User Manual¶

In this chapter we will discuss the parts of WORC in more detail, mostly focussed on the inputs, outputs, and the various workflow possible. We will give a more complete overview of the system and describe the more advanced features.

Interacting with WORC¶

The WORC toolbox is build around of one main object, the WORC object. This object provides all functionality of the toolbox. However, to make certain functionalities easier to use and limit the complexity, we have constructed two facades. The SimpleWORC facade is the simplest to interact with and provides all functionality required for conducting basic experiments. The BasicWORC object is based on the SimpleWORC object and provides several more advances functions. The specific functionalities of these two facades and the WORC object itself can be found in this section.

For documentation on SimpleWORC and BasicWORC, please look at the documentation within those modules: WORC.facade.simpleworc and WORC.facade.basicworc. Many of the functions are actually wrappers to interact with the WORC object, and therefore use the functionality described below. For basic usage, only using SimpleWORC and it’s respective documentation should be sufficient.

The WORC Object¶

import WORC
network = WORC.WORC('somename')

It’s attributes are split in a couple of categories. We will not discuss the WORC.defaultconfig() function here, which generates the default configuration, as it is listed in a separate page, see the Config chapter. More detailed documentation of the various functions can be found in the docstrings of WORC.WORC: we will mostly focus on the attributes, inputs, outputs and workflows here.

Input file definitions¶

Attributes: Sources¶

There are numerous WORC attributes which serve as source nodes for the FASTR network. These are:

images_train and images_test
segmentations_train and segmentations_test
semantics_train and semantics_test
labels_train and labels_test
masks_train and masks_test
features_train and features_test
metadata_train and metadata_test
Elastix_Para
fastrconfigs

When using a single dataset for both training and evaluation, you should supply all sources in train objects. You can use several kinds of validation methods (e.g.cross validation) to compute the performance on this dataset. Optionally, you can supply a separate training and test set.

Each source should be given as a dictionary of strings corresponding to the source files. Each element should correspond to a single object, e.g. tumor, or patient. The keys are used to match the features to the label and semantics sources, so make sure these correspond to the label file. The values should refer to the actual source files corresponding to the FASTR formats, see http://fastr.readthedocs.io/en/stable/fastr.reference.html#ioplugin-reference.

You can off course have multiple images or ROIs per object, e.g. a liver ROI and a tumor ROI. This can be easily done by appending to the sources. For example:

images1 = {'patient1': vfs://example/MR.nii, 'patient2': vfs://example/MR.nii}
segmentations1 = {'patient1': vfs://example/tumor.nii, 'patient2': vfs://example/tumor.nii}
segmentations2 = {'patient1': vfs://example/liver.nii, 'patient2': vfs://example/liver.nii}

network.images_train.append(images1)
network.images_train.append(images1)

network.segmentations_train.append(segmentations1)
network.segmentations_train.append(segmentations2)

When using multiple sequences per patients (e.g. T1 and T2), the same appending procedure can be used.

Note

You have to make sure the images and segmentation sources match in size.

Note

You have to supply a configuration file for each image or feature source you append. Thus, in above example, you need to append two configurations!

Note

When you use multiple image sequences, you can supply a ROI for each sequence by appending to to segmentations object. Alternatively, when you do not supply a segmentation for a specific sequence, WORC will use Elastix to align this sequence to another through image registration. It will then warp the segmentation from this sequence to the sequence for which you did not supply a segmentation. WORC will always align these sequences with no segmentations to the first sequence, i.e. the first object in the images_train list. Hence make sure you supply the sequence for which you have a ROI as the first object.

Images and segmentations¶

The minimal input for a Radiomics pipeline consists of either images (plus a segmentation if you have not implemented an automatic segmentation tool) or features plus a label file (and a configuration, but you can just use the default one.

If you supply these, features will be computed within the segmentations on the images. They are read out using SimpleITK, which supports various image formats such as DICOM, NIFTI, TIFF, NRRD and MHD.

Semantics¶

Semantic features are used in the PREDICT CalcFeatures tool. You can supply these as a .csv listing your features per patient. The first column should always be named Patient and contain the Patient ID. The other columns should contain a label for the feature and their values. For example:

Patient	Label1	Label2
patient1	1	0
patient2	2	1
patient3	1	5

Similar to the patient labels, the semantic features are matched to the correct image/features by the name of the image/features. So in this case, your sources should look as following:

images_train = {'patient1': 'source1.nii.gz', 'patient2': 'source2.nii.gz', ...}
segmentations_train = {'patient1': 'seg1.nii.gz', 'patient2': 'seg2.nii.gz', ...}

Labels¶

The labels are used in classification. For PREDICT, these should be supplied as a .txt file. Similar to the semantics, the first column should head Patient and contain the patient ID. The next columns can contain things you want to predict. Hence the format is similar to the semantics file.

Masks¶

WORC contains a segmentation preprocessing tool, called segmentix. This tool is still under development. The idea is that you can manipulate your segmentation, e.g. using dilation, then use a mask to make sure it is still valid. Currently, you can only let it take a ring of a certain radius around your ROI and mask it.

Features¶

If you already computed your features, e.g. from a previous run, you can directly supply the features instead of the images and segmentations and skip the feature computation step. These should be stored in .hdf5 files matching the PREDICT CalcFeatures format.

Metadata¶

This source can be used if you want to use tags from the DICOM header as features, e.g. patient age and sex. In this case, this source should contain a single DICOM per patient from which the tags that are read. Check the PREDICT.imagefeatures.patient_feature module for the currently implemented tags.

Elastix_Para¶

If you have multiple images for each patient, e.g. T1 and T2, but only a single segmentation, you can use image registration to align and transform the segmentation to the other modality. This is done in WORC using Elastix http://elastix.isi.uu.nl/. In this source, you can supply a parameter file for Elastix to be used in the registration in .txt. format. Alternatively, you can use SimpleElastix to generate a parameter map and pass this object to WORC. Note: WORC assume your segmentation is made on the first WORC.images source you supply. The segmentation will be alingned to all other image sources.

Attributes: Settings¶

There are several attributes in WORC which define how your pipeline is executed:

fastr_plugin
fastr_tmpdir
Tools: additional workflows are stored here. Currently only includes a pipeline for image registration without any Radiomics.
CopyMetadata: Whether to automatically copy the metadata info (e.g. direction of cosines) from the images to the segmentations before applying transformix.

An explanation of the FASTR settings is given below.

Attributes: Functions¶

The WORC.configs() attribute contains the configparser files, which you can easily edit. The WORC.set() function saves these objects in a temporary folder and converts the filename into as FASTR source, which is then put in the WORC.fastrconfigs() objects. Hence you do not need to edit the fastrconfigs object manually.

FASTR settings¶

There are two WORC attributes which contain settings on running FASTR. In WORC.fastr_plugin, you can specify which Execution Plugin should be used: see also http://fastr.readthedocs.io/en/stable/fastr.reference.html#executionplugin-reference.

The default is the ProcessPollExecution plugin. The WORC.fastr_tempdir sets the temporary directory used in your run.

Construction and execution commands¶

After supplying your sources, you need to build the FASTR network. This can be done through the WORC.build() command. Depending on your sources, several nodes will be added and linked. This creates the WORC.network() object, which is a fastr.network() object. You can edit this network freely, e.g. add another source or node. You can print the network with the WORC.network.draw_network() command.

Next, we have to tell the network which sources should be used in the source nodes. This can be done through the WORC.set() command. This will put your supplied sources into the source nodes and also creates the needed sink nodes. You can check these by looking at the created WORC.source_data_data and WORC.sink objects.

Finally, after completing above steps, you can execute the network through the WORC.execute() command.

Evaluation of your network¶

In WORC, there are two options for testing your fitted models:

Single dataset: cross-validation (currently only random-split)
Separate train and test dataset: bootstrapping on test dataset

Within these evaluation settings, the following performance evaluation methods are used:

Confidence intervals on several metrics:
For classification:
1. Area under the curve (AUC) of the receiver operating characteristic (ROC) curve. In a multiclass setting, weuse the multiclass AUC from the TADPOLE Challenge.
2. Accuracy.
3. Balanced classification accuracy as defined by the TADPOLE Challenge.
4. F1-score
5. Sensitivity, aka recall or true positive rate
6. Specificity, aka true negative rate
7. Negative predictive value (NPV)
8. Precision, aka Positive predictive value (PPV)
For regression:
1. R2-score
2. Mean Squared Error (MSE)
3. Intraclass Correlation Coefficient (ICC)
4. Pearson correlation coefficient and p-value
5. Spearmand correlation coefficient and p-value
For survival, in addition to the regression scores: a) Concordance index b) Cox regression coefficient and p-value

In cross-validation, by default, 95% confidence intervals for the mean performance measures are constructed using the corrected resampled t-test base on all cross-validation iterations, thereby taking into account that the samples in the cross-validation splits are not statistically independent. See als Nadeau C, Bengio Y. Inference for the generalization error. In Advances in Neural Information Processing Systems, 2000; 307–313.

In bootstrapping, 95% confidence intervals are created using the ‘’standard’’ method according to a normal distribution: see Table 6, method 1 in Efron B., Tibshirani R. Bootstrap Methods for Standard Errors, Confidence Intervals, and Other Measures of Statistical Accuracy, Statistical Science Vol.1, No,1, 54-77, 1986.
ROC curve with 95% confidence intervals using the fixed-width bands method, see Macskassy S. A., Provost F., Rosset S. ROC Confidence Bands: An Empirical Evaluation. In: Proceedings of the 22nd international conference on Machine learning. 2005.
Univariate statistical testing of the features using:
1. A student t-test
2. A Welch test
3. A Wilcoxon test
4. A Mann-Whitney U test
The uncorrected p-values for all these tests are reported in a single excel sheet. Pick the right test and significance level based on your assumptions. Normally, we make use of the Mann-Whitney U test, as our features do not have to be normally distributed, it’s nonparametric, and assumes independent samples.
Ranking patients from typical to atypical as determined by the model, based on either:

a) The percentage of times a patient was classified correctly when occuring in the test set. Patients always correctly classified can be seen as typical examples; patients always classified incorrectly as atypical. b) The mean posterior of the patient when occuring in the test set.

These measures can only be used in classification. Besides an Excel with the rankings, snapshots of the middle slice of the image + segmentation are saved with the ground truth label and the percentage/posterior in the filename. In this way, one can scroll through the patients from typical to atypical to distinguish a pattern.
A barchart of how often certain features groups were selected in the optimal methods. Only useful when using groupwise feature selection.

By default, only the first evaluation method, e.g. metric computation, is used. The other methods can simply be added to WORC by using the add_evaluation() function, either directly in WORC or through the facade:

import WORC
network = WORC.WORC('somename')
label_type = 'name_of_label_predicted_for_evaluation'
...
network.add_evaluation(label_type)

import WORC
from WORC import SimpleWORC
experiment = SimpleWORC('somename')
...
experiment.add_evaluation()

Debugging¶

As WORC is based on fastr, debugging is similar to debugging a fastr pipeline: see therefore also the fastr debugging guidelines.

If you run into any issue, please create an issue on the WORC Github.

Example data¶

For many files used in typical WORC experiments, we provide example data. Some of these can be found in the exampledata folder within the WORC package. To save memory, for several types the example data is not included, but a script is provided to create the example data. This script (create_examplt_data) can be foud in the exampledata folder as well.