Impact of Pre-processing and Local Feature Extraction on Feature-Based Registration of Whole-Slide Images

Feature-based registration has gained increasing popularity in digital pathology as a means of achieving initial global, low-resolution alignment between image pairs. Despite its widespread adoption, the specific design choices within registration pipelines are often insufficiently justified. This study presents a comprehensive benchmarking analysis on consecutive multi-stained whole-slide images to evaluate the performance of various pre-processing and local feature extraction methods. Both traditional and deep learning-based techniques are assessed to determine whether the latter consistently outperform the former. The findings underscore the critical importance of both pre-processing and feature description steps in influencing overall alignment quality. Notably, Grayscale conversion consistently surpasses Hematoxylin deconvolution as a pre-processing approach. While detector selection has a relatively minor impact on performance, descriptor choice plays a crucial role. Among the most robust descriptors identified are two deep learning-based methods ( SuperPoint and DISK ), as well as two classical algorithms ( BRIEF and O-BRIEF ), which deliver competitive results with lower computational demands. In more challenging registration scenarios, however, SuperPoint emerges as the most effective descriptor.