A Multi-stage Anti-blur Network for Unsupervised Deformable Image Registration

Su, Yao

Etd

A Multi-stage Anti-blur Network for Unsupervised Deformable Image Registration

Public

Image registration, the task of aligning two or more images into one coordinate system, serves as an essential pre-process for medical diagnosis, researches, and analysis of neuroimaging data. Recently, deep learning-based methods for image registration have attracted much attention, due to their real-time processing speeds. However, previous research mainly focused on improving registration accuracy but rarely discuss sharpness preservation. In this paper, we investigate the problem of anti-blur image registration, where the goal is to achieve high registration accuracy while preserving sharpness. The problem is significant because the analysis on neuroimaging data, e.g., brain network discovery, are critically impacted by the correctness of spatial alignment and the distortion of voxel signals. The task is also highly challenging because the state-of-the-art paradigm based on multi-stage deformation and the improvement on registration unavoidably come with sharpness loss. To address this issue, we propose a novel solution, called Anti-Blur Multi-Stage Registration Network (ABN), for deformable image registration under unsupervised settings. Compared to existing multi-stage methods, which only learn the deformation for the current warped image recursively, ABN learns global deformations at each stage by fusing local deformations, thus allowing it to directly resample from the moving source image and preserve the sharpness. Extensive experiments on natural and medical image datasets demonstrate that ABN can effectively preserve image sharpness. Higher registration accuracy is also observed when comparing to multiple state-of-the-art methods.

Creator