README.md

In robertogattabs/RadAgent: Package for handling DICOM RT files

RadAgent

Preliminary Operations

First of all we have to load the library :)

library(RadAgent)

Then we have to load the CSV with the clinical outcome. The csv can also contains a set of clinical variable you may wish to add to the model. This is because, in general, in Radiomics investigation we are interested to find out the BEST model, which usually includes clinical and image features. One note: the first line of the csv must contains the header

Loading the clinical outcome

csv.Dino <- read.csv("~/original.csv.Dino.csv")

The first thing to do is to save the column with the clinical, dicotomic, outcome (0/1), in an array and name the positions with the names of the folders of the corresponding DICOM Studies. In my case such names are numbers from 2 to 40 with some holes (I deleted some folders due to a not acceptable pixelSpacing)

arr.clinicalOutcome <- csv.Dino$Outcome
names(arr.clinicalOutcome) <- as.character( c(seq(2,6),seq(9,17),seq(19,24),seq(26,32),seq(34,40)) )

Loading other clinical covariates

This step is optional: you can train a Radiomics model based only on features acquired from the DICOM images. However, as example, now I load other clinical covariates, such as Age, Sex, status of the pelivic lymphnodes, cT and cN. I build a new matrix called 'other.covariates.matrix':

# assign the AGE, present in the csv (numeric)
arr.eta <- csv.Dino$eta
# assign teh SEX, present in csv (0,1)
arr.sesso <- csv.Dino$sesso
# assign the status of pelvic lynfonode, present in the csv (0,1)
arr.STATO.LINFONODI.INGUINALI <- csv.Dino$STATO.LINFONODI.INGUINALI
# assigne the clinical T stage, present in the CSV (categorical 1,2,3,4)
arr.xT <- csv.Dino$xT
# assigne the clinical N stage, present in the CSV (categorical 1,2,3)
arr.xN <- csv.Dino$N
# build the matrix with the clinical variables to include in the model
other.covariates.matrix <- cbind( "eta"=arr.eta, "sesso"=arr.sesso, 
                                  "STATO.LINFONODI.INGUINALI"=arr.STATO.LINFONODI.INGUINALI, 
                                  "xT"=arr.xT, "xN"=arr.xN )                                 

RADIOMICS agent

RadAgent extends the moddicom package [1], providing the entire pipeline for Radiomics investigation. Because of that, it requires a set of DICOM studies formatted in a way compatible with moddicom. This means, for example (a) axial or almost-axial projections only, (b) voxels disjointed and adiacents, (c) no missing image slices and (d) a DICOM RTStruct object contaning the contoured ROI (Region Of Interest).

Starting from those, RadAgent: extract the image features on the found foldes, given and entry point on the filesystem. The features are compliant with the Image biomarker standardisation initiative [2]. given and entry point on the filesystem, it loads the images from the found foldes. applies the LoGs filtering, according to the given sigmas; for each filtered image, it extracts the features, compliant with the Image biomarker standardisation initiative [2]. for each feature, changing the sigmas in the LoG filtering, it calculate makes a Mann Withney test with the clinical outcome. Then it takes the sigmas able to give the highest performances to that feature. In case of high variation of the p.value for small variations of the sigma, it can discharge the feature (doubt of overfitting) it check the couple of features and in case of high correlation (pearson test) it discharge the feature with the lowest performance in predicting the results; it begins to build a model with a feedforeward feature selection strategy, until the number of requested feature in the model is reached. the trained model is a logist regressor, which performances are calculated in terms of AUC of the ROC curve.

A RadAgent object can be easyly instantiated with

obj <- RAD.RadiomicAgent()

The default values for the internal parameters (a rich set of threshold of the previously cited steps) can be set with:

obj$set.param()

The model can now be trained with:

m.features <- obj$scoutFeatures(pathName = "/images/test4RadAgent/",
                                ROIName = "GTV",feature.family = c("stat","morph","glcm","rlm","szm"),
                                arr.sigma=c( 3.2, 3.4, 3.6, 3.8, 4, 4.2, 4.4,  4.6, 4.8, 5, 5.2, 5.4, 5.6, 5.8, 6 ),
                                arr.clinicalOutcome=arr.clinicalOutcome, 
                                matrice.altre.covariate = matrice.altre.covariate,
                                cache.fileName = "dino.RData",
                                forceRecalculus = FALSE,)

The meaning of the parameters is the following:

• pathName : the folder containin the subfolders with the DICOM studies. Each subfolder name must have a corresponding named position in the clinical array outcome;
• ROIName : the ROIName that shuold be extracted;
• feature.family : the features families you want to extract;
• arr.sigma : the array with the sigma for the LoG filtering. Sigmas are expressed in mm, so be sure to adopt sigma values comparable (larger) than your pixelspacing;
• arr.clinicalOutcome : the array containing the clinical outcomes;
• matrice.altre.covariate : the matrix containing the other clinical variables you want to use to enrich the model (optional);
• cache.fileName : the name of the file adopted to cache the computation. This is useful because if you want to interrupt the computation, you can: when the script will be run again, if "force-recalculus" is set to FALSE, it will load from this cache file what was already calculated;
• forceRecalculus : do not exploit the cache: if you find a cache delete it and calculate all from scratch;

Important note: the computation can take a loooooong time (hours or days, depending on how much sigmas you want to investigate). I suggest you to set the forceRecalculus to FALSE.

After the computation, you can retrieve a model for example with two covariate with:

lst.models <- obj$featureSelection(  number.of.cov =  2)

Interpreting the results

The returned lst.models is a list containing the following elements (the most important):

• tmp.str.covariate : the string of the covariates in the "winning model";
• covariate.incluse : the same of tmp.str.covariate but in form of array;
• lista.modelli : a list, described below (this is probably the most interesting result);
• arr.accuracy : an array containing the accuracy of the best models. The accuracy is calculated with a threshold of .5. Please, use this carefully because in case of unbalanced outcomes this can lead you to misunderstand and overestimate the performances of the models;
• arr.sensitivity : this is the array of the sensitivity (see arr.accuracy for details);
• arr.specificity : this is the array of the specificity (see arr.accuracy for details);
• arr.AUC : this is the array with the AUC of the best models;

The most interesting element, lista.modelli, is a list which contains a set of "good" models, providing for each:

• modello : the R Regressor model (with coefficients, Degrees of Freedom, Akaike information criterion, etc..). This can be easily used to predict the outcome for new nets of incoming data.
• roc : the ROC curve (which can be easily plot with the plot command). This also contains the information on the confidence interval;
• auc : the numeric value of the AUC of the roc;
• confusion.matric : the confusion matrix built using a default threshold of .5. Differend confusion matrices can actually be built directly using the model provided in modello and the predict function of R (e.g.: see [3]).

Here is an example of a monovariate Logist Model ROC, built on a set of 36 PET images of patients affected from rectal cancer. The model uses two image features extracted after a LoG filtered with sigma equal to 4. The two features are F_cm.inv.diff.norm and F_cm.contrast, see [2] for details.

plot(lst.models$lista.modelli$`s_4_F_cm.inv.diff.norm+s_4_F_cm.contrast`$roc)

references:

[1] Dinapoli N, Alitto AR, Vallati M, Gatta R, Autorino R, Boldrini L, Damiani A, Valentini V. Moddicom: a complete and easily accessible library for prognostic evaluations relying on image features. Conf Proc IEEE Eng Med Biol Soc. 2015 Aug;2015:771-4. doi: 10.1109/EMBC.2015.7318476.

[2] Zwanenburg A, Leger S, Vallières M, Löck S. Image biomarker standardisation initiative, eprint arXiv:1612.07003. Available (29/06/2018) at: https://arxiv.org/abs/1612.07003

[3] http://stat.ethz.ch/R-manual/R-devel/library/stats/html/predict.lm.html

robertogattabs/RadAgent documentation built on June 30, 2018, 12:02 a.m.