reproduction commands and the known limitations.