Sensitivity-Specifity loss paper: http://www.rogertam.ca/Brosch_MICCAI_2015.pdf tf code: https://github.com/NifTK/NiftyNet/bl

square_volumes will square the weight term. The paper recommends square_volumes=True; I don't (just an intuition)

Sensitivity-Specifity loss paper: http://www.rogertam.ca/Brosch_MICCAI_2015.pdf tf code: https://github.com/NifTK/NiftyNet

compute boudary loss only compute the loss of foreground ref: https://github.com/LIVIAETS/surface-loss/blob/108bd9892adca476e

net_output must be (b, c, x, y(, z))) gt must be a label map (shape (b, 1, x, y(, z)) OR shape (b, x, y(, z))) or one hot encoding (b, c, x,

net_output must be (b, c, x, y(, z))) gt must be a label map (shape (b, 1, x, y(, z)) OR shape (b, x, y(, z))) or one hot encoding (b, c, x

net_output must be (b, c, x, y(, z))) gt must be a label map (shape (b, 1, x, y(, z)) OR shape (b, x, y(, z))) or one hot encoding (b, c, x,

create self.output_folder modify self.output_folder if you are doing cross-validation (one folder per fold) set self.tr_gen

compute boudary loss only compute the loss of foreground ref: https://github.com/LIVIAETS/surface-loss/blob/108bd9892adca4

This is specific to the U-Net and must be adapted for other network architectures :return:

This is what actually configures the entire experiment. The plans file is generated by experiment planning :return:

Don't use this. If you need softmax output, use preprocess_predict_nifti and set softmax_output_file. :param data: :param do_

This is a suggestion for if your dataset is a dictionary (my personal standard) :return:

Flattens a given tensor such that the channel axis is first. The shapes are transformed as follows: (N, C, D, H, W) -> (C, N * D * H * W)

Flattens a given tensor such that the channel axis is first. The shapes are transformed as follows: (N, C, D, H, W) -> (C, N * D * H * W)

GT.shape = (batch_size, x,y,z) only for binary segmentation

GT.shape = (batch_size, x,y,z) only for binary segmentation

GT.shape = (batch_size, x,y,z) only for binary segmentation

compute the distance transform map of foreground in ground gruth. input: segmentation, shape = (batch_size, class, x, y, z) output: the f

compute the distance transform map of foreground in prediction. input: segmentation, shape = (batch_size, class, x, y, z) output: the for

compute the signed distance map of binary mask input: segmentation, shape = (batch_size, x, y, z) output: the Signed Distance Map (SDM)

Can be implemented, does not have to :return:

net_output must be (b, c, x, y(, z))) gt must be a label map (shape (b, 1, x, y(, z)) OR shape (b, x, y(, z))) or one hot encoding (b, c, x,

copy from: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunet/training/loss_functions/dice_loss.py net_output must be (b, c, x, y(, z)))

net_output must be (b, c, x, y(, z))) gt must be a label map (shape (b, 1, x, y(, z)) OR shape (b, x, y(, z))) or one hot encoding (b, c, x,

used for if the checkpoint is already in ram :param saved_model: :param train: :return:

Computes gradient of the Lovasz extension w.r.t sorted errors See Alg. 1 in paper

Computes gradient of the Lovasz extension w.r.t sorted errors See Alg. 1 in paper

Saves a checkpoint every save_ever epochs. :return:

Should probably by improved :return:

Used to predict new unseen data. Not used for the preprocessing of the training/test data :param input_files: :return:

Can be implemented, does not have to :param output_torch: :param target_npy: :return:

If self.all_val_eval_metrics is unused (len=0) then we fall back to using -self.all_val_losses for the MA to determine early stopping

used to swap between folds for inference (ensemble of models from cross-validation) DO NOT USE DURING TRAINING AS THIS WILL NOT UPDAT

A generic class that can train almost any neural network (RNNs excluded). It provides basic functionality such as the training loop,

:param deterministic: :param fold: can be either [0 ... 5) for cross-validation, 'all' to train on all available training data or