configuration_file.rst

Configuration File

All parameters used for loading data, training and predicting are contained within a single JSON configuration file. This section describes how to set up this configuration file.

For convenience, here is an generic configuration file: config_config.json.

Below are other, more specific configuration files:

• config_classification.json: Trains a classification model.
• config_sctTesting.json: Trains a 2D segmentation task with the U-Net architecture.
• config_spineGeHemis.json: Trains a segmentation task with the HeMIS-UNet architecture.
• config_tumorSeg.json: Trains a segmentation task with a 3D U-Net architecture.

General parameters

command

Run the specified command. Choices: "train" or "test", to train and evaluate a model respectively.

gpu

Integer. ID of the GPU to use.

log_directory

Folder name that will contain the output files (e.g., trained model, predictions, results).

model_name

Folder name containing the trained model (ONNX format) and its configuration file, located within "log_directory/", eg "log_directory/seg_gm_t2star/seg_gm_t2star.onnx" and "log_directory/seg_gm_t2star/seg_gm_t2star.json", respectively. When possible, the folder name will follow the following convention: task_(animal)_region_(contrast) with

task = {seg, label, find}
animal = {human, dog, cat, rat, mouse, ...}
region = {sc, gm, csf, brainstem, ...}
contrast = {t1, t2, t2star, dwi, ...}

debugging

Bool. Extended verbosity and intermediate outputs.

Loader parameters

bids_path

String. Path of the BIDS folder.

target_suffix

List. Suffix list of the derivative file containing the ground-truth of interest (e.g. ["_seg-manual", "_lesion-manual"]). The length of this list controls the number of output channels of the model (i.e. out_channel). If the list has a length greater than 1, then a multi-class model will be trained.

contrasts

• train_validation: List. List of image contrasts (e.g. T1w, T2w) loaded for the training and validation. If multichannel is true, this list represents the different channels of the input tensors (i.e. its length equals model's in_channel). Otherwise, the contrasts are mixed and the model has only one input channel (i.e. model's in_channel=1).
• test: List. List of image contrasts (e.g. T1w, T2w) loaded in the testing dataset. Same comment than for train_validation regarding multichannel.
• balance: Dict. Enables to weight the importance of specific channels (or contrasts) in the dataset: e.g. {"T1w": 0.1} means that only 10% of the available T1w images will be included into the training/validation/test set. Please set multichannel to false if you are using this parameter.

multichannel

Bool. Indicated if more than a contrast (e.g. T1w and T2w) is used by the model. See details in both train_validation and test for the contrasts that are input.

slice_axis

Choice between "sagittal", "coronal", and "axial". Sets the slice orientation for on which the model will be used.

slice_filter

Dict. Discard a slice from the dataset if it meets a condition, see below.

• filter_empty_input: Bool. Discard slices where all voxel intensities are zeros.
• filter_empty_mask: Bool. Discard slices where all voxel labels are zeros.

roi

Dict. of parameters about the region of interest

• suffix: String. Suffix of the derivative file containing the ROI used to crop (e.g. "_seg-manual") with ROICrop as transform. Please use null if you do not want to use an ROI to crop.
• slice_filter_roi: int. If the ROI mask contains less than slice_filter_roi non-zero voxels, the slice will be discarded from the dataset. This feature helps with noisy labels, e.g., if a slice contains only 2-3 labeled voxels, we do not want to use these labels to crop the image. This parameter is only considered when using "ROICrop".

soft_gt

Bool. Indicates if a soft mask will be used as ground-truth to train and / or evaluate a model. In particular, the masks are not binarized after interpolations implied by preprocessing or data-augmentation operations.

Split dataset

fname_split

String. File name of the log (joblib) that contains the list of training/validation/testing subjects. This file can later be used to re-train a model using the same data splitting scheme. If null, a new splitting scheme is performed.

random_seed

Int. Seed used by the random number generator to split the dataset between training/validation/testing. The use of the same seed ensures the same split between the sub-datasets, which is useful to reproduce results.

method

{"per_patient", "per_center"}. "per_patient": all subjects are shuffled, then split between train/validation/test according to "train_fraction" and "test_fraction", regardless their institution. "per_center": all subjects are split so as not to mix institutions between the train/validation/test sets according to "train_fraction" and "center_test". The latter option enables to ensure the model is working across domains (institutions). Note: the institution information is contained within the institution_id column in the participants.tsv file.

balance

String (Optional). Metadata contained in "participants.tsv" file with categorical values. Each category will be evenly distributed in the training, validation and testing datasets.

train_fraction

Float. Between 0 and 1 representing the fraction of the dataset used as training set.

test_fraction

Float. Between 0 and 1 representing the fraction of the dataset used as test set. This parameter is only used if the method is "per_patient".

center_test

List of strings. Each string corresponds to an institution/center to only include in the testing dataset (not validation). This parameter is only used if the method is "per_center". If used, the file bids_dataset/participants.tsv needs to contain a column institution_id, which associates a subject with an institution/center.

Training parameters

batch_size

Strictly positive integer.

loss

• name: Name of the loss function class. See ivadomed.losses
• Other parameters that could be needed in the Loss function definition: see attributes of the Loss function of interest (e.g. "gamma": 0.5 for FocalLoss).

training_time

• num_epochs: Strictly positive integer.
• early_stopping_epsilon: Float. If the validation loss difference during one epoch (i.e. abs(validation_loss[n] - validation_loss[n-1] where n is the current epoch) is inferior to this epsilon for early_stopping_patience consecutive epochs, then training stops.
• early_stopping_patience: Strictly positive integer. Number of epochs after which the training is stopped if the validation loss improvement is smaller than early_stopping_epsilon.

scheduler

• initial_lr: Float. Initial learning rate.

• scheduler_lr:

  • name: Choice between: "CosineAnnealingLR", "CosineAnnealingWarmRestarts" and "CyclicLR". Please find documentation here.
  • Other parameters that are needed for the scheduler of interest (e.g. "base_lr": 1e-5, "max_lr": 1e-2 for "CosineAnnealingLR").

balance_samples

Bool. Balance positive and negative labels in both the training and the validation datasets.

mixup_alpha

Float. Alpha parameter of the Beta distribution, see original paper on the Mixup technique.

transfer_learning

• retrain_model: Filename of the pretrained model (path/to/pretrained-model). If null, no transfer learning is performed and the network is trained from scratch.
• retrain_fraction: Float between 0. and 1. Controls the fraction of the pre-trained model that will be fine-tuned. For instance, if set to 0.5, the second half of the model will be fine-tuned while the first layers will be frozen.
• reset: boolean. if true, the weights of the layers that are not frozen are reset. If false, they are kept as loaded.

Architecture

Architectures for both segmentation and classification are available and described in the architectures section. If the selected architecture is listed in the loader file, a classification (not segmentation) task is run. In the case of a classification task, the ground truth will correspond to a single label value extracted from target, instead being an array (the latter being used for the segmentation task).

default_model (Mandatory)

Dictionary. Define the default model (Unet) and mandatory parameters that are common to all available architectures. For custom architectures (see below), the default parameters are merged with the parameters that are specific to the tailored architecture.

• name: Unet (default)
• dropout_rate: Float (e.g. 0.4).
• batch_norm_momentum: Float (e.g. 0.1).
• depth: Strictly positive integer. Number of down-sampling operations. - relu (optional): Bool. Sets final activation to normalized ReLU (relu between 0 and 1).

FiLMedUnet (Optional)

• applied: Bool. Set to true to use this model.
• metadata: String. Choice between "mri_params", "contrasts" (i.e. image-based metadata) or the name of a column from the participants.tsv file (i.e. subject-based metadata). "mri_params": Vectors of [FlipAngle, EchoTime, RepetitionTime, Manufacturer] (defined in the json of each image) are input to the FiLM generator. "contrast": Image contrasts (according to config/contrast_dct.json) are input to the FiLM generator.

HeMISUnet (Optional)

• applied: Bool. Set to true to use this model.
• missing_probability: Float between 0 and 1. Initial probability of missing image contrasts as model's input (e.g. 0.25 results in a quarter of the image contrasts, i.e. channels, that will not been sent to the model for training).
• missing_probability_growth: Float. Controls missing probability growth at each epoch: at each epoch, the missing_probability is modified with the exponent missing_probability_growth.

UNet3D (Optional)

• length_3D: (Int, Int, Int). Size of the 3D patches used as model's input tensors.
• stride_3D: [Int, Int, Int]. Voxels' shift over the input matrix to create patches. Ex: Stride of [1, 2, 3] will cause a patch translation of 1 voxel in the 1st dimension, 2 voxels in the 2nd dimension and 3 voxels in the 3rd dimension at every iteration until the whole input matrix is covered.
• attention_unet (optional): Bool. Use attention gates in the Unet's decoder.
• n_filters (optional): Int. Number of filters in the first convolution of the UNet. This number of filters will be doubled at each convolution.

Cascaded Architecture Features

object_detection_params (Optional)

• object_detection_path: String. Path to object detection model and the configuration file. The folder, configuration file, and model need to have the same name (e.g. findcord_tumor/, findcord_tumor/findcord_tumor.json, and findcord_tumor/findcord_tumor.onnx, respectively). The model's prediction will be used to generate bounding boxes.
• safety_factor: List. List of length 3 containing the factors to multiply each dimension of the bounding box. Ex: If the original bounding box has a size of 10x20x30 with a safety factor of [1.5, 1.5, 1.5], the final dimensions of the bounding box will be 15x30x45 with an unchanged center.

Transformations

Transformations applied during data augmentation. Transformations are sorted in the order they are applied to the image samples. For each transformation, the following parameters are customizable: -applied_to: list betweem "im", "gt", "roi". If not specified, then the transformation is applied to all loaded samples. Otherwise, only applied to the specified types: eg ["gt"] implies that this transformation is only applied to the ground-truth data. -dataset_type: list between "training", "validation", "testing". If not specified, then the transformation is applied to the three sub-datasets. Otherwise, only applied to the specified subdatasets: eg ["testing"] implies that this transformation is only applied to the testing sub-dataset.

Available transformations:

• NumpyToTensor
• CenterCrop2D (parameters: size)
• ROICrop2D (parameters: size)
• NormalizeInstance
• RandomAffine (parameters: degrees (Positive integer), translate (List of floats between 0. and 1.), scale (List of floats between 0. and 1.))
• RandomShiftIntensity (parameters: shift_range)
• ElasticTransform (parameters: alpha_range, sigma_range, p)
• Resample (parameters: wspace, hspace, dspace)
• AdditionGaussianNoise (parameters: mean, std)
• DilateGT (parameters: dilation_factor) Float. Controls the number of iterations of ground-truth dilation depending on the size of each individual lesion, data augmentation of the training set. Use 0 to disable.
• HistogramClipping (parameters: min_percentile, max_percentile)
• Clahe (parameters: clip_limit, kernel_size)
• RandomReverse

Uncertainty

Uncertainty computation is performed if n_it>0 and at least epistemic or aleatoric is true. Note: both epistemic and aleatoric can be true.

epistemic

Bool. Model-based uncertainty with Monte Carlo Dropout.

aleatoric

Bool. Image-based uncertainty with test-time augmentation.

n_it

Integer. Number of Monte Carlo iterations. Set to 0 for no uncertainty computation.

Postprocessing

binarize_prediction

Dict. Binarizes predictions according to the given threshold thr. Predictions below the threshold become 0, and predictions above or equal to threshold become 1.

• thr: Float. Threshold is between 0 and 1. To use soft predictions (i.e. no binarisation, float between 0 and 1) for metric computation, indicate -1.

fill_holes

Dict. Fill holes in the predictions. No parameters required (i.e., {}).

keep_largest

Dict. Keeps only the largest connected object in prediction. Only nearest neighbors are connected to the center, diagonally-connected elements are not considered neighbors. No parameters required (i.e., {})

remove_noise

Dict. Sets to zero prediction values strictly below the given threshold thr.

• thr: Float. Threshold is between 0 and 1. Threshold set to -1 will not apply this postprocessing step.

remove_small

Dict. Remove small objects from the prediction. An object is defined as a group of connected voxels. Only nearest neighbors are connected to the center, diagonally-connected elements are not considered neighbors.

• unit: String. Either "vox" for voxels or "mm3". Indicates the unit used to define the minimal object size.
• thr: Int. Minimal object size.

threshold_uncertainty

Dict. Removes the most uncertain predictions (set to 0) according to a threshold thr using the uncertainty file with the suffix suffix. To apply this method, uncertainty needs to be evaluated on the predictions with the uncertainty <Uncertainty> parameter.

• thr: Float. Threshold is between 0 and 1. Threshold set to -1 will not apply this postprocessing step.
• suffix: String. Indicates the suffix of an uncertainty file. Choices: _unc-vox.nii.gz for voxel-wise uncertainty, _unc-avgUnc.nii.gz for structure-wise uncertainty derived from mean value of _unc-vox.nii.gz within a given connected object, _unc-cv.nii.gz for structure-wise uncertainty derived from coefficient of variation, _unc-iou.nii.gz for structure-wise measure of uncertainty derived from the Intersection-over-Union of the predictions, or _soft.nii.gz to threshold on the average of Monte Carlo iterations.

Evaluation parameters

Dict. Parameters to get object detection metrics (true positive and false detection rates), and this, for defined object sizes.

targetSize

• unit: String. Either "vox" for voxels or "mm3". Indicates the unit used to define the target object sizes.
• thr: List. Containing int values. These values will create several consecutive target size bins. For instance with a list of two values, we will have three target size bins: minimal size to first list element, first list element to second list element, and second list element to infinity.

overlap

• unit: String. Either "vox" for voxels or "mm3". Indicates the unit used to define the overlap.
• thr: Int. Minimal object size overlapping to be considered a TP, FP, or FN.

Examples

Examples of configuration files: config_config.json.

In particular:

• config_classification.json. Is dedicated to classification task.
• config_sctTesting.json. Is a user case of 2D segmentation using a U-Net model.
• config_spineGeHemis.json. Shows how to use the HeMIS-UNet.

- config_tumorSeg.json. Runs a 3D segmentation using a 3D UNet. =======