FastSurferCNN.models.networks

class FastSurferCNN.models.networks.FastSurferCNN(params, padded_size)

Main Fastsurfer CNN Network.

Attributes

classifier

Initialized Classification Block.

Methods

forward(x[, scale_factor, scale_factor_out])

Feedforward through graph.

forward(x, scale_factor=None, scale_factor_out=None)

Feedforward through graph.

Parameters:

xTensor

Input image [N, C, H, W].

scale_factorTensor, optional

[N, 1] Defaults to None.

scale_factor_outTensor, optional

Tensor representing the scale factor for the output. Defaults to None.

Returns:

outputTensor

Prediction logits.

class FastSurferCNN.models.networks.FastSurferCNNBase(params, padded_size=256)

Network Definition of Fully Competitive Network network.

• Spatial view aggregation (input 7 slices of which only middle one gets segmented)

• Same Number of filters per layer (normally 64)

• Dense Connections in blocks

• Unpooling instead of transpose convolutions

• Concatenationes are replaced with Maxout (competitive dense blocks)

• Global skip connections are fused by Maxout (global competition)

• Loss Function (weighted Cross-Entropy and dice loss)

Attributes

encode1, encode2, encode3, encode4	Competitive Encoder Blocks.
decode1, decode2, decode3, decode4	Competitive Decoder Blocks.
bottleneck	Bottleneck Block.

Methods

forward(x[, scale_factor, scale_factor_out])

Feedforward through graph.

forward(x, scale_factor=None, scale_factor_out=None)

Feedforward through graph.

Parameters:

xTensor

Input image [N, C, H, W] representing the input data.

scale_factorTensor, optional

[N, 1] Defaults to None.

scale_factor_outTensor, optional

Tensor representing the scale factor for the output. Defaults to None.

Returns:

decoder_output1Tensor

Prediction logits.

class FastSurferCNN.models.networks.FastSurferVINN(params, padded_size=256)

Network Definition of Fully Competitive Network.

• Spatial view aggregation (input 7 slices of which only middle one gets segmented)

• Same Number of filters per layer (normally 64)

• Dense Connections in blocks

• Unpooling instead of transpose convolutions

• Concatenationes are replaced with Maxout (competitive dense blocks)

• Global skip connections are fused by Maxout (global competition)

• Loss Function (weighted Cross-Entropy and dice loss)

Attributes

height	The height of segmentation model (after interpolation layer).
width	The width of segmentation model.
out_tensor_shape	Out tensor dimensions for interpolation layer.
interpolation_mode	Interpolation mode for up/downsampling in flex networks.
crop_position	Crop positions for up/downsampling in flex networks.
inp_block	Initialized input dense block.
outp_block	Initialized output dense block.
interpol1	Initialized 2d input interpolation block.
interpol2	Initialized 2d output interpolation block.
classifier	Initialized Classification Block.

Methods

forward(x, scale_factor[, scale_factor_out])

Feedforward through graph.

forward(x, scale_factor, scale_factor_out=None)

Feedforward through graph.

Parameters:

xTensor

Input image [N, C, H, W].

scale_factorTensor

Tensor of shape [N, 1] representing the scale factor for each image in the batch.

scale_factor_outTensor, optional

Tensor representing the scale factor for the output. Defaults to None.

Returns:

logitsTensor

Prediction logits.

FastSurferCNN.models.networks.build_model(cfg)

Build requested model.

Parameters:

cfgyacs.config.CfgNode

Node of configs to be used.

Returns:

model

Object of the initialized model.