LKGDIR: A Deformable Image Registration Network with Large Kernel Ghost Module
Haonan Zou, Hong Zheng, Xin Sheng Jiang, Bin Jiang, Qingling Xia · 2024
Deformable image registration (DIR) holds significant importance in many fields. However, current deep learning (DL)-based registration methods still face several challenges, such as 1) inadequate extraction of global information of images due to the low receptive field and 2) large number of model parameters and GPU memory usage. This paper proposes a novel unsupervised deformable image registration network based on large kernel ghost module (LKGDIR), to improve the ability to extract global information and reduce the number of model parameters and GPU memory usage. Experimental results on 3D brain dataset show that compared to the three state-of-the-art DL-based models including VoxelMorph, LKU-Net and TransMorph, our model achieves superior results in terms of the dice similarity coefficient (DSC) and GPU memory usage. Additionally, the number of parameters is smaller than those of LKU-Net and TransMorph, demonstrating strong competitiveness in accuracy while featuring lightweight nature. The code of our method is available at https://github.com/NextoNexus/LKGDIR.