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.
Run the specified command. Choices: "train"
or "test"
, to train and evaluate a model respectively.
Integer. ID of the GPU to use.
Folder name that will contain the output files (e.g., trained model, predictions, results).
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, ...}
Bool. Extended verbosity and intermediate outputs.
String. Path of the BIDS folder.
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.
train_validation
: List. List of image contrasts (e.g.T1w
,T2w
) loaded for the training and validation. Ifmultichannel
istrue
, this list represents the different channels of the input tensors (i.e. its length equals model'sin_channel
). Otherwise, the contrasts are mixed and the model has only one input channel (i.e. model'sin_channel=1
).test
: List. List of image contrasts (e.g.T1w
,T2w
) loaded in the testing dataset. Same comment than fortrain_validation
regardingmultichannel
.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 availableT1w
images will be included into the training/validation/test set. Please setmultichannel
tofalse
if you are using this parameter.
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.
Choice between "sagittal"
, "coronal"
, and "axial"
. Sets the slice orientation for on which the model will be used.
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.
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"
) withROICrop
as transform. Please usenull
if you do not want to use an ROI to crop.slice_filter_roi
: int. If the ROI mask contains less thanslice_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"
.
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.
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.
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.
{"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.
String (Optional). Metadata contained in "participants.tsv" file with categorical values. Each category will be evenly distributed in the training, validation and testing datasets.
Float. Between 0
and 1
representing the fraction of the dataset used as training set.
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"
.
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.
Strictly positive integer.
name
: Name of the loss function class. Seeivadomed.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
forFocalLoss
).
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 forearly_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 thanearly_stopping_epsilon
.
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"
).
Bool. Balance positive and negative labels in both the training and the validation datasets.
Float. Alpha parameter of the Beta distribution, see original paper on the Mixup technique.
retrain_model
: Filename of the pretrained model (path/to/pretrained-model
). Ifnull
, 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.
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).
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).
applied
: Bool. Set totrue
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 toconfig/contrast_dct.json
) are input to the FiLM generator.
applied
: Bool. Set totrue
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, themissing_probability
is modified with the exponentmissing_probability_growth
.
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.
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
, andfindcord_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 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.
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. Use0
to disable.HistogramClipping
(parameters:min_percentile
,max_percentile
)Clahe
(parameters:clip_limit
,kernel_size
)RandomReverse
Uncertainty computation is performed if n_it>0
and at least epistemic
or aleatoric
is true
. Note: both epistemic
and aleatoric
can be true
.
Bool. Model-based uncertainty with Monte Carlo Dropout.
Bool. Image-based uncertainty with test-time augmentation.
Integer. Number of Monte Carlo iterations. Set to 0 for no uncertainty computation.
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.
Dict. Fill holes in the predictions. No parameters required (i.e., {}).
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., {})
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.
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.
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.
Dict. Parameters to get object detection metrics (true positive and false detection rates), and this, for defined object sizes.
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.
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 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. =======