Abstract
Histological analysis plays a crucial role in understanding tissue structureand pathology. While recent advancements in registration methods have improved2D histological analysis, they often struggle to preserve critical 3D spatialrelationships, limiting their utility in both clinical and researchapplications. Specifically, constructing accurate 3D models from 2D slicesremains challenging due to tissue deformation, sectioning artifacts,variability in imaging techniques, and inconsistent illumination. Deeplearning-based registration methods have demonstrated improved performance butsuffer from limited generalizability and require large-scale training data. Incontrast, non-deep-learning approaches offer better generalizability but oftencompromise on accuracy. In this study, we introduced ZeroReg3D, a novelzero-shot registration pipeline tailored for accurate 3D reconstruction fromserial histological sections. By combining zero-shot deep learning-basedkeypoint matching with optimization-based affine and non-rigid registrationtechniques, ZeroReg3D effectively addresses critical challenges such as tissuedeformation, sectioning artifacts, staining variability, and inconsistentillumination without requiring retraining or fine-tuning. The code has beenmade publicly available at https://github.com/hrlblab/ZeroReg3D