Abstract
Histopathology often serves as the gold standard in medical diagnosis but lacks a spatial three-dimensional axis of information due to the inherent process of two-dimensional tissue slide preparation. Recovering this third dimension would enable correlation with in vivo imaging, improve multimodal data integration, and open new doors for three-dimensional quantitative tissue analysis. This study presents the RAPID framework, which tackles this task of registering whole slide images of whole-mount histopathology specimens into a three-dimensional stack. Our proposed framework leverages a DINOv2-pretrained ViT-L14 foundational model to consistently detect anatomical features, which are used to align a stack of unregistered whole slide images (WSIs) to obtain a three-dimensional reconstruction. RAPID is optimized to work with high-resolution WSIs at 0.25
m/pixel, leading to gigapixel reconstructions that can be used for cell-level downstream tasks. We validate our framework on various external validation sets and find that RAPID obtains an accurate reconstruction in 91.7-93.5% of cases, thereby significantly outperforming a current state-of-the-art reconstruction method. Unlike existing methods limited to serial sections, RAPID handles the sparse sampling intervals (3000-4000
m) routinely used in clinical pathology. We demonstrate its utility for three-dimensional radiology-pathology correlation, enabling direct volumetric comparison between histopathology and pre-operative imaging.