Point-Based Registration for Multi-stained Histology Images

Abstract

Image registration is a basic task in biological image processing. Different stained histology images contain different clinical information, which could assist pathologists to diagnose a certain disease. It is necessary to improve the accuracy of image registration. In this paper, we present a robust registration method that consists of three steps: 1) extracting match points; 2) a pre-alignment consisting of a rigid transformation and an affine transformation on the coarse level; 3) an accurate non-rigid registration optimized by the extracted points. The existing methods use the features of the image pair to initial alignment. We proposed a new metric for the non-rigid transformation which adding the part of optimizing extracting points into the original metric. We evaluate our method on the dataset from the ANHIR Registration Challenge and use MrTRE (median relative target registration error) to measure the performance on the training data. The test result illustrates that the presented method is accurate and robust.