When Does Supervision Beat Reconstruction? Magnification Shift and the Robustness of AI Quality Control in Digital Pathology

Abstract Purpose Automated quality control of whole-slide images must stay reliable when scanning conditions change. We compared supervised probes on foundation-model features with generative reconstruction for artifact detection. Methods A linear probe on UNI2-h features and diffusion-reconstruction detectors adapted from PixCell used the same 16 annotated training slides. We evaluated them in cross-validation, on a 24-slide benchmark, on a blind 53-slide multi-cohort benchmark scored by its organizers, and on two external sets (289 cases; 2,127 images). Resampling to half resolution and back mimicked half magnification on the same tissue. We also assessed calibration and monitoring of the probe on known-clean slides. Results At matching magnification the probe was the better detector. It flagged no tissue on clean cross-validation slides, and on the blind benchmark its precision was 0.29 against 0.10 at similar Dice. Under resampling its false positives on clean tissue rose 227-fold on the benchmark and in every organ of the external images, while the reconstruction detectors’ did not; at matched false positives the ranking reversed. Its calibration error rose from 0.05 to 0.30. A second encoder, Phikon-v2, showed a 3.8-fold rise. Training the UNI2-h probe with resampled copies removed most of its rise but lowered accuracy on the original images. Rules combining both detectors inherited the probe’s failure. On known-clean slides the probe’s flagged-tissue fraction separated the 5× external set and resampled images (AUROC 0.94 and 0.89). Conclusions Laboratories using foundation-model probes for quality control should verify acquisition magnification, monitor the probe on clean tissue, and keep a reconstruction detector as a safeguard.