Published August 10, 2026 | Version v1

Foundation-model embeddings as a reproducible readout of Ki-67 from H&E in triple-negative breast cancer: a multi-encoder measurement audit

Authors/Creators

  • 1. Independent Researcher

Description

Background. Ki-67 is a proliferation marker routinely scored on immunohistochemistry (IHC) with substantial inter-observer variability. Predicting Ki-67 directly from haematoxylin-and-eosin (H&E) morphology has been demonstrated, but most existing work reports point performance (R^2, MAE) without characterising the pipeline as a measurement instrument -- i.e. without reproducibility, agreement, precision, and failure-mode analysis. We conduct a measurement audit of a deliberately simple readout -- frozen pathology foundation-model embeddings followed by a linear regressor -- restricted to the triple-negative breast cancer (TNBC) subtype.

Methods. In the BCNB cohort we identified 117 TNBC (PAM50 Basal) cases with pathologist Ki-67 IHC labels. Whole-slide H&E patches were encoded with two frozen foundation models (UNI, Phikon-v2) and mean- or max-pooled per slide. The pooled embedding is read out by an anchor-axis engine (standing-wave anchor, SWA): Ki-67 quantile-slice anchors define axes through embedding space, made non-redundant by Gram-Schmidt orthogonalisation, and a Ridge model maps the anchor coordinates to the Ki-67 fraction. A commodity PCA(50)+Ridge readout is retained as a control. We report a nine-test battery spanning concurrent validity, inter-encoder agreement, precision, discriminant validity, and internal/external generalisation.

Results. The SWA anchor-axis readout achieved concurrent validity R^2=0.515 (95% CI 0.380-0.633; r=0.718) for UNI max-pool, essentially bias-free, and at statistical parity with the PCA+Ridge control. Two independently pretrained encoders agreed with ICC(2,1)=0.945 (95% CI 0.888-0.973). Discrimination of high Ki-67 (>15%) reached AUC=0.960. On an independent second cohort (ACROBAT, n=111), a naive external correlation of r=0.49 fell to r~0.32 after removing a scanner-batch confound and automated-label artefacts -- a real but weak external signal, cautioning that unadjusted external estimates are inflated by batch effects.

Conclusions. A frozen-encoder anchor-axis readout behaves as an essentially unbiased instrument for TNBC Ki-67 within a single cohort and modality, matching a commodity baseline while providing an interpretable coordinate system, and approaches the noise ceiling of visual Ki-67 labelling. We position this as a reproducibility audit, not a state-of-the-art accuracy claim. This is a research measurement and is not intended or validated for clinical use.

Files

submission_EN.pdf

Files (1.3 MB)

Name Size
md5:b4030ce944c80b7574d0612710fe61a8
1.3 MB Preview Download

Additional details

Dates

Updated
2026-08-10