Distortion-Free Registration of Whole-Slide Images via Tile-Based Refinement

Registering Hematoxylin and Eosin (H&E)-stained tissue images with Immunohistochemistry (IHC)-stained counterparts is a critical step for accurate pathological interpretation. Existing registration methods in digital pathology have primarily focused on aligning corresponding cells at the same pixel coordinates. While this approach achieves low quantitative registration error, nonlinear transformations often distort cellular morphology and tissue architecture, limiting the applicability of such methods in clinical practice. To address this issue, we propose a tile-based local registration method that minimizes image distortion. Our approach relies solely on translation, rotation, and uniform scaling, thereby preserving structural integrity during registration. Experimental results on the ACROBAT 2023 benchmark datasets demonstrate that the proposed method achieves lower registration errors compared to conventional rigid registration algorithms, with a 34.99% reduction in error.