MIR.models.HyperTransMorph

Hyper-TransMorph: TransMorph conditioned on a regularization hyperparameter.

Convolution and linear weights are predicted from an embedding of the hyperparameter value (hypernetwork conditioning), so a single trained network covers a range of regularization strengths.

Chen, J., Du, Y., He, Y., Segars, W. P., Li, Y., & Frey, E. C. (2021). TransMorph: Transformer for unsupervised medical image registration. arXiv preprint arXiv:2111.10480. Swin-Transformer code retrieved from: https://github.com/SwinTransformer/Swin-Transformer-Semantic-Segmentation Original paper: Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., … & Guo, B. (2021). Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030. Modified and tested by: Junyu Chen jchen245@jhmi.edu Johns Hopkins University

Functions

window_partition(x, window_size)	Partition a feature map into non-overlapping 3D windows.
window_reverse(windows, window_size, H, W, L)	Merge 3D windows back into a feature map (inverse of window_partition()).

Classes

BasicLayer(*args, **kwargs)	A basic Swin Transformer layer for one stage.
Conv3dReLU(*args, **kwargs)	Hyper-conditioned convolution, LeakyReLU, then normalization.
CustomConv(*args, **kwargs)	Hyper-conditioned 3D convolution with weights predicted from parameters.
DecoderBlock(*args, **kwargs)	Decoder block: 2x trilinear upsampling, optional skip concat, two convs.
HyperBlocks(*args, **kwargs)	MLP that embeds hyperparameters for hypernetwork conditioning.
HyperLinear(*args, **kwargs)	Linear layer with weights predicted from hyperparameters.
HyperTransMorphTVF(*args, **kwargs)	Hyper-conditioned TransMorph with time-varying (progressive) flow estimation.
HyperTransMorphTVFSPR(*args, **kwargs)	Hyper-conditioned TransMorph TVF with spatially-varying regularization.
Mlp(*args, **kwargs)	Feed-forward MLP block used inside transformer blocks.
PatchEmbed(*args, **kwargs)	Image to Patch Embedding.
PatchMerging(*args, **kwargs)	Patch Merging Layer.
RegistrationHead(*args, **kwargs)	Hyper-conditioned convolution producing a displacement field.
SinPositionalEncoding3D(*args, **kwargs)	3D sinusoidal positional encoding for channels-first feature maps.
SinusoidalPositionEmbedding(*args, **kwargs)	1D sinusoidal position embedding over the token axis.
SwinTransformer(*args, **kwargs)	Hyper-conditioned 3D Swin Transformer encoder.
SwinTransformerBlock(*args, **kwargs)	Swin Transformer Block.
WindowAttention(*args, **kwargs)	Window based multi-head self attention (W-MSA) module with relative position bias.