Work

CORE: A coarse-to-fine registration engine for multi-stain whole-slide images

Year: 2026

Type: article

Abstract: Accurate registration of multi-stained whole-slide images (WSIs) enables the integration of complementary morphological and molecular information, improving both clinical diagnostic and prognostic analysis. It also enables efficient transfer of annotations between consecutive or re-stained slides, significantly reducing annotation time and cost, as well as three-dimensional tissue reconstruction. Despite its importance, WSI registration remains challenging due to variations in slide preparation across stains and complex tissue deformations. Existing methods are often tailored to specific staining modalities and fail to generalize across diverse settings, highlighting the need for a robust and generalizable cell-level multi-stain registration method. In this work, we introduce CORE, a novel coarse-to-fine framework for accurate cell-centric registration across six multimodal WSI datasets, encompassing 30 distinct staining types, including Hematoxylin & Eosin (H&E), periodic acid–Schiff (PAS), multiplex immunohistochemistry (mIHC), multiplex immunofluorescence (mIF) and Cyclic immunofluorescence (Cyc-IF). CORE first performs coarse global registration by extracting tissue masks through prompt-based segmentation to remove artifacts and non-tissue regions, followed by rigid alignment using tissue morphology and a pre-trained feature extractor. The resulting alignment is refined using a shape-aware point-set registration model applied to automatically detected nuclei centroids, enabling fine-grained rigid alignment. Finally, Coherent Point Drift (CPD) is used to estimate a non-linear displacement field for non-rigid cellular alignment. Using nuclei correspondences throughout the pipeline, CORE achieves accurate and robust cell-level alignment across modalities. Experimental results show that CORE consistently outperforms state-of-the-art methods in accuracy, robustness, and generalization across both bright-field and immunofluorescence WSIs.

Source: Biomedical Signal Processing and Control

DOI: 10.1016/j.bspc.2026.111292