HPCReg-Net: Unsupervised U-Net Integrating Dilated Convolution and Residual Attention for Hippocampus Registration

Abstract

The hippocampus plays a critical role in the human brain, which is mainly responsible for memory and other functions. Accurate and effective hippocampal registration significantly impacts all kinds of hippocampus-related analysis, particularly in the widely-used multi-atlas hippocampus segmentation and the associated clinical decision. However, the existing registration methods suffer from high computational cost and insufficient registration performance. A 3D unsupervised U-Net registration model HPCReg-Net was proposed under a coarse-fine registration strategy in this study, which combined dilated convolution and residual attention module. Specifically, in the coarse registration stage, a new gap-filling mechanism was designed to solve the semantic gap problem in U-Net, and a novel residual attention module was devised to characterize hippocampal deformation. In the fine registration stage, a cascaded dilated convolution module was adopted to capture the considerable deformation and position information of voxels from each pair of images. The experimental results based on ADNI data sets show that our method can improve hippocampal registration greatly in both computational efficiency and registration precision.