EdgeReg: Edge-assisted Unsupervised Medical Image Registration

Medical image registration (MIR) is essential for various clinical diagnoses and treatments. Despite the rapid progress in deep learning-based MIR techniques, most methods focus on directly optimizing the raw image intensity information. In this paper, we explore the usage of edge information of anatomical structures associated with the spatial location of image intensities to assist in registration, termed EdgeReg. The intuition is that the edge information can provide additional rich boundary information to the raw images, enhancing the network's feature representation. Additionally, as the edge images are strictly spatially consistent with the raw images, additional supervised information can be added to network training. Specifically, we first extract the edge images from the raw moving and fixed images using the Sobel operator and feed these images into a lightweight feature extractor to merge the image intensity and edge information. The enriched features are subsequently input into established registration networks. Finally, similarity loss is applied to both the raw and edge images. Extensive experiments show that EdgeReg is compatible with various networks across diverse datasets and dimensions (2D and 3D), achieving superior registration performance. In particular, EdgeReg does not rely on segmentation labels and is trained in an unsupervised paradigm. Therefore, edge information is a beneficial assistance for unsupervised MIR. The code is available at https://github.com/PerceptionComputingLab/EdgeReg.