Iterative Cross-Scanner Registration for Whole Slide Images

The successful registration of digitized microscopic images is required for many applications in digital pathology. In particular, the registration of specimens scanned by different slide scanning systems may be beneficial to transfer expert annotations from one image domain to another and thereby reduce labeling effort. We present an iterative approach to register microscopic specimens digitized with multiple scanning systems, aiming to compute an optimal global transformation for the images at highest resolution. For this purpose, an initial registration based on a down-scaled version of the images is followed by a patch-based iterative update scheme. We make use of the hierarchical structure of digitized whole slide images to gradually approximate the optimal transformation. By using kernel density estimation to weight local transformation estimates, the influence of registration errors can be further mitigated. We validate our method on five histologic and five cytologic samples, each scanned with four different scanning systems. Furthermore, we perform first experiments on samples stained with different stain combinations. Our experiments demonstrate the potential of the proposed method for a variety of datasets and application fields.