Logo image
Back
1343 Novel Hematoxylin-based Coregistration for Multi-stain Whole Slide Images
Abstract   Peer reviewed

1343 Novel Hematoxylin-based Coregistration for Multi-stain Whole Slide Images

Mathew Francis, Rohan Gajanan Chandraghatgi and Mark Zarella
Laboratory investigation, v 106(3 Supplement), 105637
Mar 2026
DOI: https://doi.org/10.1016/j.labinv.2025.105637

Abstract

Background Digital pathology has enabled the use of advanced computational algorithms for the challenging task of coregistering serial sections to improve visualization, create a 3D volume, and transfer insights from one slide to another. Coregistration algorithms have evolved, but it is still challenging when attempting to register immunohistochemistry (IHC) images to H&E images. Differences in staining patterns can lead to failure of the algorithm. Design We used the common hematoxylin channel present in IHC and H&E rather than brightfield images for coregistration. The IHC and H&E brightfield images were stain separated using non-negative matrix factorization into their constituent stains (Figure 1). Rigid coregistration was performed after a low-pass anti-aliasing filtering, downscaling, and histogram equalization (Figure 1). Rigid registration used a one-plus-one evolutionary optimizer to align the common information between the hematoxylin channels. The algorithm was tested on 920 breast cancer samples. H&E images were coregistered to a complete breast panel that not only included ER, PR, and Ki-67 stained serial sections but also p53 (nuclear), Her2 (membrane), and BCL-2 (cytoplasmic), which enabled us to test differential performance across different staining patterns and expressions. Performance was tested against conventional grayscale-based coregistration methods. A novel coregistration consistency metric was used to evaluate pairwise consistency between all combinations of stains. Results The proposed coregistration using the hematoxylin channel was superior in terms of more precise coregistration and fewer failures compared to the grayscale-based coregistration methods in most images by an average of 18% (Figure 2). Stain separation decreased co-registration errors for both resections and biopsies compared to grayscale images. No single stain was responsible for the errors across all cases; however, in certain instances, one stain contributed significantly to the majority of errors. For example, in Figure 2A, the absence of tissue in the PR image was the primary source of errors. In Figure 2C, the most significant errors arose from failures in non-negative matrix factorization.

Metrics

1 Record Views

Details

Logo image