GigaPath-RoMa: Pathology-Adapted Dense Correspondence for Gigapixel Whole-Slide Image Registration

Abstract

Registering whole-slide images (WSIs) is essential for integrating complementary information across histological stains, yet remains challenging because of gigapixel resolution, pronounced cross-stain appearance discrepancies, nonlinear deformation, tissue tearing, and partial tissue loss. Although general-purpose registration methods have advanced substantially, they are not tailored to histopathological images and remain difficult to apply directly at gigapixel scale. In this work, we present GigaPath-RoMa, a pathology-adapted dense correspondence framework for WSI registration. We adapt RoMa, a robust dense feature matching model, to the pathology domain through synthetic dense transformations and quality-aware certainty supervision. Furthermore, to scale dense matching to gigapixel WSIs, we introduce a global-to-local strategy that estimates a global correspondence field from a downsampled 4096-pixel WSI view and refines it across increasing WSI resolutions. Experiments on two histological registration datasets demonstrate that GigaPath-RoMa consistently outperforms representative classical and learning-based methods, with particularly strong improvements in robustness and tail registration errors. These results establish pathology-adapted dense correspondence as an effective basis for accurate gigapixel WSI registration.

Keywords