cs.LGFeb 27, 2026

Histopathology Image Normalization via Latent Manifold Compaction

Authors: Xiaolong Zhang, Jianwei Zhang, Selim Sevim, Emek Demir, Ece Eksi, Xubo Song

Organizations: Knight Cancer Institute, Oregon Health and Science University, Oregon, USA · Brenden-Colson Center for Pancreatic Care, Oregon Health and Science University, Oregon, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Batch effects arising from technical variations in histopathology staining protocols, scanners, and acquisition pipelines pose a persistent challenge for computational pathology, hindering cross-batch generalization and limiting reliable deployment of models across clinical sites. In this work, we introduce Latent Manifold Compaction (LMC), an unsupervised representation learning framework that performs image harmonization by learning batch-invariant embeddings from a single source dataset through explicit compaction of stain-induced latent manifolds. This allows LMC to generalize to target domain data unseen during training. Evaluated on three challenging public and in-house benchmarks, LMC substantially reduces batch-induced separations across multiple datasets and consistently outperforms state-of-the-art normalization methods in downstream cross-batch classification and detection tasks, enabling superior generalization.

Figures & tables

Figure 1: (a) Manifold generated by varying H and E intensities of each image induces a local 2D manifold in latent space. All points on the manifold are compacted to a single point ZnZ^{n} , which is a stain-invariant representation that reduces batch effects while preserving biological structure. (b) LMC applies H&E stain-space augmentations, encodes paired views with a Vision Transformer (ViT) backbone, and enforces batch-invariant representations using a contrastive loss.
Figure 2: Camelyon16 visualization and quantification of batch separation results. (a-d) Qualitative UMAPs. (e) Quantitative batch separations measured by CFD and W2. ◯\bigcirc -Normal, △\triangle -Tumor.
Figure 3: Camelyon16 RAD-trained Model Cross-batch Test (on UNI): ROC and AUC. Our model (solid black line) achieves the best AUC among all compared methods.
ClassesUnnormalizedMacenkoStainFuserLMC
G30.1%57.8%2.5%58.0%
G4-cribriform subtype0.0%3.6%0.0%30.6%
G4-poorly formed subtype99.9 %37.3%79.7%54.3%
G4-fused subtype0.0%0.0%1.4%17.4%
G4-glomeruloid subtype0.0%0.0%0.0%5.6%
Benign0.9%12.6%65.4 %44.3%
Table 1: BR-trained Model Cross-batch Test (on BL) Accuracy (by Class)
MethodHSHXRAverage
Unnormalized0.2940.4430.369
Macenko0.4530.5110.482
StainFuser0.4180.4600.439
LMC0.6000.6510.626
Table 2: MIDOG21 ACS-trained Model Cross-batch Test (on HS & HXR) F1 Scores

Explore similar work

CardsList

  1. Mitigating Batch Effects in Histopathology via Language-Mediated Robust Embedding Generation

    Jun 27, 2026Yishu Zhang, Shushan Wu, Zhenzhong Zhang +8Computational PathologyPathology Foundation Models

    University of North Carolina at Chapel Hill

  2. Magnification-Invariant Image Classification via Domain Generalization and Stable Sparse Embedding Signatures

    Apr 28, 2026Ifeanyi Ezuma, Olusiji MedaiyeseHealthcareCross-Domain Generalization

    Ifeanyi Ezuma, Olusiji Medaiyese

  3. Transformation Behavior of Images in Latent Space

    Jun 23, 2026Christian Zöllner, Mozzam Motiwala, Aysel Ahadova +4Histopathology Image ClassificationInvariant Representation Learning

    Department of Applied Tumor Biology Institute of Pathology Heidelberg University Hospital · Department of Applied Tumor Biology, Institute of Pathology, Heidelberg University Hospital, Im Neuenheimer Feld 224, 69120, Heidelberg, Germany · National Center for Hereditary Tumor Syndromes, University Hospital Bonn, Venusberg-Campus 1, 53127, Bonn, Germany +5