Electron Microscopy Image Registration Using Correlation Volume

Large deformation poses a big challenge to current electron microscopy (EM) image registration approaches. In this paper, we address this challenge by proposing an unsupervised deep method with correlation volumes. We first build a feature pyramid for two EM images (the fixed image and the moving image to be registered). The features are extracted by a weight-sharing convolutional encoder and processed with the crossed spatial attention module. We then construct the correlation volume to explore the correlation between enhanced features from the fixed and moving features at each level. We further propose a residual estimator based on the gated recurrent unit (GRU) to update the displacement field. Experimental results on three public EM datasets show that our approach outperforms state-of-the-art EM image registration methods in terms of accuracy and cross-dataset generalization, especially when the deformation is large.