LessNet

Overview

LessNet is a light-weight convolutional registration network (0.1 M parameters) ported from xi-jia/LessNet. MIR keeps the backbone close to the reference implementation and converts its normalized-grid prediction to a voxel displacement, so the output plugs into every MIR utility.

Variants

  • LessNet: base model; SVF=True integrates the field with VecInt.

  • LessNetSPR: adds a spatially varying regularization head and follows the TransMorphTVFSPR return convention.

  • LessNetUNet: the bare backbone.

Default configuration

get_LessNet_default_config follows the upstream reproduction command: start_channel = 8, NCC_mok similarity (using_l2 = 2), smth_labda = 5.0, lr = 1e-4. get_LessNet_SVF_config and get_LessNet_SPR_config enable the respective variants.

Usage

from MIR.models import LessNet, LessNetSPR
from MIR.models.configs_LessNet import get_LessNet_default_config, get_LessNet_SVF_config, get_LessNet_SPR_config

model = LessNet(get_LessNet_default_config())
flow = model((mov, fix))

svf_model = LessNet(get_LessNet_SVF_config(), gen_output=True)
warped, flow = svf_model((mov, fix))

spr_model = LessNetSPR(get_LessNet_SPR_config(), SVF=True)
pos_flow, neg_flow, wts = spr_model((mov, fix))

Key modules

  • MIR.models.LessNet, MIR.models.configs_LessNet

References

  • Jia, Bartlett, Zhang, Lu, Qiu, Duan. U-Net vs transformer: is U-Net outdated in medical image registration? MLMI 2022 (LessNet).