Fourier-Net and Fourier-Net+

Overview

Fourier-Net predicts a band-limited deformation in the Fourier domain and recovers the full-resolution displacement by zero-padding, which makes it very light and fast. Fourier-Net+ additionally works on low-frequency inputs. Both are ported from xi-jia/Fourier-Net with the output converted to MIR voxel displacements.

Variants

  • FourierNet, FourierNetPlus: base models; SVF=True integrates the field.

  • FourierNetSPR, FourierNetPlusSPR: spatially varying regularization heads, TransMorphTVFSPR return convention.

  • FourierNetBackbone, FourierNetPlusBackbone: bare backbones.

Default configuration

FourierNet: start_channel = 8, NCC_mok, smth_labda = 1.0, lr = 1e-4. FourierNetPlus: start_channel = 16, low_img_size = (80, 96, 112), smth_labda = 5.0. Set config.img_size and, for Fourier-Net+, config.low_img_size to half of it.

Usage

from MIR.models import FourierNet, FourierNetPlus, FourierNetPlusSPR
from MIR.models.configs_FourierNet import (get_FourierNet_default_config,
                                           get_FourierNetPlus_default_config, get_FourierNetPlus_SPR_config)

model = FourierNet(get_FourierNet_default_config())
flow = model((mov, fix))

plus = FourierNetPlus(get_FourierNetPlus_default_config(), gen_output=True)
warped, flow = plus((mov, fix))

spr = FourierNetPlusSPR(get_FourierNetPlus_SPR_config())
pos_flow, neg_flow, wts = spr((mov, fix))

Key modules

  • MIR.models.FourierNet, MIR.models.configs_FourierNet

References

  • Jia, Bartlett, Zhang, Lu, Qiu, Duan. Fourier-Net: fast image registration with band-limited deformation. AAAI 2023.

  • Jia, Lu, Qiu, Duan. Fourier-Net+: leveraging band-limited representation for efficient 3D medical image registration. 2023.