Multi-magnification networks for deformable image registration on histopathology images

Abstract

We present an end-to-end unsupervised deformable registration approach for high-resolution histopathology images with different stains. Our method comprises two sequential registration networks, where the local affine network can handle small deformations, and the non-rigid network is able to align texture details further. Both networks adopt the multi-magnification structure to improve registration accuracy. We train the proposed networks separately and evaluate them on the dataset provided by the University Hospital Frankfurt, which contains 41 multi-stained histopathology whole-slide images. By comparing with methods using the single-magnification structure, we confirm that the proposed multi-view architecture can significantly improve the performance of the local affine registration algorithm. Moreover, the proposed method achieves high registration accuracy of contents at the cell level and is potentially applicable to other medical image alignment tasks. A PyTorch implementation of the proposed algorithm is available at https://anonymous.4open.science/r/MMDReg-27A2/README.md.

Tasks

Reproductions