Backbones and heads

Overview

The building blocks shared by the registration models can be used on their own, for example to attach a segmentation or synthesis head to pretrained registration features.

Backbones

  • MIR.models.Swin_Transformer.SwinTransformer: 3-D Swin Transformer encoder (TransMorph).

  • MIR.models.Deformable_Swin_Transformer.DefSwinTransformer and Deformable_Swin_Transformer_v2.DefSwinTransformerV2: deformable cross-attention encoders (TransMorph-DCA).

  • MIR.models.TransMorph.Conv3dReLU, DecoderBlock, RegistrationHead: decoder blocks and the zero-initialized registration head.

  • MIR.models.VFA.Encoder / Decoder, MIR.models.LessNet.LessNetUNet, MIR.models.FourierNet.FourierNetBackbone.

Heads

  • MIR.models.Segmentation_Heads.AdvancedDecoder3D: segmentation decoder on registration features (used with VFA, see the VFA_image_seg tutorial).

  • MIR.models.Synthesis_Heads: SynthesisHead3D, SynthesisHead3DAdvanced, EfficientAdvancedSynthHead3D, ConvNeXtSynthHead3D (VFA_image_synthesis).

  • MIR.models.Selfsupervised_Learning_Heads: SSLHeadNLvl, SSLHead1Lvl for self-supervised pretraining of the encoders.

  • MIR.models.lddmm.SPRHead and KernelWeightHead: model-agnostic heads that predict a spatial regularization map or kernel-mixture weights from the images and a field.