niftynet.network.dense_vnet module

class DenseVNet(num_classes, hyperparams={}, w_initializer=None, w_regularizer=None, b_initializer=None, b_regularizer=None, acti_func='relu', name='DenseVNet')

Bases: niftynet.network.base_net.BaseNet

implementation of Dense-V-Net:

Gibson et al., “Automatic multi-organ segmentation on abdominal CT with dense V-networks” 2018

### Diagram

DFS = Dense Feature Stack Block

• Initial image is first downsampled to a given size.
• Each DFS+SD outputs a skip link + a downsampled output.
• All outputs are upscaled to the initial downsampled size.
• If initial prior is given add it to the output prediction.

Input

–[ DFS ]———————–[ Conv ]————[ Conv ]——[+]–>

| | |

—–[ DFS ]—————[ Conv ]—— | |

| |

—–[ DFS ]——-[ Conv ]——— |

[ Prior ]—

The layer DenseFeatureStackBlockWithSkipAndDownsample layer implements [DFS + Conv + Downsampling] in a single module, and outputs 2 elements:

• Skip layer: [ DFS + Conv]
• Downsampled output: [ DFS + Down]

create_network()

layer_op(input_tensor, is_training=True, layer_id=-1, keep_prob=0.5, **unused_kwargs)

class SpatialPriorBlock(prior_shape, output_shape, name='spatial_prior_block')

Bases: niftynet.layer.base_layer.TrainableLayer

layer_op()

class DenseFeatureStackBlock(n_dense_channels, kernel_size, dilation_rates, use_bdo, name='dense_feature_stack_block', **kwargs)

Bases: niftynet.layer.base_layer.TrainableLayer

Dense Feature Stack Block

• Stack is initialized with the input from above layers.
• Iteratively the output of convolution layers is added to the feature stack.
• Each sequential convolution is performed over all the previous stacked channels.

Diagram example:

feature_stack = [Input] feature_stack = [feature_stack, conv(feature_stack)] feature_stack = [feature_stack, conv(feature_stack)] feature_stack = [feature_stack, conv(feature_stack)] … Output = [feature_stack, conv(feature_stack)]

create_block()

layer_op(input_tensor, is_training=True, keep_prob=None)

class DenseFeatureStackBlockWithSkipAndDownsample(n_dense_channels, kernel_size, dilation_rates, n_seg_channels, n_down_channels, use_bdo, name='dense_feature_stack_block', **kwargs)

Bases: niftynet.layer.base_layer.TrainableLayer

Dense Feature Stack with Skip Layer and Downsampling

• Downsampling is done through strided convolution.

—[ DenseFeatureStackBlock ]———-[ Conv ]——- Skip layer

——————– Downsampled Output

See DenseFeatureStackBlock for more info.

create_block()

layer_op(input_tensor, is_training=True, keep_prob=None)