API Reference

This document is for developers of ivadomed, it contains the API functions.

Loader API

loader.film

normalize_metadata(ds_in: BidsDataset | Bids3DDataset | MRI2DSegmentationDataset, clustering_models: dict, debugging: bool, metadata_type: str, train_set: bool = False) → list, OneHotEncoder | list

Categorize each metadata value using a KDE clustering method, then apply a one-hot-encoding.

Parameters:

• ds_in (BidsDataset) – Dataset BidsDataset, Bids3D, MRI2D with metadata.

• clustering_models (dict) – Pre-trained clustering model that has been trained on metadata of the training set.

• debugging (bool) – If True, extended verbosity and intermediate outputs.

• metadata_type (str) – Choice between ‘mri_params’, ‘constrasts’ or the name of a column from the participants.tsv file.

• train_set (bool) – Indicates if the input dataset is the training dataset (True) or the validation or testing dataset (False).

Returns:

Dataset with normalized metadata. If train_set is True, then the one-hot-encoder model is also

returned.

Return type:

BidsDataset

class Kde_model

Bases: object

Kernel Density Estimation.

Apply this clustering method to metadata values, using (sklearn implementation.)

Attributes:

• kde (sklearn.neighbors.KernelDensity)

• minima (float) – Local minima.

__init__() → None

clustering_fit(dataset: list, key_lst: List[str]) → dict

This function creates clustering models for each metadata type, using Kernel Density Estimation algorithm.

Parameters:

• dataset (list) – data

• key_lst (list of str) – names of metadata to cluster

Returns:

Clustering model for each metadata type in a dictionary where the keys are the metadata names.

Return type:

dict

check_isMRIparam(mri_param_type: str, mri_param: dict, subject: str, metadata: dict) → bool

Check if a given metadata belongs to the MRI parameters.

Parameters:

• mri_param_type (str) – Metadata type name.

• mri_param (dict) – List of MRI params names.

• subject (str) – Current subject name.

• metadata (dict) – Metadata.

Returns:

True if mri_param_type is part of mri_param.

Return type:

bool

get_film_metadata_models(ds_train: MRI2DSegmentationDataset, metadata_type: str, debugging: bool = False)

Get FiLM models.

This function pulls the clustering and one-hot encoder models that are used by FiLMedUnet. It also calls the normalization of metadata.

Parameters:

• ds_train (MRI2DSegmentationDataset) – training dataset

• metadata_type (string) – eg mri_params, contrasts

• debugging (bool) –

Returns:

dataset, one-hot encoder and KDE model

Return type:

MRI2DSegmentationDataset, OneHotEncoder, KernelDensity

store_film_params(gammas: dict, betas: dict, metadata_values: list, metadata: list, model: nn.Module, film_layers: list, depth: int, film_metadata: str)

Store FiLM params.

Parameters:

• gammas (dict) –

• betas (dict) –

• metadata_values (list) – list of the batch sample’s metadata values (e.g., T2w, astrocytoma)

• metadata (list) –

• model (nn.Module) –

• film_layers (list) –

• depth (int) –

• film_metadata (str) – Metadata of interest used to modulate the network (e.g., contrast, tumor_type).

Returns:

gammas, betas, metadata_values

Return type:

dict, dict, list

save_film_params(gammas: dict, betas: dict, metadata_values: list, depth: int, ofolder: str) → None

Save FiLM params as npy files.

These parameters can be further used for visualisation purposes. They are saved in the ofolder with .npy format.

Parameters:

• gammas (dict) –

• betas (dict) –

• metadata_values (list) – list of the batch sample’s metadata values (eg T2w, T1w, if metadata type used is

• contrast) –

• depth (int) –

• ofolder (str) –

loader.loader

load_dataset(bids_df: BidsDataframe, data_list: List[str], transforms_params: dict, model_params: dict, target_suffix: List[str], roi_params: dict, contrast_params: dict, slice_filter_params: dict, patch_filter_params: dict, slice_axis: str, multichannel: bool, dataset_type: str = 'training', requires_undo: bool = False, metadata_type: str | None = None, object_detection_params: dict | None = None, soft_gt: bool = False, device: device | None = None, cuda_available: bool | None = None, is_input_dropout: bool = False, **kwargs) → Bids3DDataset

Get loader appropriate loader according to model type. Available loaders are Bids3DDataset for 3D data, BidsDataset for 2D data and HDF5Dataset for HeMIS.

Parameters:

• bids_df (BidsDataframe) – Object containing dataframe with all BIDS image files and their metadata.

• data_list (list) – Subject names list.

• transforms_params (dict) – Dictionary containing transformations for “training”, “validation”, “testing” (keys), eg output of imed_transforms.get_subdatasets_transforms.

• model_params (dict) – Dictionary containing model parameters.

• target_suffix (list of str) – List of suffixes for target masks.

• roi_params (dict) – Contains ROI related parameters.

• contrast_params (dict) – Contains image contrasts related parameters.

• slice_filter_params (dict) – Contains slice_filter_params, see Configuration File for more details.

• patch_filter_params (dict) – Contains patch_filter_params, see Configuration File for more details.

• slice_axis (string) – Choice between “axial”, “sagittal”, “coronal” ; controls the axis used to extract the 2D data from 3D NifTI files. 2D PNG/TIF/JPG files use default “axial.

• multichannel (bool) – If True, the input contrasts are combined as input channels for the model. Otherwise, each contrast is processed individually (ie different sample / tensor).

• metadata_type (str) – Choice between None, “mri_params”, “contrasts”.

• dataset_type (str) – Choice between “training”, “validation” or “testing”.

• requires_undo (bool) – If True, the transformations without undo_transform will be discarded.

• object_detection_params (dict) – Object dection parameters.

• soft_gt (bool) – If True, ground truths are not binarized before being fed to the network. Otherwise, ground

• thresholded (truths are) –

• device (torch.device) – Device to use for the model training.

• cuda_available (bool) – If True, cuda is available.

• is_input_dropout (bool) – Return input with missing modalities.

Returns:

BidsDataset

Note: For more details on the parameters transform_params, target_suffix, roi_params, contrast_params, slice_filter_params, patch_filter_params and object_detection_params see Configuration File.

loader.utils

split_dataset(df: pd.DataFrame, split_method: str, data_testing: dict, random_seed: int, train_frac: float = 0.8, test_frac: float = 0.1)

Splits dataset into training, validation and testing sets by applying train, test and validation fractions according to the split_method. The “data_testing” parameter can be used to specify the data_type and data_value to include in the testing set, the dataset is then split as not to mix the data_testing between the training/validation set and the testing set.

Parameters:

• df (pd.DataFrame) – Dataframe containing all BIDS image files indexed and their metadata.

• split_method (str) – Used to specify on which metadata to split the dataset (eg. “participant_id”, “sample_id”, etc.)

• data_testing (dict) – Used to specify data_type and data_value to include in the testing set.

• random_seed (int) – Random seed to ensure reproducible splits.

• train_frac (float) – Between 0 and 1. Represents the train set proportion.

• test_frac (float) – Between 0 and 1. Represents the test set proportion.

Returns:

Train, validation and test filenames lists.

Return type:

list, list, list

get_new_subject_file_split(df: pd.DataFrame, split_method: str, data_testing: dict, random_seed: int, train_frac: float, test_frac: float, path_output: str, balance: str, subject_selection: dict = None)

Randomly split dataset between training / validation / testing.

Randomly split dataset between training / validation / testing and save it in path_output + “/split_datasets.joblib”.

Parameters:

• df (pd.DataFrame) – Dataframe containing all BIDS image files indexed and their metadata.

• split_method (str) – Used to specify on which metadata to split the dataset (eg. “participant_id”, “sample_id”, etc.)

• data_testing (dict) – Used to specify the data_type and data_value to include in the testing set.

• random_seed (int) – Random seed.

• train_frac (float) – Training dataset proportion, between 0 and 1.

• test_frac (float) – Testing dataset proportionm between 0 and 1.

• path_output (str) – Output folder.

• balance (str) – Metadata contained in “participants.tsv” file with categorical values. Each category will be

• training (evenly distributed in the) –

• datasets. (validation and testing) –

• subject_selection (dict) – Used to specify a custom subject selection from a dataset.

Returns:

Training, validation and testing filenames lists.

Return type:

list, list list

get_subdatasets_subject_files_list(split_params: dict, df: pd.DataFrame, path_output: str, subject_selection: dict = None)

Get lists of subject filenames for each sub-dataset between training / validation / testing.

Parameters:

• split_params (dict) – Split parameters, see Configuration File for more details.

• df (pd.DataFrame) – Dataframe containing all BIDS image files indexed and their metadata.

• path_output (str) – Output folder.

• subject_selection (dict) – Used to specify a custom subject selection from a dataset.

Returns:

Training, validation and testing filenames lists.

Return type:

list, list list

imed_collate(batch: dict) → dict | list | str | Tensor

Collates data to create batches

Parameters:

batch (dict) – Contains input and gt data with their corresponding metadata.

Returns:

Collated data.

Return type:

list or dict or str or tensor

filter_roi(roi_data: ndarray, nb_nonzero_thr: int) → bool

Filter slices from dataset using ROI data.

This function filters slices (roi_data) where the number of non-zero voxels within the ROI slice (e.g. centerline, SC segmentation) is inferior or equal to a given threshold (nb_nonzero_thr).

Parameters:

• roi_data (nd.array) – ROI slice.

• nb_nonzero_thr (int) – Threshold.

Returns:

True if the slice needs to be filtered, False otherwise.

Return type:

bool

orient_img_hwd(data: ndarray, slice_axis: int) → ndarray

Orient a given RAS image to height, width, depth according to slice axis.

Parameters:

• data (ndarray) – RAS oriented data.

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

Returns:

Array oriented with the following dimensions: (height, width, depth).

Return type:

ndarray

orient_img_ras(data: ndarray, slice_axis: int) → ndarray

Orient a given array with dimensions (height, width, depth) to RAS orientation.

Parameters:

• data (ndarray) – Data with following dimensions (Height, Width, Depth).

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

Returns:

Array oriented in RAS.

Return type:

ndarray

orient_shapes_hwd(data: list | tuple, slice_axis: int) → ndarray

Swap dimensions according to match the height, width, depth orientation.

Parameters:

• data (list or tuple) – Shape or numbers associated with each image dimension (e.g. image resolution).

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

Returns:

Reoriented vector.

Return type:

ndarray

update_metadata(metadata_src_lst: list, metadata_dest_lst: list) → list

Update metadata keys with a reference metadata.

A given list of metadata keys will be changed and given the values of the reference metadata.

Parameters:

• metadata_src_lst (list) – List of source metadata used as reference for the destination metadata.

• metadata_dest_lst (list) – List of metadate that needs to be updated.

Returns:

updated metadata list.

Return type:

list

reorient_image(arr: np.ndarray, slice_axis: int, nib_ref: nib, nib_ref_canonical: nib) → nd.ndarray

Reorient an image to match a reference image orientation.

It reorients a array to a given orientation and convert it to a nibabel object using the reference nibabel header.

Parameters:

• arr (ndarray) – Input array, array to re orient.

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

• nib_ref (nibabel) – Reference nibabel object, whose header is used.

• nib_ref_canonical (nibabel) – nib_ref that has been reoriented to canonical orientation (RAS).

get_file_extension(filename: str) → str | None

Get file extension if it is supported :param filename: Path of the file. :type filename: str

Returns:

File extension

Return type:

str

update_filename_to_nifti(filename: str) → str

Update filename extension to ‘nii.gz’ if not a NifTI file.

This function is used to help make non-NifTI files (e.g. PNG/TIF/JPG) compatible with NifTI-only pipelines. The expectation is that a NifTI version of the file has been created alongside the original file, which allows the extension to be cleanly swapped for a .nii.gz extension.

Parameters:

filename (str) – Path of original file.

Returns:

Path of the corresponding NifTI file.

Return type:

str

dropout_input(seg_pair: dict) → dict

Applies input-level dropout: zero to all channels minus one will be randomly set to zeros. This function verifies if some channels are already empty. Always at least one input channel will be kept.

Parameters:

seg_pair (dict) – Batch containing torch tensors (input and gt) and metadata.

Returns:

Batch containing torch tensors (input and gt) and metadata with channel(s) dropped.

Return type:

seg_pair (dict)

create_temp_directory() → str

Creates a temporary directory and returns its path. This temporary directory is only deleted when explicitly requested.

Returns:

Path of the temporary directory.

Return type:

str

get_obj_size(obj) → int

Returns the size of an object in bytes. Used to gauge whether storing object in memory vs write to disk.

Source: https://stackoverflow.com/a/53705610

Parameters:

obj –

Returns:

Object Detection API

object_detection.utils

get_bounding_boxes(mask)

Generates a 3D bounding box around a given mask. :param mask: Mask of the ROI. :type mask: Numpy array

Returns:

Bounding box coordinate (x_min, x_max, y_min, y_max, z_min, z_max).

Return type:

list

adjust_bb_size(bounding_box, factor, resample=False)

Modifies the bounding box dimensions according to a given factor.

Parameters:

• bounding_box (list or tuple) – Coordinates of bounding box (x_min, x_max, y_min, y_max, z_min, z_max).

• factor (list or tuple) – Multiplicative factor for each dimension (list or tuple of length 3).

• resample (bool) – Boolean indicating if this resize is for resampling.

Returns:

New coordinates (x_min, x_max, y_min, y_max, z_min, z_max).

Return type:

list

resize_to_multiple(shape, multiple, length)

Modify a given shape so each dimension is a multiple of a given number. This is used to avoid dimension mismatch with patch training. The return shape is always larger then the initial shape (no cropping).

Parameters:

• shape (tuple or list) – Initial shape to be modified.

• multiple (tuple or list) – Multiple for each dimension.

• length (tuple or list) – Patch length.

Returns:

New dimensions.

Return type:

list

generate_bounding_box_file(subject_path_list, model_path, path_output, gpu_id=0, slice_axis=0, contrast_lst=None, keep_largest_only=True, safety_factor=None)

Creates json file containing the bounding box dimension for each images. The file has the following format: {“path/to/img.nii.gz”: [[x1_min, x1_max, y1_min, y1_max, z1_min, z1_max], [x2_min, x2_max, y2_min, y2_max, z2_min, z2_max]]} where each list represents the coordinates of an object on the image (2 instance of a given object in this example).

Parameters:

• subject_path_list (list) – List of all subjects in all the BIDS directories.

• model_path (string) – Path to object detection model.

• path_output (string) – Output path.

• gpu_id (int) – If available, GPU number.

• slice_axis (int) – Slice axis (0: sagittal, 1: coronal, 2: axial).

• contrast_lst (list) – Contrasts.

• keep_largest_only (bool) – Boolean representing if only the largest object of the prediction is kept.

• safety_factor (list or tuple) – Factors to multiply each dimension of the bounding box.

Returns:

Dictionary containing bounding boxes related to their image.

Return type:

dict

resample_bounding_box(metadata, transform)

Resample bounding box.

Parameters:

• metadata (dict) – Dictionary containing the metadata to be modified with the resampled coordinates.

• transform (Compose) – Transformations possibly containing the resample params.

adjust_transforms(transforms, seg_pair, length=None, stride=None)

This function adapts the transforms by adding the BoundingBoxCrop transform according the specific parameters of an image. The dimensions of the crop are also adapted to fit the length and stride parameters if the 3D loader is used.

Parameters:

• transforms (Compose) – Prepreocessing transforms.

• seg_pair (dict) – Segmentation pair (input, gt and metadata).

• length (list or tuple) – Patch size of the 3D loader.

• stride (list or tuple) – Stride value of the 3D loader.

Returns:

Modified transforms.

Return type:

Compose

adjust_undo_transforms(transforms, seg_pair, index=0)

This function adapts the undo transforms by adding the BoundingBoxCrop to undo transform according the specific parameters of an image.

Parameters:

• transforms (Compose) – Transforms.

• seg_pair (dict) – Segmentation pair (input, gt and metadata).

• index (int) – Batch index of the seg_pair.

load_bounding_boxes(object_detection_params, subject_path_list, slice_axis, constrast_lst)

Verifies if bounding_box.json exists in the output path, if so loads the data, else generates the file if a valid detection model path exists.

Parameters:

• object_detection_params (dict) – Object detection parameters.

• subject_path_list (list) – List of all subjects in all the BIDS directories.

• slice_axis (int) – Slice axis (0: sagittal, 1: coronal, 2: axial).

• constrast_lst (list) – Contrasts.

Returns:

bounding boxes for every subject in BIDS directory

Return type:

dict

verify_metadata(metadata, has_bounding_box)

Validates across all metadata that the ‘bounding_box’ param is present.

Parameters:

• metadata (dict) – Image metadata.

• has_bounding_box (bool) – If ‘bounding_box’ is present across all metadata, True, else False.

Returns:

Boolean indicating if ‘bounding_box’ is present across all metadata.

Return type:

bool

bounding_box_prior(fname_mask, metadata, slice_axis, safety_factor=None)

Computes prior steps to a model requiring bounding box crop. This includes loading a mask of the ROI, orienting the given mask into the following dimensions: (height, width, depth), extracting the bounding boxes and storing the information in the metadata.

Parameters:

• fname_mask (str) – Filename containing the mask of the ROI

• metadata (dict) – Dictionary containing the image metadata

• slice_axis (int) – Slice axis (0: sagittal, 1: coronal, 2: axial)

• safety_factor (list or tuple) – Factors to multiply each dimension of the bounding box.

compute_bb_statistics(bounding_box_path)

Measures min, max and average, height, width, depth and volume of bounding boxes from a json file

Parameters:

bounding_box_path (string) – Path to json file.

Evaluation API

evaluate(bids_df, path_output, target_suffix, eval_params)

Evaluate predictions from inference step.

Parameters:

• bids_df (BidsDataframe) – Object containing dataframe with all BIDS image files and their metadata.

• path_output (str) – Folder where the output folder “results_eval” is be created.

• target_suffix (list) – List of suffixes that indicates the target mask(s).

• eval_params (dict) – Evaluation parameters.

Returns:

results for each image.

Return type:

pd.Dataframe

class Evaluation3DMetrics(data_pred, data_gt, dim_lst, params=None)

Bases: object

Computes 3D evaluation metrics.

Parameters:

• data_pred (ndarray) – Network prediction mask.

• data_gt (ndarray) – Ground-truth mask.

• dim_lst (list) – Resolution (mm) along each dimension.

• params (dict) – Evaluation parameters.

Attributes:

• data_pred (ndarray) – Network prediction mask.

• data_gt (ndarray) – Ground-truth mask.

• n_classes (int) – Number of classes.

• px (float) – Resolution (mm) along the first axis.

• py (float) – Resolution (mm) along the second axis.

• pz (float) – Resolution (mm) along the third axis.

• bin_struct (ndarray) – Binary structure.

• size_min (int) – Minimum size of objects. Objects that are smaller than this limit can be removed if “removeSmall” is in params.

• object_detection_metrics (bool) – Indicate if object detection metrics (lesions true positive and false detection rates) are computed or not.

• overlap_vox (int) – A prediction and ground-truth are considered as overlapping if they overlap for at least this amount of voxels.

• overlap_ratio (float) – A prediction and ground-truth are considered as overlapping if they overlap for at least this portion of their volumes.

• data_pred_label (ndarray) – Network prediction mask that is labeled, ie each object is filled with a different value.

• data_gt_label (ndarray) – Ground-truth mask that is labeled, ie each object is filled with a different value.

• n_pred (int) – Number of objects in the network prediction mask.

• n_gt (int) – Number of objects in the ground-truth mask.

• data_painted (ndarray) – Mask where each predicted object is labeled depending on whether it is a TP or FP.

__init__(data_pred, data_gt, dim_lst, params=None)

label_per_size(data)

Get data with labels corresponding to label size.

Parameters:

data (ndarray) – Input data.

Returns:

ndarray

get_vol(data)

Get volume.

get_rvd()

Relative volume difference.

The volume is here defined by the physical volume, in mm3, of the non-zero voxels of a given mask. Relative volume difference equals the difference between the ground-truth and prediction volumes, divided by the ground-truth volume. Optimal value is zero. Negative value indicates over-segmentation, while positive value indicates under-segmentation.

get_avd()

Absolute volume difference.

The volume is here defined by the physical volume, in mm3, of the non-zero voxels of a given mask. Absolute volume difference equals the absolute value of the Relative Volume Difference. Optimal value is zero.

get_ltpr(label_size=None, class_idx=0)

Lesion True Positive Rate / Recall / Sensitivity.

Parameters:

• label_size (int) – Size of label.

• class_idx (int) – Label index. If monolabel 0, else ranges from 0 to number of output channels - 1.

Note: computed only if n_obj >= 1 and “object_detection_metrics” evaluation parameter is True.

get_lfdr(label_size=None, class_idx=0)

Lesion False Detection Rate / 1 - Precision.

Parameters:

• label_size (int) – Size of label.

• class_idx (int) – Label index. If monolabel 0, else ranges from 0 to number of output channels - 1.

Note: computed only if n_obj >= 1 and “object_detection_metrics” evaluation parameter is True.

run_eval()

Stores evaluation results in dictionary

Returns:

dictionary containing evaluation results, data with each object painted a different color

Return type:

dict, ndarray

Losses API

class MultiClassDiceLoss(classes_of_interest=None)

Bases: Module

Multi-class Dice Loss.

Inspired from https://arxiv.org/pdf/1802.10508.

Parameters:

classes_of_interest (list) – List containing the index of a class which its dice will be added to the loss. If is None all classes are considered.

Attributes:

• classes_of_interest (list) – List containing the index of a class which its dice will be added to the loss. If is None all classes are considered.

• dice_loss (DiceLoss) – Class computing the Dice loss.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(classes_of_interest=None)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(prediction, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class DiceLoss(smooth=1.0)

Bases: Module

DiceLoss.

See also

Milletari, Fausto, Nassir Navab, and Seyed-Ahmad Ahmadi. “V-net: Fully convolutional neural networks for volumetric medical image segmentation.” 2016 fourth international conference on 3D vision (3DV). IEEE, 2016.

Parameters:

smooth (float) – Value to avoid division by zero when images and predictions are empty.

Attributes:

smooth (float) – Value to avoid division by zero when images and predictions are empty.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(smooth=1.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(prediction, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class BinaryCrossEntropyLoss

Bases: Module

(BinaryCrossEntropyLoss).

Attributes:

loss_fct (BCELoss) – Binary cross entropy loss function from torch library.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__()

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(prediction, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class FocalLoss(gamma=2, alpha=0.25, eps=1e-07)

Bases: Module

FocalLoss.

See also

Lin, Tsung-Yi, et al. “Focal loss for dense object detection.” Proceedings of the IEEE international conference on computer vision. 2017.

Parameters:

• gamma (float) – Value from 0 to 5, Control between easy background and hard ROI training examples. If set to 0, equivalent to cross-entropy.

• alpha (float) – Value from 0 to 1, usually corresponding to the inverse of class frequency to address class imbalance.

• eps (float) – Epsilon to avoid division by zero.

Attributes:

• gamma (float) – Value from 0 to 5, Control between easy background and hard ROI training examples. If set to 0, equivalent to cross-entropy.

• alpha (float) – Value from 0 to 1, usually corresponding to the inverse of class frequency to address class imbalance.

• eps (float) – Epsilon to avoid division by zero.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(gamma=2, alpha=0.25, eps=1e-07)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class FocalDiceLoss(beta=1, gamma=2, alpha=0.25)

Bases: Module

FocalDiceLoss.

See also

Wong, Ken CL, et al. “3D segmentation with exponential logarithmic loss for highly unbalanced object sizes.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018.

Parameters:

• beta (float) – Value from 0 to 1, indicating the weight of the dice loss.

• gamma (float) – Value from 0 to 5, Control between easy background and hard ROI training examples. If set to 0, equivalent to cross-entropy.

• alpha (float) – Value from 0 to 1, usually corresponding to the inverse of class frequency to address class imbalance.

Attributes:

• beta (float) – Value from 0 to 1, indicating the weight of the dice loss.

• gamma (float) – Value from 0 to 5, Control between easy background and hard ROI training examples. If set to 0, equivalent to cross-entropy.

• alpha (float) – Value from 0 to 1, usually corresponding to the inverse of class frequency to address class imbalance.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(beta=1, gamma=2, alpha=0.25)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class GeneralizedDiceLoss(epsilon=1e-05, include_background=True)

Bases: Module

GeneralizedDiceLoss.

See also

Sudre, Carole H., et al. “Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations.” Deep learning in medical image analysis and multimodal learning for clinical decision support. Springer, Cham, 2017. 240-248.

Parameters:

• epsilon (float) – Epsilon to avoid division by zero.

• include_background (bool) – If True, then an extra channel is added, which represents the background class.

Attributes:

• epsilon (float) – Epsilon to avoid division by zero.

• include_background (bool) – If True, then an extra channel is added, which represents the background class.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(epsilon=1e-05, include_background=True)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class TverskyLoss(alpha=0.7, beta=0.3, smooth=1.0)

Bases: Module

Tversky Loss.

See also

Salehi, Seyed Sadegh Mohseni, Deniz Erdogmus, and Ali Gholipour. “Tversky loss function for image segmentation using 3D fully convolutional deep networks.” International Workshop on Machine Learning in Medical Imaging. Springer, Cham, 2017.

Parameters:

• alpha (float) – Weight of false positive voxels.

• beta (float) – Weight of false negative voxels.

• smooth (float) – Epsilon to avoid division by zero, when both Numerator and Denominator of Tversky are zeros.

Attributes:

• alpha (float) – Weight of false positive voxels.

• beta (float) – Weight of false negative voxels.

• smooth (float) – Epsilon to avoid division by zero, when both Numerator and Denominator of Tversky are zeros.

Notes

• setting alpha=beta=0.5: Equivalent to DiceLoss.

• default parameters were suggested by https://arxiv.org/pdf/1706.05721.pdf .

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(alpha=0.7, beta=0.3, smooth=1.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

tversky_index(y_pred, y_true)

Compute Tversky index.

Parameters:

• y_pred (torch Tensor) – Prediction.

• y_true (torch Tensor) – Target.

Returns:

Tversky index.

Return type:

float

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class FocalTverskyLoss(alpha=0.7, beta=0.3, gamma=1.33, smooth=1.0)

Bases: TverskyLoss

Focal Tversky Loss.

See also

Abraham, Nabila, and Naimul Mefraz Khan. “A novel focal tversky loss function with improved attention u-net for lesion segmentation.” 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE, 2019.

Parameters:

• alpha (float) – Weight of false positive voxels.

• beta (float) – Weight of false negative voxels.

• gamma (float) – Typically between 1 and 3. Control between easy background and hard ROI training examples.

• smooth (float) – Epsilon to avoid division by zero, when both Numerator and Denominator of Tversky are zeros.

Attributes:

gamma (float) – Typically between 1 and 3. Control between easy background and hard ROI training examples.

Notes

• setting alpha=beta=0.5 and gamma=1: Equivalent to DiceLoss.

• default parameters were suggested by https://arxiv.org/pdf/1810.07842.pdf .

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(alpha=0.7, beta=0.3, gamma=1.33, smooth=1.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class L2loss

Bases: Module

Euclidean loss also known as L2 loss. Compute the sum of the squared difference between the two images.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__()

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class AdapWingLoss(theta=0.5, alpha=2.1, omega=14, epsilon=1)

Bases: Module

Adaptive Wing loss Used for heatmap ground truth.

See also

Wang, Xinyao, Liefeng Bo, and Li Fuxin. “Adaptive wing loss for robust face alignment via heatmap regression.” Proceedings of the IEEE International Conference on Computer Vision. 2019.

Parameters:

• theta (float) – Threshold between linear and non linear loss.

• alpha (float) – Used to adapt loss shape to input shape and make loss smooth at 0 (background).

• properties. (It needs to be slightly above 2 to maintain ideal) –

• omega (float) – Multiplicating factor for non linear part of the loss.

• epsilon (float) – factor to avoid gradient explosion. It must not be too small

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(theta=0.5, alpha=2.1, omega=14, epsilon=1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class LossCombination(losses_list, params_list=None)

Bases: Module

Loss that sums other implemented losses.

Parameters:

• losses_list (list) – list of losses that will be summed. Elements should be string.

• params_list (list) – list of params for the losses, contain None or dictionnary definition of params for the loss

• used. (at same index. If no params list is given all default parameter will be) –

• (e.g. – params_list = [None,{“param1:0.5”}])

• ["L2loss" (losses_list =) – params_list = [None,{“param1:0.5”}])

• "DiceLoss"] – params_list = [None,{“param1:0.5”}])

Returns:

sum of losses computed on (input,target) with the params

Return type:

tensor

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(losses_list, params_list=None)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input, target)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Main API

run_command(context, n_gif=0, thr_increment=None, resume_training=False, no_patch=False, overlap_2d=None)

Run main command.

This function is central in the ivadomed project as training / testing / evaluation commands are run via this function. All the process parameters are defined in the config.

Parameters:

• context (dict) – Dictionary containing all parameters that are needed for a given process. See Configuration File for more details.

• n_gif (int) – Generates a GIF during training if larger than zero, one frame per epoch for a given slice. The parameter indicates the number of 2D slices used to generate GIFs, one GIF per slice. A GIF shows predictions of a given slice from the validation sub-dataset. They are saved within the output path.

• thr_increment (float) – A threshold analysis is performed at the end of the training using the trained model and the training + validation sub-dataset to find the optimal binarization threshold. The specified value indicates the increment between 0 and 1 used during the ROC analysis (e.g. 0.1).

• resume_training (bool) – Load a saved model (“checkpoint.pth.tar” in the output directory specified with flag “–path-output” or via the config file “output_path”. This training state is saved everytime a new best model is saved in the log argument) for resume training directory.

• no_patch (bool) – If True, 2D patches are not used while segmenting with models trained with patches (command “–segment” only). Default: False (i.e. segment with patches). The “no_patch” option supersedes the “overlap_2D” option.

• overlap_2d (list of int) – Custom overlap for 2D patches while segmenting (command “–segment” only). Default model overlap is used otherwise.

Returns:

• If “train” command: Returns floats: best loss score for both training and validation.

• If “test” command: Returns a pandas Dataframe: of metrics computed for each subject of the testing sub-dataset and return the prediction metrics before evaluation.

• If “segment” command: No return value.

Return type:

float or pandas.DataFrame or None

Metrics API

class MetricManager(metric_fns)

Bases: object

Computes specified metrics and stores them in a dictionary.

Parameters:

metric_fns (list) – List of metric functions.

Attributes:

• metric_fns (list) – List of metric functions.

• result_dict (dict) – Dictionary storing metrics.

• num_samples (int) – Number of samples.

__init__(metric_fns)

__call__(prediction, ground_truth)

Call self as a function.

numeric_score(prediction, groundtruth)

Computation of statistical numerical scores:

• FP = Soft False Positives

• FN = Soft False Negatives

• TP = Soft True Positives

• TN = Soft True Negatives

Robust to hard or soft input masks. For example::

prediction=np.asarray([0, 0.5, 1]) groundtruth=np.asarray([0, 1, 1]) Leads to FP = 1.5

Note: It assumes input values are between 0 and 1.

Parameters:

• prediction (ndarray) – Binary prediction.

• groundtruth (ndarray) – Binary groundtruth.

Returns:

FP, FN, TP, TN

Return type:

float, float, float, float

dice_score(im1, im2, empty_score=nan)

Computes the Dice coefficient between im1 and im2.

Compute a soft Dice coefficient between im1 and im2, ie equals twice the sum of the two masks product, divided by the sum of each mask sum. If both images are empty, then it returns empty_score.

Parameters:

• im1 (ndarray) – First array.

• im2 (ndarray) – Second array.

• empty_score (float) – Returned value if both input array are empty.

Returns:

Dice coefficient.

Return type:

float

mse(im1, im2)

Compute the Mean Squared Error.

Compute the Mean Squared Error between the two images, i.e. sum of the squared difference.

Parameters:

• im1 (ndarray) – First array.

• im2 (ndarray) – Second array.

Returns:

Mean Squared Error.

Return type:

float

hausdorff_score(prediction, groundtruth)

Compute the directed Hausdorff distance between two N-D arrays.

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

Returns:

Hausdorff distance.

Return type:

float

precision_score(prediction, groundtruth, err_value=0.0)

Positive predictive value (PPV).

Precision equals the number of true positive voxels divided by the sum of true and false positive voxels. True and false positives are computed on soft masks, see "numeric_score".

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

• err_value (float) – Value returned in case of error.

Returns:

Precision score.

Return type:

float

recall_score(prediction, groundtruth, err_value=0.0)

True positive rate (TPR).

Recall equals the number of true positive voxels divided by the sum of true positive and false negative voxels. True positive and false negative values are computed on soft masks, see "numeric_score".

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

• err_value (float) – Value returned in case of error.

Returns:

Recall score.

Return type:

float

specificity_score(prediction, groundtruth, err_value=0.0)

True negative rate (TNR).

Specificity equals the number of true negative voxels divided by the sum of true negative and false positive voxels. True negative and false positive values are computed on soft masks, see "numeric_score".

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

• err_value (float) – Value returned in case of error.

Returns:

Specificity score.

Return type:

float

intersection_over_union(prediction, groundtruth, err_value=0.0)

Intersection of two (soft) arrays over their union (IoU).

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

• err_value (float) – Value returned in case of error.

Returns:

IoU.

Return type:

float

accuracy_score(prediction, groundtruth, err_value=0.0)

Accuracy.

Accuracy equals the number of true positive and true negative voxels divided by the total number of voxels. True positive/negative and false positive/negative values are computed on soft masks, see "numeric_score".

Parameters:

• prediction (ndarray) – First array.

• groundtruth (ndarray) – Second array.

Returns:

Accuracy.

Return type:

float

multi_class_dice_score(im1, im2)

Dice score for multi-label images.

Multi-class Dice score equals the average of the Dice score for each class. The first dimension of the input arrays is assumed to represent the classes.

Parameters:

• im1 (ndarray) – First array.

• im2 (ndarray) – Second array.

Returns:

Multi-class dice.

Return type:

float

plot_roc_curve(tpr, fpr, opt_thr_idx, fname_out)

Plot ROC curve.

Parameters:

• tpr (list) – True positive rates.

• fpr (list) – False positive rates.

• opt_thr_idx (int) – Index of the optimal threshold.

• fname_out (str) – Output filename.

plot_dice_thr(thr_list, dice_list, opt_thr_idx, fname_out)

Plot Dice results against thresholds.

Parameters:

• thr_list (list) – Thresholds list.

• dice_list (list) – Dice results.

• opt_thr_idx (int) – Index of the optimal threshold.

• fname_out (str) – Output filename.

Postprocessing API

nifti_capable(wrapped)

Decorator to make a given function compatible with input being Nifti objects.

Parameters:

wrapped – Given function.

Returns:

Functions’ return.

binarize_with_low_threshold(wrapped)

Decorator to set low values (< 0.001) to 0.

Parameters:

wrapped – Given function.

Returns:

Functions’ return.

multilabel_capable(wrapped)

Decorator to make a given function compatible multilabel images.

Parameters:

wrapped – Given function.

Returns:

Functions’ return.

threshold_predictions(predictions, thr=0.5)

Threshold a soft (i.e. not binary) array of predictions given a threshold value, and returns a binary array.

Parameters:

• predictions (ndarray or nibabel object) – Image to binarize.

• thr (float) – Threshold value: voxels with a value < to thr are assigned 0 as value, 1 otherwise.

Returns:

ndarray or nibabel (same object as the input) containing only zeros or ones. Output type is int.

Return type:

ndarray

keep_largest_object(predictions)

Keep the largest connected object from the input array (2D or 3D).

Parameters:

predictions (ndarray or nibabel object) – Input segmentation. Image could be 2D or 3D.

Returns:

ndarray or nibabel (same object as the input).

keep_largest_object_per_slice(predictions, axis=2)

Keep the largest connected object for each 2D slice, along a specified axis.

Parameters:

• predictions (ndarray or nibabel object) – Input segmentation. Image could be 2D or 3D.

• axis (int) – 2D slices are extracted along this axis.

Returns:

ndarray or nibabel (same object as the input).

fill_holes(predictions, structure=(3, 3, 3))

Fill holes in the predictions using a given structuring element. Note: This function only works for binary segmentation.

Parameters:

• predictions (ndarray or nibabel object) – Input binary segmentation. Image could be 2D or 3D.

• structure (tuple of integers) – Structuring element, number of ints equals number of dimensions in the input array.

Returns:

ndrray or nibabel (same object as the input). Output type is int.

mask_predictions(predictions, mask_binary)

Mask predictions using a binary mask: sets everything outside the mask to zero.

Parameters:

• predictions (ndarray or nibabel object) – Input binary segmentation. Image could be 2D or 3D.

• mask_binary (ndarray) – Numpy array with the same shape as predictions, containing only zeros or ones.

Returns:

ndarray or nibabel (same object as the input).

coordinate_from_heatmap(nifti_image, thresh=0.3)

Retrieve coordinates of local maxima in a soft segmentation. :param nifti_image: nifti image of the soft segmentation. :type nifti_image: nibabel object :param thresh: Relative threshold for local maxima, i.e., after normalizing :type thresh: float :param the min and max between 0 and 1: :param respectively.:

Returns:

A list of computed coordinates found by local maximum. each element will be a list composed of [x, y, z]

Return type:

list

label_file_from_coordinates(nifti_image, coord_list)

Creates a nifti object with single-voxel labels. Each label has a value of 1. The nifti object as the same orientation as the input. :param nifti_image: Path to the image which affine matrix will be used to generate a new image with :type nifti_image: nibabel object :param labels.: :param coord_list: list of coordinates. Each element is [x, y, z]. Orientation should be the same as the image :type coord_list: list

Returns:

A nifti object containing the singe-voxel label of value 1. The matrix will be the same size as nifti_image.

Return type:

nib_pred

remove_small_objects(data, bin_structure, size_min)

Removes all unconnected objects smaller than the minimum specified size.

Parameters:

• data (ndarray) – Input data.

• bin_structure (ndarray) – Structuring element that defines feature connections.

• size_min (int) – Minimal object size to keep in input data.

Returns:

Array with small objects.

Return type:

ndarray

class Postprocessing(postprocessing_params, data_pred, dim_lst, filename_prefix)

Bases: object

Postprocessing steps manager

Parameters:

• postprocessing_params (dict) – Indicates postprocessing steps (in the right order)

• data_pred (ndarray) – Prediction from the model.

• dim_lst (list) – Dimensions of a voxel in mm.

• filename_prefix (str) – Path to prediction file without suffix.

Attributes:

• postprocessing_params (dict) – Indicates postprocessing steps (in the right order)

• data_pred (ndarray) – Prediction from the model.

• px (float) – Resolution (mm) along the first axis.

• py (float) – Resolution (mm) along the second axis.

• pz (float) – Resolution (mm) along the third axis.

• filename_prefix (str) – Path to prediction file without suffix.

• n_classes (int) – Number of classes.

• bin_struct (ndarray) – Binary structure.

__init__(postprocessing_params, data_pred, dim_lst, filename_prefix)

apply()

Parse postprocessing parameters and apply postprocessing steps to data.

binarize_prediction(thr)

Binarize output.

binarize_maxpooling()

Binarize by setting to 1 the voxel having the max prediction across all classes.

uncertainty(thr, suffix)

Removes the most uncertain predictions.

Parameters:

• thr (float) – Uncertainty threshold.

• suffix (str) – Suffix of uncertainty filename.

remove_small(unit, thr)

Remove small objects

Parameters:

• unit (str) – Indicates the units of the objects: “mm3” or “vox”

• thr (int or list) – Minimal object size to keep in input data.

fill_holes()

Fill holes in the predictions

keep_largest()

Keep largest object in prediction

remove_noise(thr)

Remove prediction values under the given threshold

Parameters:

thr (float) – Threshold under which predictions are set to 0.

Testing API

test(model_params, dataset_test, testing_params, path_output, device, cuda_available=True, metric_fns=None, postprocessing=None)

Main command to test the network.

Parameters:

• model_params (dict) – Model’s parameters.

• dataset_test (imed_loader) – Testing dataset.

• testing_params (dict) – Testing parameters.

• path_output (str) – Folder where predictions are saved.

• device (torch.device) – Indicates the CPU or GPU ID.

• cuda_available (bool) – If True, CUDA is available.

• metric_fns (list) – List of metrics, see ivadomed.metrics.

• postprocessing (dict) – Contains postprocessing steps.

Returns:

result metrics.

Return type:

dict

run_inference(test_loader, model, model_params, testing_params, ofolder, cuda_available, i_monte_carlo=None, postprocessing=None)

Run inference on the test data and save results as nibabel files.

Parameters:

• test_loader (torch DataLoader) –

• model (nn.Module) –

• model_params (dict) –

• testing_params (dict) –

• ofolder (str) – Folder where predictions are saved.

• cuda_available (bool) – If True, CUDA is available.

• i_monte_carlo (int) – i_th Monte Carlo iteration.

• postprocessing (dict) – Indicates postprocessing steps.

Returns:

Prediction, Ground-truth of shape n_sample, n_label, h, w, d.

Return type:

ndarray, ndarray

threshold_analysis(model_path, ds_lst, model_params, testing_params, metric='dice', increment=0.1, fname_out='thr.png', cuda_available=True)

Run a threshold analysis to find the optimal threshold on a sub-dataset.

Parameters:

• model_path (str) – Model path.

• ds_lst (list) – List of loaders.

• model_params (dict) – Model’s parameters.

• testing_params (dict) – Testing parameters

• metric (str) – Choice between “dice” and “recall_specificity”. If “recall_specificity”, then a ROC analysis is performed.

• increment (float) – Increment between tested thresholds.

• fname_out (str) – Plot output filename.

• cuda_available (bool) – If True, CUDA is available.

Returns:

optimal threshold.

Return type:

float

get_gt(filenames)

Get ground truth data as numpy array.

Parameters:

filenames (list) – List of ground truth filenames, one per class.

Returns:

4D numpy array.

Return type:

ndarray

Training API

train(model_params, dataset_train, dataset_val, training_params, path_output, device, wandb_params=None, cuda_available=True, metric_fns=None, n_gif=0, resume_training=False, debugging=False)

Main command to train the network.

Parameters:

• model_params (dict) – Model’s parameters.

• dataset_train (imed_loader) – Training dataset.

• dataset_val (imed_loader) – Validation dataset.

• training_params (dict) –

• path_output (str) – Folder where log files, best and final models are saved.

• device (str) – Indicates the CPU or GPU ID.

• cuda_available (bool) – If True, CUDA is available.

• metric_fns (list) – List of metrics, see ivadomed.metrics.

• n_gif (int) – Generates a GIF during training if larger than zero, one frame per epoch for a given slice. The parameter indicates the number of 2D slices used to generate GIFs, one GIF per slice. A GIF shows predictions of a given slice from the validation sub-dataset. They are saved within the output path.

• resume_training (bool) – Load a saved model (“checkpoint.pth.tar” in the path_output) for resume training. This training state is saved everytime a new best model is saved in the log directory.

• debugging (bool) – If True, extended verbosity and intermediate outputs.

Returns:

best_training_dice, best_training_loss, best_validation_dice,

best_validation_loss.

Return type:

float, float, float, float

get_sampler(ds, balance_bool, metadata)

Get sampler.

Parameters:

• ds (BidsDataset) – BidsDataset object.

• balance_bool (bool) – If True, a sampler is generated that balance positive and negative samples.

Returns:

Returns BalancedSampler, Bool: Sampler and boolean for shuffling (set to False). Otherwise: Returns None and True.

Return type:

If balance_bool is True

get_scheduler(params, optimizer, num_epochs=0)

Get scheduler.

Parameters:

• params (dict) – scheduler parameters, see PyTorch documentation

• optimizer (torch optim) –

• num_epochs (int) – number of epochs.

Returns:

torch.optim, bool, which indicates if the scheduler is updated for each batch (True), or for each epoch (False).

get_loss_function(params)

Get Loss function.

Parameters:

params (dict) – See ivadomed.losses.

Returns:

imed_losses object.

get_metadata(metadata, model_params)

Get metadata during batch loop.

Parameters:

• metadata (batch) –

• model_params (dict) –

Returns:

If FiLMedUnet, Returns a list of metadata, that have been transformed by the One Hot Encoder. If HeMISUnet, Returns a numpy array where each row represents a sample and each column represents a contrast.

load_checkpoint(model, optimizer, gif_dict, scheduler, fname)

Load checkpoint.

This function check if a checkpoint is available. If so, it updates the state of the input objects.

Parameters:

• model (nn.Module) – Init model.

• optimizer (torch.optim) – Model’s optimizer.

• gif_dict (dict) – Dictionary containing a GIF of the training.

• scheduler (_LRScheduler) – Learning rate scheduler.

• fname (str) – Checkpoint filename.

Returns:

nn.Module, torch, dict, int, float, _LRScheduler, int

Transformations API

multichannel_capable(wrapped)

Decorator to make a given function compatible multichannel images.

Parameters:

wrapped – Given function.

Returns:

Functions’ return.

two_dim_compatible(wrapped)

Decorator to make a given function compatible 2D or 3D images.

Parameters:

wrapped – Given function.

Returns:

Functions’ return.

class ImedTransform

Bases: object

Base class for transforamtions.

__call__(sample, metadata=None)

Call self as a function.

class Compose(dict_transforms, requires_undo=False)

Bases: object

Composes transforms together.

Composes transforms together and split between images, GT and ROI.

self.transform is a dict:

• keys: “im”, “gt” and “roi”

• values torchvision_transform.Compose objects.

Attributes:

• dict_transforms (dict) – Dictionary where the keys are the transform names and the value their parameters.

• requires_undo (bool) – If True, does not include transforms which do not have an undo_transform implemented yet.

Parameters:

transform (dict) – Keys are “im”, “gt”, “roi” and values are torchvision_transforms.Compose of the transformations of interest.

__init__(dict_transforms, requires_undo=False)

__call__(sample, metadata, data_type='im', preprocessing=False)

Call self as a function.

class UndoCompose(compose)

Bases: object

Undo the Compose transformations.

Call the undo transformations in the inverse order than the “do transformations”.

Attributes:

compose (torchvision_transforms.Compose)

Parameters:

transforms (torchvision_transforms.Compose) –

__init__(compose)

__call__(sample, metadata, data_type='gt')

Call self as a function.

class UndoTransform(transform)

Bases: object

Call undo transformation.

Attributes:

transform (ImedTransform)

Parameters:

transform (ImedTransform) –

__init__(transform)

__call__(sample)

Call self as a function.

class NumpyToTensor

Bases: ImedTransform

Converts nd array to tensor object.

undo_transform(sample, metadata=None)

Converts Tensor to nd array.

__call__(sample, metadata=None)

Converts nd array to Tensor.

class Resample(hspace, wspace, dspace=1.0)

Bases: ImedTransform

Resample image to a given resolution.

Parameters:

• hspace (float) – Resolution along the first axis, in mm.

• wspace (float) – Resolution along the second axis, in mm.

• dspace (float) – Resolution along the third axis, in mm.

• interpolation_order (int) – Order of spline interpolation. Set to 0 for label data. Default=2.

__init__(hspace, wspace, dspace=1.0)

undo_transform(sample, metadata=None)

Resample to original resolution.

__call__(sample, metadata=None)

Resample to a given resolution, in millimeters.

class NormalizeInstance

Bases: ImedTransform

Normalize a tensor or an array image with mean and standard deviation estimated from the sample itself.

__call__(sample, metadata=None)

Call self as a function.

class CroppableArray

Bases: ndarray

Zero padding slice past end of array in numpy.

Adapted From: https://stackoverflow.com/a/41155020/13306686

__getitem__(item)

Return self[key].

class Crop(size)

Bases: ImedTransform

Crop data.

Parameters:

size (tuple of int) – Size of the output sample. Tuple of size 2 if dealing with 2D samples, 3 with 3D samples.

Attributes:

size (tuple of int) – Size of the output sample. Tuple of size 3.

__init__(size)

__call__(sample, metadata)

Call self as a function.

class CenterCrop(size)

Bases: Crop

Make a centered crop of a specified size.

__call__(sample, metadata=None)

Call self as a function.

class ROICrop(size)

Bases: Crop

Make a crop of a specified size around a Region of Interest (ROI).

__call__(sample, metadata=None)

Call self as a function.

class DilateGT(dilation_factor)

Bases: ImedTransform

Randomly dilate a ground-truth tensor.

Parameters:

• dilation_factor (float) – Controls the number of dilation iterations. For each individual lesion, the number of

• follows (dilation iterations is computed as) – nb_it = int(round(dilation_factor * sqrt(lesion_area)))

• 0 (If dilation_factor <=) –

• performed. (then no dilation will be) –

__init__(dilation_factor)

__call__(sample, metadata=None)

Call self as a function.

class BoundingBoxCrop(size)

Bases: Crop

Crops image according to given bounding box.

__call__(sample, metadata)

Call self as a function.

class RandomAffine(degrees=0, translate=None, scale=None)

Bases: ImedTransform

Apply Random Affine transformation.

Parameters:

• degrees (float) – Positive float or list (or tuple) of length two. Angles in degrees. If only a float is provided, then rotation angle is selected within the range [-degrees, degrees]. Otherwise, the list / tuple defines this range.

• translate (list of float) – List of floats between 0 and 1, of length 2 or 3 depending on the sample shape (2D or 3D). These floats defines the maximum range of translation along each axis.

• scale (list of float) – List of floats between 0 and 1, of length 2 or 3 depending on the sample shape (2D or 3D). These floats defines the maximum range of scaling along each axis.

Attributes:

• degrees (tuple of floats)

• translate (list of float)

• scale (list of float)

__init__(degrees=0, translate=None, scale=None)

__call__(sample, metadata=None)

Call self as a function.

class RandomReverse

Bases: ImedTransform

Make a randomized symmetric inversion of the different values of each dimensions.

__call__(sample, metadata=None)

Call self as a function.

class RandomShiftIntensity(shift_range, prob=0.1)

Bases: ImedTransform

Add a random intensity offset.

Parameters:

• shift_range (tuple of floats) – Tuple of length two. Specifies the range where the offset that is applied is randomly selected from.

• prob (float) – Between 0 and 1. Probability of occurence of this transformation.

__init__(shift_range, prob=0.1)

__call__(sample, metadata=None)

Call self as a function.

class ElasticTransform(alpha_range, sigma_range, p=0.1)

Bases: ImedTransform

Applies elastic transformation.

See also

Simard, Patrice Y., David Steinkraus, and John C. Platt. “Best practices for convolutional neural networks applied to visual document analysis.” Icdar. Vol. 3. No. 2003. 2003.

Parameters:

• alpha_range (tuple of floats) – Deformation coefficient. Length equals 2.

• sigma_range (tuple of floats) – Standard deviation. Length equals 2.

__init__(alpha_range, sigma_range, p=0.1)

__call__(sample, metadata=None)

Call self as a function.

class AdditiveGaussianNoise(mean=0.0, std=0.01)

Bases: ImedTransform

Adds Gaussian Noise to images.

Parameters:

• mean (float) – Gaussian noise mean.

• std (float) – Gaussian noise standard deviation.

__init__(mean=0.0, std=0.01)

__call__(sample, metadata=None)

Call self as a function.

class Clahe(clip_limit=3.0, kernel_size=(8, 8))

Bases: ImedTransform

Applies Contrast Limited Adaptive Histogram Equalization for enhancing the local image contrast.

See also

Zuiderveld, Karel. “Contrast limited adaptive histogram equalization.” Graphics gems (1994): 474-485.

Default values are based on:

See also

Zheng, Qiao, et al. “3-D consistent and robust segmentation of cardiac images by deep learning with spatial propagation.” IEEE transactions on medical imaging 37.9 (2018): 2137-2148.

Parameters:

• clip_limit (float) – Clipping limit, normalized between 0 and 1.

• kernel_size (tuple of int) – Defines the shape of contextual regions used in the algorithm. Length equals image

• dimension (ie 2 or 3 for 2D or 3D, respectively) –

__init__(clip_limit=3.0, kernel_size=(8, 8))

__call__(sample, metadata=None)

Call self as a function.

class HistogramClipping(min_percentile=5.0, max_percentile=95.0)

Bases: ImedTransform

Performs intensity clipping based on percentiles.

Parameters:

• min_percentile (float) – Between 0 and 100. Lower clipping limit.

• max_percentile (float) – Between 0 and 100. Higher clipping limit.

__init__(min_percentile=5.0, max_percentile=95.0)

__call__(sample, metadata=None)

Call self as a function.

class RandomGamma(log_gamma_range, p=0.5)

Bases: ImedTransform

Randomly changes the contrast of an image by gamma exponential

Parameters:

• log_gamma_range (tuple of floats) – Log gamma range for changing contrast. Length equals 2.

• p (float) – Probability of performing the gamma contrast

__init__(log_gamma_range, p=0.5)

__call__(sample, metadata=None)

Call self as a function.

class RandomBiasField(coefficients, order, p=0.5)

Bases: ImedTransform

Applies a random MRI bias field artifact to the image via torchio.RandomBiasField()

Parameters:

• coefficients (float) – Maximum magnitude of polynomial coefficients

• order – Order of the basis polynomial functions

• p (float) – Probability of applying the bias field

__init__(coefficients, order, p=0.5)

__call__(sample, metadata=None)

Call self as a function.

class RandomBlur(sigma_range, p=0.5)

Bases: ImedTransform

Applies a random blur to the image

Parameters:

• sigma_range (tuple of floats) – Standard deviation range for the gaussian filter

• p (float) – Probability of performing blur

__init__(sigma_range, p=0.5)

__call__(sample, metadata=None)

Call self as a function.

get_subdatasets_transforms(transform_params)

Get transformation parameters for each subdataset: training, validation and testing.

Parameters:

transform_params (dict) –

Returns:

Training, Validation and Testing transformations.

Return type:

dict, dict, dict

get_preprocessing_transforms(transforms)

Checks the transformations parameters and selects the transformations which are done during preprocessing only.

Parameters:

transforms (dict) – Transformation dictionary.

Returns:

Preprocessing transforms.

Return type:

dict

apply_preprocessing_transforms(transforms, seg_pair, roi_pair=None) → Tuple[dict, dict]

Applies preprocessing transforms to segmentation pair (input, gt and metadata).

Parameters:

• transforms (Compose) – Preprocessing transforms.

• seg_pair (dict) – Segmentation pair containing input and gt.

• roi_pair (dict) – Segementation pair containing input and roi.

Returns:

Segmentation pair and roi pair.

Return type:

tuple

prepare_transforms(transform_dict, requires_undo=True)

This function separates the preprocessing transforms from the others and generates the undo transforms related.

Parameters:

• transform_dict (dict) – Dictionary containing the transforms and there parameters.

• requires_undo (bool) – Boolean indicating if transforms can be undone.

Returns:

transform lst containing the preprocessing transforms and regular transforms, UndoCompose

object containing the transform to undo.

Return type:

list, UndoCompose

tio_transform(x, transform)

Applies TorchIO transformations to a given image and returns the transformed image and history.

Parameters:

• x (np.ndarray) – input image

• transform (tio.transforms.Transform) – TorchIO transform

Returns:

transformed image, history of parameters used for the applied transformation

Return type:

np.ndarray, list

Utils API

class Metavar(value)

Bases: Enum

This class is used to display intuitive input types via the metavar field of argparse.

initialize_wandb(wandb_params)

Initializes WandB and based upon the parameters sets it up or disables it for experimental tracking

Parameters:

wandb_params (dict) – wandb parameters

Returns:

True if wandb tracking is enabled

Return type:

bool, wandb_tracking

cuda(input_var, cuda_available=True, non_blocking=False)

Passes input_var to GPU.

Parameters:

• input_var (Tensor) – either a tensor or a list of tensors.

• cuda_available (bool) – If False, then return identity

• non_blocking (bool) –

Returns:

Tensor

unstack_tensors(sample)

Unstack tensors.

Parameters:

sample (Tensor) –

Returns:

list of Tensors.

Return type:

list

generate_sha_256(context: dict, df, file_lst: List[str]) → None

generate sha256 for a training file

Parameters:

• context (dict) – configuration context.

• df (pd.DataFrame) – Dataframe containing all BIDS image files indexed and their metadata.

• file_lst (List[str]) – list of strings containing training files

save_onnx_model(model, inputs, model_path)

Convert PyTorch model to ONNX model and save it as model_path.

Parameters:

• model (nn.Module) – PyTorch model.

• inputs (Tensor) – Tensor, used to inform shape and axes.

• model_path (str) – Output filename for the ONNX model.

define_device(gpu_id)

Define the device used for the process of interest.

Parameters:

gpu_id (int) – GPU ID.

Returns:

True if cuda is available.

Return type:

Bool, device

display_selected_model_spec(params)

Display in terminal the selected model and its parameters.

Parameters:

params (dict) – Keys are param names and values are param values.

display_selected_transfoms(params, dataset_type)

Display in terminal the selected transforms for a given dataset.

Parameters:

• params (dict) –

• dataset_type (list) – e.g. [‘testing’] or [‘training’, ‘validation’]

plot_transformed_sample(before, after, list_title=None, fname_out='', cmap='jet')

Utils tool to plot sample before and after transform, for debugging.

Parameters:

• before (ndarray) – Sample before transform.

• after (ndarray) – Sample after transform.

• list_title (list of str) – Sub titles of before and after, resp.

• fname_out (str) – Output filename where the plot is saved if provided.

• cmap (str) – Matplotlib colour map.

check_exe(name)

Ensure that a program exists.

Parameters:

name (str) – Name or path to program.

Returns:

path of the program or None

Return type:

str or None

exception ArgParseException

Bases: Exception

get_arguments(parser, args)

Get arguments from function input or command line.

Parameters:

• parser (argparse.ArgumentParser) – ArgumentParser object

• args (list) – either a list of arguments or None. The list should be formatted like this: [“-d”, “SOME_ARG”, “–model”, “SOME_ARG”]

format_path_data(path_data)

Parameters:

path_data (list or str) – Either a list of paths, or just one path.

Returns:

A list of paths

Return type:

list

similarity_score(a: str, b: str) → float

use DiffLIb SequenceMatcher to resolve the similarity between text. Help make better choice in terms of derivatives. :param a: a string :param b: another string

Returns: a score indicative of the similarity between the sequence.

init_ivadomed()

Initialize the ivadomed for typical terminal usage.

get_timestamp() → str

Return a datetime string in the format YYYY-MM-DDTHHMMSS.(sub-precision) Returns:

get_system_memory() → float

Return the system memory in GB. Returns:

get_win_system_memory() → float

Obtain the amount of memory available on Windows system. Returns: memory in GB Source: https://stackoverflow.com/a/21589439

get_linux_system_memory() → float

Obtain the amount of memory available on Linux system. Returns: memory in GB Source: https://stackoverflow.com/a/28161352

get_mac_system_memory() → float

Obtain the amount of memory available on MacOS system. Returns: memory in GB Source: https://apple.stackexchange.com/a/4296

Visualize API

overlap_im_seg(img, seg)

Overlap image (background, greyscale) and segmentation (foreground, jet).

class LoopingPillowWriter(fps=5, metadata=None, codec=None, bitrate=None)

Bases: PillowWriter

finish()

Finish any processing for writing the movie.

class AnimatedGif(size)

Bases: object

Generates GIF.

Parameters:

size (tuple) – Size of frames.

Attributes:

• fig (plt)

• size_x (int)

• size_y (int)

• images (list) – List of frames.

__init__(size)

save_color_labels(gt_data, binarize, gt_filename, output_filename, slice_axis)

Saves labels encoded in RGB in specified output file.

Parameters:

• gt_data (ndarray) – Input image with dimensions (Number of classes, height, width, depth).

• binarize (bool) – If True binarizes gt_data to 0 and 1 values, else soft values are kept.

• gt_filename (str) – GT path and filename.

• output_filename (str) – Name of the output file where the colored labels are saved.

• slice_axis (int) – Indicates the axis used to extract slices: “axial”: 2, “sagittal”: 0, “coronal”: 1.

Returns:

RGB labels.

Return type:

ndarray

convert_labels_to_RGB(grid_img)

Converts 2D images to RGB encoded images for display in tensorboard.

Parameters:

grid_img (Tensor) – GT or prediction tensor with dimensions (batch size, number of classes, height, width).

Returns:

RGB image with shape (height, width, 3).

Return type:

tensor

save_img(writer, epoch, dataset_type, input_samples, gt_samples, preds, wandb_tracking=False, is_three_dim=False)

Saves input images, gt and predictions in tensorboard (and wandb depending upon the inputs in the config file).

Parameters:

• writer (SummaryWriter) – Tensorboard’s summary writer.

• epoch (int) – Epoch number.

• dataset_type (str) – Choice between Training or Validation.

• input_samples (Tensor) – Input images with shape (batch size, number of channel, height, width, depth) if 3D else (batch size, number of channel, height, width)

• gt_samples (Tensor) – GT images with shape (batch size, number of channel, height, width, depth) if 3D else (batch size, number of channel, height, width)

• preds (Tensor) – Model’s prediction with shape (batch size, number of channel, height, width, depth) if 3D else (batch size, number of channel, height, width)

• is_three_dim (bool) – True if 3D input, else False.

save_feature_map(batch, layer_name, path_output, model, test_input, slice_axis)

Save model feature maps.

Parameters:

• batch (dict) –

• layer_name (str) –

• path_output (str) – Output folder.

• model (nn.Module) – Network.

• test_input (Tensor) –

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

class HookBasedFeatureExtractor(submodule, layername, upscale=False)

Bases: Module

This function extracts feature maps from given layer. Helpful to observe where the attention of the network is focused.

https://github.com/ozan-oktay/Attention-Gated-Networks/tree/a96edb72622274f6705097d70cfaa7f2bf818a5a

Parameters:

• submodule (nn.Module) – Trained model.

• layername (str) – Name of the layer where features need to be extracted (layer of interest).

• upscale (bool) – If True output is rescaled to initial size.

Attributes:

• submodule (nn.Module) – Trained model.

• layername (str) – Name of the layer where features need to be extracted (layer of interest).

• outputs_size (list) – List of output sizes.

• outputs (list) – List of outputs containing the features of the given layer.

• inputs (list) – List of inputs.

• inputs_size (list) – List of input sizes.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

__init__(submodule, layername, upscale=False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Inference API

onnx_inference(model_path: str, inputs: tensor) → tensor

Run ONNX inference

Parameters:

• model_path (str) – Path to the ONNX model.

• inputs (Tensor) – Batch of input image.

Returns:

Network output.

Return type:

Tensor

get_preds(context: dict, fname_model: str, model_params: dict, cuda_available: bool, device: device, batch: dict) → tensor

Returns the predictions from the given model.

Parameters:

• context (dict) – configuration dict.

• fname_model (str) – name of file containing model.

• model_params (dict) – dictionary containing model parameters.

• cuda_available (bool) – True if cuda is available.

• device (torch.device) – Device used for prediction.

• batch (dict) – dictionary containing input, gt and metadata

Returns:

predictions from the model.

Return type:

tensor

get_onehotencoder(context: dict, folder_model: str, options: dict, ds: Dataset) → dict

Returns one hot encoder which is needed to update the model parameters when FiLMedUnet is applied.

Parameters:

• context (dict) – Configuration dict.

• folder_model (str) – Foldername which contains trained model and its configuration file.

• options (dict) – Contains film metadata information.

• ds (Dataset) – Dataset used for the segmentation.

Returns:

onehotencoder used in the model params.

Return type:

dict

pred_to_nib(data_lst: List[ndarray], z_lst: List[int], fname_ref: str, fname_out: str, slice_axis: int, debug: bool = False, kernel_dim: str = '2d', bin_thr: float = 0.5, discard_noise: bool = True, postprocessing: dict | None = None) → Nifti1Image

Save the network predictions as nibabel object.

Based on the header of fname_ref image, it creates a nibabel object from the Network predictions (data_lst).

Parameters:

• data_lst (list of np arrays) – Predictions, either 2D slices either 3D patches.

• z_lst (list of ints) – Slice indexes to reconstruct a 3D volume for 2D slices.

• fname_ref (str) – Filename of the input image: its header is copied to the output nibabel object.

• fname_out (str) – If not None, then the generated nibabel object is saved with this filename.

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

• debug (bool) – If True, extended verbosity and intermediate outputs.

• kernel_dim (str) – Indicates whether the predictions were done on 2D or 3D patches. Choices: ‘2d’, ‘3d’.

• bin_thr (float) – If positive, then the segmentation is binarized with this given threshold. Otherwise, a soft segmentation is output.

• discard_noise (bool) – If True, predictions that are lower than 0.01 are set to zero.

• postprocessing (dict) – Contains postprocessing steps to be applied.

Returns:

NiBabel object containing the Network prediction.

Return type:

nibabel.Nifti1Image

pred_to_png(pred_list: list, target_list: list, subj_path: str, suffix: str = '', max_value: int = 1)

Save the network predictions as PNG files with suffix “_target_pred”.

Parameters:

• pred_list (list of np arrays) – list of 2D predictions.

• target_list (list of str) – list of target suffixes.

• subj_path (str) – Path of the subject filename in output folder without extension (e.g. “path_output/pred_masks/sub-01_sample-01_SEM”).

• suffix (str) – additional suffix to append to the filename (e.g. “_pred.png”)

• max_value (int) – Maximum mask value of the float mask to use during the conversion to uint8.

process_transformations(context: dict, fname_roi: str, fname_prior: str, metadata: dict, slice_axis: int, fname_images: list) → dict

Sets the transformation based on context parameters. When ROI is not provided center-cropping is applied.

If there is an object_detection_path, then we modify the metadata to store transformation data.

Parameters:

• context (dict) – configuration dictionary.

• fname_roi (str) – filename containing region for cropping image prior to segmentation.

• fname_prior (str) – prior image filename.

• metadata (dict) – metadata used in setting bounding box when we have object_detection_params.

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

• fname_images (list) – list of image filenames (e.g. .nii.gz) to segment.

Returns:

metadata.

Return type:

dict

set_option(options: dict, postpro: dict, context: dict, key: str)

Generalized function that sets postprocessing option based on given list of options.

When given key already exists in options, we initialize the key value for the postprocessing dictionary Otherwise, when the key is already found in the postprocessing attritute of the context, we remove it

Parameters:

• options (dict) – Contains postprocessing steps information.

• postpro (dict) – Postprocessing settings.

• context (dict) – Configuration dict.

• key (str) – The key of the postprocessing option we wish to set.

Returns:

postprocessing settings.

Return type:

dict

set_postprocessing_options(options: dict, context: dict)

Updates the postprocessing options based on existing settings found in options.

Parameters:

• options (dict) – Contains postprocessing steps information.

• context (dict) – Configuration dict.

segment_volume(folder_model: str, fname_images: list, gpu_id: int = 0, options: dict | None = None)

Segment an image.

Segment an image (fname_image) using a pre-trained model (folder_model). If provided, a region of interest (fname_roi) is used to crop the image prior to segment it.

Parameters:

• folder_model (str) – foldername which contains (1) the model (‘folder_model/folder_model.pt’) to use (2) its configuration file (‘folder_model/folder_model.json’) used for the training, see https://github.com/neuropoly/ivadomed/wiki/configuration-file

• fname_images (list) – list of image filenames (e.g. .nii.gz) to segment. Multichannel models require multiple images to segment, e.i., len(fname_images) > 1.

• gpu_id (int) – Number representing gpu number if available. Currently does NOT support multiple GPU segmentation.

• options (dict) –

  This can optionally contain any of the following key-value pairs:

  • ’binarize_prediction’: (float) Binarize segmentation with specified threshold. Predictions below the threshold become 0, and predictions above or equal to threshold become 1. Set to -1 for no thresholding (i.e., soft segmentation).

  • ’binarize_maxpooling’: (bool) Binarize by setting to 1 the voxel having the maximum prediction across all classes. Useful for multiclass models.

  • ’fill_holes’: (bool) Fill small holes in the segmentation.

  • ’keep_largest’: (bool) Keep the largest connected-object for each class from the output segmentation.

  • ’remove_small’: (list of str) Minimal object size to keep with unit (mm3 or vox). A single value can be provided or one value per prediction class. Single value example: [“1mm3”], [“5vox”]. Multiple values example: [“10”, “20”, “10vox”] (remove objects smaller than 10 voxels for class 1 and 3, and smaller than 20 voxels for class 2).

  • ’pixel_size’: (list of float) List of microscopy pixel size in micrometers. Length equals 2 [PixelSizeX, PixelSizeY] for 2D or 3 [PixelSizeX, PixelSizeY, PixelSizeZ] for 3D, where X is the width, Y the height and Z the depth of the image.

  • ’pixel_size_units’: (str) Units of pixel size (Must be either “mm”, “um” or “nm”)

  • ’no_patch’: (bool) 2D patches are not used while segmenting with models trained with patches. The “no_patch” option supersedes the “overlap_2D” option. This option may not be suitable with large images depending on computer RAM capacity.

  • ’overlap_2D’: (list of int) List of overlaps in pixels for 2D patching. Length equals 2 [OverlapX, OverlapY], where X is the width and Y the height of the image.

  • ’metadata’: (str) Film metadata.

  • ’fname_prior’: (str) An image filename (e.g., .nii.gz) containing processing information (e.g., spinal cord segmentation, spinal location or MS lesion classification, spinal cord centerline), used to crop the image prior to segment it if provided. The segmentation is not performed on the slices that are empty in this image.

Returns:

List of nibabel objects containing the soft segmentation(s), one per prediction class, List of target suffix associated with each prediction in pred_list

Return type:

list, list

split_classes(nib_prediction)

Split a 4D nibabel multi-class segmentation file in multiple 3D nibabel binary segmentation files.

Parameters:

nib_prediction (nibabelObject) – 4D nibabel object.

Returns:

list of nibabelObject.

reconstruct_3d_object(context: dict, batch: dict, undo_transforms: UndoCompose, preds: tensor, preds_list: list, kernel_3D: bool, is_2d_patch: bool, slice_axis: int, slice_idx_list: list, data_loader: DataLoader, fname_images: list, i_batch: int, last_sample_bool: bool, weight_matrix: tensor, volume: tensor, image: tensor)

Reconstructs the 3D object from the current batch, and returns the list of predictions and targets.

Parameters:

• context (dict) – configuration dict.

• batch (dict) – Dictionary containing input, gt and metadata

• undo_transforms (UndoCompose) – Undo transforms so prediction match original image resolution and shape

• preds (tensor) – Subvolume predictions

• preds_list (list of tensor) – list of subvolume predictions.

• kernel_3D (bool) – true when using 3D kernel.

• is_2d_patch (bool) – Indicates if 2d patching is used.

• slice_axis (int) – Indicates the axis used for the 2D slice extraction: Sagittal: 0, Coronal: 1, Axial: 2.

• slice_idx_list (list of int) – list of indices for the axis slices.

• data_loader (DataLoader) – DataLoader object containing batches using in object construction.

• fname_images (list) – list of image filenames (e.g. .nii.gz) to segment.

• i_batch (int) – index of current batch.

• last_sample_bool (bool) – flag to indicate whether this is the last sample in the 3D volume

• weight_matrix (tensor) – the weight matrix

• volume (tensor) – the volume tensor that is being partially reconstructed through the loop

• image (tensor) – the image tensor that is being partially reconstructed through the loop

Returns:

list of predictions target_list (list): list of targets last_sample_bool (bool): flag to indicate whether this is the last sample in the 3D volume weight_matrix (tensor): the weight matrix. Must be returned as passing tensor by reference is NOT reliable. volume (tensor): the volume tensor that is being partially reconstructed through the loop. Must be returned as passing tensor by reference is NOT reliable. image (tensor): the vimage tensor that is being partially reconstructed through the loop. Must be returned as passing tensor by reference is NOT reliable.

Return type:

pred_list (list)

volume_reconstruction(batch: dict, pred: tensor, undo_transforms: UndoCompose, smp_idx: int, volume: tensor | None = None, weight_matrix: tensor | None = None)

Reconstructs volume prediction from subvolumes used during training :param batch: Dictionary containing input, gt and metadata :type batch: dict :param pred: Subvolume prediction :type pred: tensor :param undo_transforms: Undo transforms so prediction match original image resolution and shap :type undo_transforms: UndoCompose :param smp_idx: Batch index :type smp_idx: int :param volume: Reconstructed volume :type volume: tensor :param weight_matrix: Weights containing the number of predictions for each voxel :type weight_matrix: tensor

Returns:

undone subvolume, metadata (dict): metadata, last_sample_bool (bool): boolean representing if its the last sample of the volume volume (tensor): representing the volume reconstructed weight_matrix (tensor): weight matrix

Return type:

pred_undo (tensor)

image_reconstruction(batch: dict, pred: tensor, undo_transforms: UndoCompose, smp_idx: int, image: tensor | None = None, weight_matrix: tensor | None = None)

Reconstructs image prediction from patches used during training :param batch: Dictionary containing input, gt and metadata :type batch: dict :param pred: Patch prediction :type pred: tensor :param undo_transforms: Undo transforms so prediction match original image resolution and shape :type undo_transforms: UndoCompose :param smp_idx: Batch index :type smp_idx: int :param image: Reconstructed image :type image: tensor :param weight_matrix: Weights containing the number of predictions for each pixel :type weight_matrix: tensor

Returns:

undone image metadata (dict): metadata last_patch_bool (bool): boolean representing if its the last patch of the image image (tensor): representing the image reconstructed weight_matrix (tensor): weight matrix

Return type:

pred_undo (tensor)

Mixup API

mixup(data, targets, alpha, debugging=False, ofolder=None)

Compute the mixup data.

See also

Zhang, Hongyi, et al. “mixup: Beyond empirical risk minimization.” arXiv preprint arXiv:1710.09412 (2017).

Parameters:

• data (Tensor) – Input images.

• targets (Tensor) – Input masks.

• alpha (float) – MixUp parameter.

• debugging (Bool) – If True, then samples of mixup are saved as png files.

• ofolder (str) – If debugging, output folder where “mixup” folder is created and samples are saved.

Returns:

Mixed image, Mixed mask.

Return type:

Tensor, Tensor

save_mixup_sample(ofolder, input_data, labeled_data, lambda_tensor)

Save mixup samples as png files in a “mixup” folder.

Parameters:

• ofolder (str) – Output folder where “mixup” folder is created and samples are saved.

• input_data (Tensor) – Input image.

• labeled_data (Tensor) – Input masks.

• lambda_tensor (Tensor) –

Uncertainty API

run_uncertainty(image_folder)

Compute uncertainty from model prediction.

This function loops across the model predictions (nifti masks) and estimates the uncertainty from the Monte Carlo samples. Both voxel-wise and structure-wise uncertainty are estimates.

Parameters:

image_folder (str) – Folder containing the Monte Carlo samples.

combine_predictions(fname_lst, fname_hard, fname_prob, thr=0.5)

Combine predictions from Monte Carlo simulations.

Combine predictions from Monte Carlo simulations and save the resulting as:

1. fname_prob, a soft segmentation obtained by averaging the Monte Carlo samples.

2. fname_hard, a hard segmentation obtained thresholding with thr.

Parameters:

• fname_lst (list of str) – List of the Monte Carlo samples.

• fname_hard (str) – Filename for the output hard segmentation.

• fname_prob (str) – Filename for the output soft segmentation.

• thr (float) – Between 0 and 1. Used to threshold the soft segmentation and generate the hard segmentation.

voxelwise_uncertainty(fname_lst, fname_out, eps=1e-05)

Estimate voxel wise uncertainty.

Voxel-wise uncertainty is estimated as entropy over all N MC probability maps, and saved in fname_out.

Parameters:

• fname_lst (list of str) – List of the Monte Carlo samples.

• fname_out (str) – Output filename.

• eps (float) – Epsilon value to deal with np.log(0).

structurewise_uncertainty(fname_lst, fname_hard, fname_unc_vox, fname_out)

Estimate structure wise uncertainty.

Structure-wise uncertainty from N MC probability maps (fname_lst) and saved in fname_out with the following suffixes:

• ‘-cv.nii.gz’: coefficient of variation

• ‘-iou.nii.gz’: intersection over union

• ‘-avgUnc.nii.gz’: average voxel-wise uncertainty within the structure.

Parameters:

• fname_lst (list of str) – List of the Monte Carlo samples.

• fname_hard (str) – Filename of the hard segmentation, which is used to compute the avgUnc by providing a mask of the structures.

• fname_unc_vox (str) – Filename of the voxel-wise uncertainty, which is used to compute the avgUnc.

• fname_out (str) – Output filename.

Maths API

rescale_values_array(arr, minv=0.0, maxv=1.0, dtype=<class 'numpy.float32'>)

Rescale the values of numpy array arr to be from minv to maxv.

Parameters:

• arr (ndarry) – Array whose values will be rescaled.

• minv (float) – Minimum value of the output array.

• maxv (float) – Maximum value of the output array.

• dtype (type) – Cast array to this type before performing the rescaling.

gaussian_kernel(kernlen=10)

Create a 2D gaussian kernel with user-defined size.

Parameters:

kernlen (int) – size of kernel

Returns:

a 2D array of size (kernlen,kernlen)

Return type:

ndarray

heatmap_generation(image, kernel_size)

Generate heatmap from image containing sing voxel label using convolution with gaussian kernel :param image: 2D array containing single voxel label :type image: ndarray :param kernel_size: size of gaussian kernel :type kernel_size: int

Returns:

2D array heatmap matching the label.

Return type:

ndarray