Poster #39: William Sullivan

A multi-scanner whole-slide image framework for detection of scanner-induced batch effects in digital pathology data

Lina Takemaru*, BS, MPS,1 Yutong Zhu*, BS, MS,2 William Sullivan*, BS,1 Jonghyun Lee*, PhD,3 Angela Ge,4 Aastha Vatsyayan, PhD,4 Portia A. Kreiger, MD,5 Amy E. Baxter, D.Phil,6 Dokyoon Kim, PhD,3 Derek A. Oldridge, MD, PhD4 1. Graduate Group in Genomics and Computational Biology, Perelman School of Medicine, University of Pennsylvania 2. Graduate Group in Bioengineering, School of Engineering and Applied Science, University of Pennsylvania 3. Department of Biostatistics, Epidemiology and Informatics, Perelman School of Medicine, University of Pennsylvania 4. Department of Pathology and Laboratory Medicine, Perelman School of Medicine, University of Pennsylvania 5. Division of Anatomic Pathology, Children’s Hospital of Philadelphia 6. Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine, University of Pennsylvania

Digital pathology foundation models (PFMs) learn morphological features from whole-slide images (WSIs) and achieve strong performance in diagnosis, prognosis, and biomarker discovery across a range of diseases. However, their robustness to technical variation across institutions is poorly understood. To test this, we evaluated scanner-induced batch effects using 103 pan-normal WSIs scanned on six imaging platforms representing various research and clinical use cases. Feature representations were extracted using 19 PFMs. We quantified scanner signal using logistic regression, assessed stain normalization and batch correction methods, evaluated supervised tissue classification under cross-scanner training conditions, and performed unsupervised clustering and attention-based multiple instance learning (ABMIL). Our results show that scanner identity is strongly encoded in PFM features. As one measure, cosine similarity between embeddings of the same image patch across scanners was highly variable and often poor (range: 0.20–0.96), suggesting substantial scanner-dependent distortions of underlying biological signals. Additionally, logistic regression predicted scanner with >97% accuracy across PFMs. Stain normalization only slightly reduced accuracy, suggesting batch effects extend beyond color variation. In supervised tissue classification, performance dropped by 14.66% ± 2.44% under cross-scanner confounding compared to balanced training, averaged across PFMs. In unsupervised analyses, scanner-driven structure dominated and obscured biological patterns. ABMIL highlighted different tissue regions depending on scanner, and unsupervised clustering often separated identical patches by scanner. Principal component analysis revealed heterogeneous batch structure: some PFMs show linear scanner variation, while others show non-linear effects. Existing batch correction methods including ComBat and Harmony reduced scanner classification accuracy but also degraded biological signal. Our results provide evidence that scanner variation is a primary source of technical batch effects and a major signal learned by PFMs, impacting both supervised and unsupervised analyses. These findings highlight the need to account for scanner variation to improve the robustness of PFMs for research applications and clinical deployment.