niftynet.layer.loss_segmentation module

Loss functions for multi-class segmentation

class LossFunction(n_class, loss_type='Dice', softmax=True, loss_func_params=None, name='loss_function')

Bases: niftynet.layer.base_layer.Layer

layer_op(prediction, ground_truth, weight_map=None)

Compute loss from prediction and ground truth, the computed loss map are weighted by weight_map.

if prediction `is list of tensors, each element of the list will be compared against `ground_truth and the weighted by weight_map. (Assuming the same gt and weight across scales)

Parameters:

• prediction – input will be reshaped into (batch_size, N_voxels, num_classes)
• ground_truth – input will be reshaped into (batch_size, N_voxels, ...)
• weight_map – input will be reshaped into (batch_size, N_voxels, ...)

Returns:

labels_to_one_hot(ground_truth, num_classes=1)

Converts ground truth labels to one-hot, sparse tensors. Used extensively in segmentation losses.

Parameters:

• ground_truth – ground truth categorical labels (rank N)
• num_classes – A scalar defining the depth of the one hot dimension (see depth of tf.one_hot)

Returns:

one-hot sparse tf tensor (rank N+1; new axis appended at the end)

undecided_loss(prediction, ground_truth, weight_map=None)

Parameters:

• prediction –
• ground_truth –
• weight_map –

Returns:

volume_enforcement(prediction, ground_truth, weight_map=None, eps=0.001, hard=False)

Computing a volume enforcement loss to ensure that the obtained volumes are close and avoid empty results when something is expected :param prediction: :param ground_truth: labels :param weight_map: potential weight map to apply :param eps: epsilon to use as regulariser :return:

volume_enforcement_fin(prediction, ground_truth, weight_map=None, eps=0.001)

Computing a volume enforcement loss to ensure that the obtained volumes are

close and avoid empty results when something is expected

Parameters:

• prediction –
• ground_truth –
• weight_map –
• eps –

Returns:

generalised_dice_loss(prediction, ground_truth, weight_map=None, type_weight='Square')

Function to calculate the Generalised Dice Loss defined in

Sudre, C. et. al. (2017) Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations. DLMIA 2017

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground truth
• weight_map –
• type_weight – type of weighting allowed between labels (choice between Square (square of inverse of volume), Simple (inverse of volume) and Uniform (no weighting))

Returns:

the loss

dice_plus_xent_loss(prediction, ground_truth, weight_map=None)

Function to calculate the loss used in https://arxiv.org/pdf/1809.10486.pdf, no-new net, Isenseee et al (used to win the Medical Imaging Decathlon).

It is the sum of the cross-entropy and the Dice-loss.

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground truth
• weight_map –

Returns:

the loss (cross_entropy + Dice)

sensitivity_specificity_loss(prediction, ground_truth, weight_map=None, r=0.05)

Function to calculate a multiple-ground_truth version of the sensitivity-specificity loss defined in “Deep Convolutional Encoder Networks for Multiple Sclerosis Lesion Segmentation”, Brosch et al, MICCAI 2015, https://link.springer.com/chapter/10.1007/978-3-319-24574-4_1

error is the sum of r(specificity part) and (1-r)(sensitivity part)

Parameters:

• prediction – the logits
• ground_truth – segmentation ground_truth.
• r – the ‘sensitivity ratio’ (authors suggest values from 0.01-0.10 will have similar effects)

Returns:

the loss

cross_entropy(prediction, ground_truth, weight_map=None)

Function to calculate the cross-entropy loss function

Parameters:

• prediction – the logits (before softmax)
• ground_truth – the segmentation ground truth
• weight_map –

Returns:

the cross-entropy loss

cross_entropy_dense(prediction, ground_truth, weight_map=None)

wasserstein_disagreement_map(prediction, ground_truth, weight_map=None, M=None)

Function to calculate the pixel-wise Wasserstein distance between the flattened prediction and the flattened labels (ground_truth) with respect to the distance matrix on the label space M.

Parameters:

• prediction – the logits after softmax
• ground_truth – segmentation ground_truth
• M – distance matrix on the label space

Returns:

the pixelwise distance map (wass_dis_map)

generalised_wasserstein_dice_loss(prediction, ground_truth, weight_map=None)

Function to calculate the Generalised Wasserstein Dice Loss defined in

Fidon, L. et. al. (2017) Generalised Wasserstein Dice Score for Imbalanced Multi-class Segmentation using Holistic Convolutional Networks.MICCAI 2017 (BrainLes)

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground_truth
• weight_map –

Returns:

the loss

dice(prediction, ground_truth, weight_map=None)

Function to calculate the dice loss with the definition given in

Milletari, F., Navab, N., & Ahmadi, S. A. (2016) V-net: Fully convolutional neural networks for volumetric medical image segmentation. 3DV 2016

using a square in the denominator

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground_truth
• weight_map –

Returns:

the loss

dice_nosquare(prediction, ground_truth, weight_map=None)

Function to calculate the classical dice loss

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground_truth
• weight_map –

Returns:

the loss

tversky(prediction, ground_truth, weight_map=None, alpha=0.5, beta=0.5)

Function to calculate the Tversky loss for imbalanced data

Sadegh et al. (2017)

Tversky loss function for image segmentation using 3D fully convolutional deep networks

Parameters:

• prediction – the logits
• ground_truth – the segmentation ground_truth
• alpha – weight of false positives
• beta – weight of false negatives
• weight_map –

Returns:

the loss

dice_dense(prediction, ground_truth, weight_map=None)

Computing mean-class Dice similarity.

Parameters:

• prediction – last dimension should have num_classes
• ground_truth – segmentation ground truth (encoded as a binary matrix) last dimension should be num_classes
• weight_map –

Returns:

1.0 - mean(Dice similarity per class)

dice_dense_nosquare(prediction, ground_truth, weight_map=None)

Computing mean-class Dice similarity with no square terms in the denominator

Parameters:

• prediction – last dimension should have num_classes
• ground_truth – segmentation ground truth (encoded as a binary matrix) last dimension should be num_classes
• weight_map –

Returns:

1.0 - mean(Dice similarity per class)