Fourier-Net: Fast Image Registration with Band-Limited Deformation

Unsupervised image registration commonly adopts U-Net style networks to predict dense displacement fields in the full-resolution spatial domain. For high-resolution volumetric data, this process is however resource-intensive and time-consuming. To tackle problem, we propose Fourier-Net, replacing expansive path a network with parameter-free model-driven decoder. Specifically, instead of our Fourier-Net learning output field domain, learn its low-dimensional representation band-limited Fourier This then decoded by devised decoder (consisting zero padding layer an inverse discrete transform layer) dense, These changes allow unsupervised contain fewer parameters computational operations, resulting faster inference speeds. evaluated on two public 3D brain datasets against various state-of-the-art approaches. example, when compared recent transformer-based method, named TransMorph, which only uses 2.2% 6.66% multiply-add achieves 0.5% higher Dice score 11.48 times speed. Code available at https://github.com/xi-jia/Fourier-Net.