Layers

Summary of layers that can be used to build networks and their characteristics

Activation

File: activation.py Possible activation layers to be specified as the acti_func field of ConvolutionLayer or DeconvolutionLayer among the following values:

Field	Equation
relu	relu_eq
relu6	relu6_eq
elu	elu_eq
softplus	softplus_eq
softsign	softsign_eq
sigmoid	sigmoid_eq
tanh	tanh_eq
prelu	prelu_eq
dropout

Batch normalisation

File: bn.py Class: BNLayer Fields:

• regularizer: None
• moving_decay: 0.9
• eps
• name Layer that applies normalisation to the input batch.

Convolution

File: convolution.py Classes: ConvLayer and ConvolutionLayer

The convolution layer (called as ConvLayer) takes as fields:

• n_output _chns : Number of output channels/features
• kernel_size : window size of the convolution
• stride : Stride with which the kernel is applied
• padding : ['SAME'/'VALID'] Padding strategy applied
• with_bias: [True/False] Application of a bias
• w_initializer: Initialisation for the weights
• w_regularizer: Regularisation strategy for the weights
• b_initializer: Initialisation strategy for the bias
• b_regularizer: Regularisation strategy for the bias

ConvolutionLayer combines the following optional layers: ConvLayer, batch normalisation, activation, Takes as fields those necessary for a ConvLayer, a BatchNormLayer, an activation layer. Dropout is applied according to the argument keep_prob applied when calling the layer.

Conditional Random Field

Cropping

File: crop.py Class: CropLayer Operates the cropping of the data it is applied to resulting in the centered part of the data given the field border cropped in each spatial dimension on both sides.

Deconvolution

File: deconvolution.py Class: DeconvLayer and DeconvolutionLayer

Fields are similar to those needed for ConvLayer

DeconvolutionLayer composes in order the following (optional) layers: deconvolution, batch normalisation, activation and dropout. See the fields and arguments of the convolution file for details.

Dilated context

Downsampling

File: downsample.py Class: DownSampleLayer Fields:

• func: ['AVG'/'MAX'/'CONSTANT'].
• kernel_size: Determines the size of the kernel that will be applied
• stride: Striding parameter of the kernel
• padding: ['SAME'/'VALID']
• name

Elementwise operations

File: elementwise.py Class: ElementwiseLayer Fields:

• func: ['SUM','CONCAT']
• initializer
• regularizer
• name

Performs elementwise operations between two outputs coming from two different network flows given as arguments.

Case of SUM operation: 0 padding is applied on the features dimension for the second argument if the first argument has more features. Projection is made in the opposite case.

Input normalisation

File: input_normalisation.py Class: HistogramNormalisationLayer Fields:

• models_ filename: Text file with the intensity landmarks trained for each modality. It corresponds to the histogram_ref _file field given on the command line/ config file. If not provided, the folders
• multimod_mask _type ['and'/'or'/'all']: Strategy applied when creating a mask used for landmark extraction from multiple modalities either as intersection (and), union (or) or considering each modality separately (all)
• norm_type: ['percentile'/'quartile'] strategy of landmarks used for the piecewise linear adaptation
• cutoff: Landmarks cutoff points to be used for the histogram matching. 2 values (min and max) should be given in the range ]0 - 1[ Default value is (0.05,0.95)
• mask_type: ['otsu_plus'/ 'otsu_minus'/ 'thresh_plus'/ 'thresh_minus']. Strategy applied to obtain the image mask
• name

Layer takes as arguments the input 5d image on which to apply the normalisation and flags indicating if any normalisation and/or whitening should be applied. The normalisation follows the method developed by Nyul et al ¹

Loss functions

Loss functions are application-specific.

File: loss_segmentation.py Class: LossFunction Fields:

• n_class: Number of classes/labels
• loss_type: ['CrossEntropy'/ 'Dice'/ 'Dice_NS'/ 'GDSC'/ 'WGDL'/ 'SensSpec'/ 'L1Loss'/ 'L2Loss'/ 'Huber'] Name of the loss to be applied
• loss_func_params: Additional parameters to be used for the specified loss
• name

Following is a brief description of the loss functions for segmentation:

Loss function	Notes	Citation	Additional Arguments
Cross-entropy
Dice Loss		"V-net: Fully convolutional neural networks for volumetric medical image segmentation", Milletari, et al, 3DV 2016.
Dice Loss (no square)	Similar to Dice Loss, but probabilities are not squared in the denominator.
Generalised Dice Loss		"Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations", Sudre, C. et. al. DLMIA 2017.	type_weight: default 'Square'. Indicates how the volume of each label is weighted. Square - Multiplication by the inverse of square of the volume / Simple - Multiplication by 1/V / Uniform - No weighting
Generalised Wasserstein Dice Loss		"Generalised Wasserstein Dice Score for Imbalanced Multi-class Segmentation using Holistic Convolutional Networks", Fidon, L. et. al. MICCAI 2017 (BrainLes).
Sensitivity-Specificity Loss		"Deep Convolutional Encoder Networks for Multiple Sclerosis Lesion Segmentation", Brosch et al, MICCAI 2015.	r: default 0.05. The 'sensitivity ratio' (authors suggest values from 0.01-0.10 will have similar effects)

File: loss_regression.py Class: LossFunction

• n_class: Number of classes/labels
• loss_type: ['L1Loss'/ 'L2Loss'/ 'Huber'/ 'RMSE'] Name of the loss to be applied
• loss_func_params: Additional parameters to be used for the specified loss
• name

Following is a brief description of the regression loss functions:

Loss function	Notes	Citation	Additional Arguments
L1 Loss
L2 Loss
Huber Loss	The Huber loss is a smooth piecewise loss function that is quadratic for |x| <= delta, and linear for |x|> delta. See https://en.wikipedia.org/wiki/Huber_loss		delta: default 1.0
Root Mean Square Error

Random flip

This layer introduces flipping along user-specified axes. This can be useful as a data-augmentation step in training.

File: rand_flip.py Class: RandomFlipLayer Fields:

• flip_axes: which axes to flip on.
• flip_probability: default 0.5. The probability of flipping along any of the specified axes.

Random rotation

File: rand_rotation.py Class: RandomRotationLayer Fields:

• min_angle: Minimum angle considered in the random range
• max_angle: Maximum angle considered in the random range

The random rotation belongs to the set of possible augmentation operations. Note The random rotation is only applied on 3d data.

Random spatial scaling

File: rand_spatial_scaling.py Class: RandomSpatialScalingLayer.py Fields:

• min_percentage: Value between 0 and 100 that indicates the range of possible random scaling to be applied.
• max_percentage: Value between 0 and 100 that indicates the range of possible random scaling to be applied.
• name

The random spatial scaling is one of the possible augmentation operations. The scaling factor is computed as [(min_percentage+100)/100, (max_percentage+100)/100]

Upsampling

File: upsample.py Class: UpSampleLayer Fields:

• func: ['REPLICATE'/'CHANNELWISE_DECONV']
• kernel_size
• stride
• w_initializer
• w_regularizer
• with_bias
• b_initializer
• b_regularizer

This upsampling layer can allow for two strategies of upsampling. Note if func is REPLICATE, kernel and strides must be of the same size. The data is locally replicated. If func is CHANNELWISE_DECONV, the channels are upsampled separately using a DeconvLayer.

Footnotes

1. Nyúl, L. G., Udupa, J. K., & Zhang, X. (2000). New variants of a method of MRI scale standardization. IEEE transactions on medical imaging, 19(2), 143-150. ↩