SlideRuler: one-slide calibration for frozen pathology foundation models

A post-hoc calibration technique using intra-slide internal controls to align gigapixel pathology embeddings across scanner hardware without retraining the backbone.

Primary source: arXiv:2610.08944: https://arxiv.org/abs/2610.08944

What’s new: Whole-slide pathology foundation models suffer from scanner-induced domain shift: the same tissue scanned on different instruments produces drifted latent embeddings. Retraining or fine-tuning foundation encoders across every clinical site is rarely feasible. In One-Slide Calibration of Pathology Foundation Models (arXiv:2610.08944, 6 Oct 2026), Hou and Huang introduce SlideRuler, a post-hoc calibration method that uses regions within a single slide as internal controls to correct acquisition shifts.

The framework learns a transfer map from paired rescans. At test time, calibration requires only a single scan: the foundation model remains completely frozen. Evaluated across two encoders and five SCORPION scanners, SlideRuler reduced mean target-to-source embedding distance by 16.3% to 38.5% compared to raw embeddings. A source-anchored variant further reduced source-feature displacement by 47.7% to 83.6% while preserving alignment gains.

Why it matters: Hospital deployments cannot afford site-specific foundation model fine-tuning every time a scanner firmware updates or new hardware arrives. Calibrating the embedding space from a single slide without gradient updates treats the foundation model as a modular component, bringing frozen pathology representations closer to plug-and-play clinical utility.

Caveats: Evaluated in a workshop setting on the SCORPION scanner benchmark; validation across a wider array of clinical slide preparations, staining protocols (H&E vs immunohistochemistry), and rare histologic subtypes is still needed. The transfer map itself requires initial paired rescan data for training.