A multi-path approach to histology volume reconstruction

Abstract

This paper presents a method for correcting erratic pairwise registrations when reconstructing a volume from 2D histology slices. Due to complex and unpredictable alterations of the content of histology images, a pairwise rigid registration between two adjacent slices may fail systematically. Conversely, a neighbouring registration, which potentially involves one of these two slices, will work. This grounds our approach: using correct spatial correspondences established through neighbouring registrations to account for direct failures. We propose to search the best alignment of every couple of adjacent slices from a finite set of transformations that involve neighbouring slices in a transitive fashion. Using the proposed method, we obtained reconstructed volumes with increased coherence compared to the classical pairwise approach, both in synthetic and real data.