Staining Independent Nonrigid Iterative Registration Method, for Microscopic Samples

Abstract

Using digital microscope scanners, gigapixel-scale images for tissue samples are scanned in a minute, which provides an opportunity for quantitative evaluation at the cellular or gene level. However, to make an accurate diagnosis for clinical or research cases, it is necessary to make serial sections and stain them using different reagents. Since digital scanning and processing are preceded by manual workflows, the orientations between the images are lost. In the absence of adjustment, we cannot compare them to each other, for colocalization or correlation analysis. A registration method is needed that organizes the samples in the same orientation. The proposed method is inspired by the traditional and deep-learning based registration methods (SURF, SIFT, ORB, SuperPoint, SuperGlue) and further developed to manage the tearing, creasing and other deformations between the samples. Based on the validation results, the basic methods give moderate results, however, by utilizing a grid-based approach and by choosing the appropriate number of recursive iterations and resolution, the methods can be improved. The proposed stain-independent, iterative, non-rigid registration method can manage not only tears, creases and deformations, but also correct structural changes between series sections.