Eosinophil detection in oesophageal biopsies with a frozen pathology foundation model and a pathologist-adjudicated reference: a single-reader development study on public whole-slide images

We assessed an eosinophil detector on a frozen pathology foundation model and how the reference affected its comparison with its annotator. On public CC0 slides split by patient, one pathologist, also the comparator, marked every eosinophil in 134 model-stratified fields of 128 um from 18 slides (16 patients), blind and then with the overlay, and adjudicated discordances with the source hidden (789 reference eosinophils, 111 don't-care points). A logistic head on frozen Midnight features scores colour and dense-head candidates (leave-one-slide-out). The primary comparison, fixed before the final model was trained, was the detector's best precision at or above the blind first pass's sensitivity minus that read's precision; no margin was set. At threshold 0.9 the detector reached sensitivity 0.762 (95% CI 0.701-0.811) and precision 0.957 (0.941-0.968); the blind first pass 0.748 (0.692-0.809) and 0.972 (0.956-0.984); primary difference -0.011 (-0.054 to +0.010). Under the strict rule the detector was less precise (-0.031, -0.062 to -0.008), as against the pre-final reference (-0.035, -0.084 to -0.003). Re-marking 30 fields blind five days later, the pathologist's sensitivity fell from 0.827 to 0.709, and with this pass as comparator the difference there was +0.008 (-0.052 to +0.040), not -0.051. Against his blind or final marks alone the detector's precision was 0.60-0.85, not 0.95-0.96. A whole-slide audit gave precision 0.848 (0.78-0.92) at score >= 0.9 for countable labels (all detections 0.68-0.88; strict 0.713), 0.31 at 0.5-0.9. No tested remedy showed a detectable gain. The reference and the comparator's reproducibility, more than the model, moved the comparison. The primary difference establishes neither equivalence nor non-inferiority and allows a five-point deficit; the model has no independent patient-level or whole-slide validation, and the clinical endpoint was not evaluated. Deposit contents and scopeRevision r7 (62 pages). This deposit contains the preprint PDF and selected ancillary analysis and figure source code (14 Python scripts) with a README, dependency list, MIT licence, and source-hash inventory. The PDF is unchanged from the verified r7 manuscript. The source archive is a partial, source-only release: it does not include raw whole-slide images, reference/label exports, trained model heads, training code, or the complete detector. Required external inputs are not included, so the archive alone is not a complete runnable reproducibility package. Portability edits were checked statically; the analyses were not rerun for this deposit. This deposit does not freeze or register the planned prospective sealed test. The study is a development study, not independent clinical validation. The preprint text and figures are licensed CC BY 4.0; the selected author code is licensed MIT as specified inside its archive.