H. Jeremy Bockholt and Mark Scully-1-National Alliance for Medical Image Computing

DetectingWhiteMatterLesionsinLupus

H. Jeremy BockholtMark Scully

Slicer3 Training Compendium

Version 2.36/25/2009

H. Jeremy Bockholt and Mark Scully-2-National Alliance for Medical Image Computing

LearningobjectiveThis tutorial demonstratesan automated, multi-levelmethod to segment whitematter brain lesions inlupus.

Following this tutorial,you’ll be able to load scansinto Slicer3, and segmentand measure the volume ofwhite matter lesions on theprovided data-set.

H. Jeremy Bockholt and Mark Scully-3-National Alliance for Medical Image Computing

Prerequisites

This tutorial assumes that you have already completedthe tutorial Data Loading and Visualization. Tutorialsfor 3DSlicer are available at the following location:

http://www.na-mic.org/Wiki/index.php/Slicer3.2:Training

H. Jeremy Bockholt and Mark Scully-4-National Alliance for Medical Image Computing

Material

This course requires the following installation:

•The current stable version of 3DSlicer 3.4 Software which can be installedfrom:

– http://slicer.org/pages/Special:SlicerDownloads•The White Matter Lesion module extension to 3DSlicer

– (see follow on instructions on slides 7-9 of this tutorial)•The Lupus Lesion Tutorial Data, which can be downloaded from:

– http://wiki.na-mic.org/Wiki/images/c/c8/LesionSegmentationTutorialData.tgz

•n.b., a reliable internet connection will be required for downloading the data

Disclaimer

It is the responsibility of the user of 3DSlicer to comply with both the terms ofthe license and with the applicable laws, regulations and rules.

H. Jeremy Bockholt and Mark Scully-5-National Alliance for Medical Image Computing

Methods

Flowchart summarizing the training pipeline for the white matterlesion classification procedure employed in this tutorial.

H. Jeremy Bockholt and Mark Scully-6-National Alliance for Medical Image Computing

MethodsThe method makes use of local morphometric features based on multiple MR

sequences, including T1-weighted, T2-weighted, and Fluid AttenuatedRecovery from ten subjects.

After preprocessing, including co-registration, brain extraction, biascorrection, and intensity standardization, 49 features were calculated foreach brain voxel based on local morphometry.

At each level of segmentation a supervised classifier takes advantage of adifferent subset of the features to conservatively segment lesion voxels,passing on more difficult voxels to the next classifier.

This multi-level approach allows for a fast lesion classification method withtunable trade-off between sensitivity and specificity, with accuracycomparable to a human rater.

When applying the above classifier to novel data sets, such as the sampledata in this tutorial, the user should follow 3 steps (described on slides11-20) to apply the classifier to the data set they wish to predict lesionsupon.

H. Jeremy Bockholt and Mark Scully-7-National Alliance for Medical Image Computing

FindingtheModule• To add the external

module, Select theExtensionsManagementWizard from theView menu within3DSlicer.

• Click next to searchthe external site forthe appropriatemodule to install.

H. Jeremy Bockholt and Mark Scully-8-National Alliance for Medical Image Computing

InstallingtheModuleSelect

LesionSegmentationApplicationsfrom the list

Click Download &Install.

Click Next after thestatus turns green

H. Jeremy Bockholt and Mark Scully-9-National Alliance for Medical Image Computing

FinishInstallingModuleClick Restart 3DSlicer now once youhave installed themodule

H. Jeremy Bockholt and Mark Scully-10-National Alliance for Medical Image Computing

Joint Intensity Standardization Volume.nhdrJoint Intensity Standardization Volume.raw.gzJoint Intensity Standardization Volume1.nhdrJoint Intensity Standardization Volume1.raw.gzJoint Intensity Standardization Volume2.nhdrJoint Intensity Standardization Volume2.raw.gzLesionSegmentTutorial.mrmlPredict Lesions Volume.nhdrPredict Lesions Volume.rawPredict Lesions Volume1.nhdrPredict Lesions Volume1.rawlesionSegmentation.modellupus002_FLAIR_reg+bias.nii.gzlupus002_T1_reg+bias.nii.gzlupus002_T2_reg+bias.nii.gzlupus002_brain_mask.nii.gzlupus003_FLAIR_reg+bias.nii.gzlupus003_T1_reg+bias.nii.gzlupus003_T2_reg+bias.nii.gzlupus003_brain_mask.nii.gzsvm.model

TutorialDataThis course is built upon two scans of patientswith lupus that have T1, T2, and FLAIR images.These images have been co-registered and brainextracted outside of the scope of this tutorial.Also, the tutorial data contains model files thatsupport the module introduced in this tutorial.Finally, results files from one subject are includedin the tutorial data. The following summary showsthe contents of the LesionSegmentationTutorialarchive once downloaded and uncompressed toyour filesystem.

H. Jeremy Bockholt and Mark Scully-11-National Alliance for Medical Image Computing

ModuleStep1:Setup

Load the scene by selecting File ImportScene feature from the menu. Navigate thefilesystem to locate the MRML scene thatyou have downloaded. By loading the sceneyou will load the reference data sets that areneeded for this tutorial.

H. Jeremy Bockholt and Mark Scully-12-National Alliance for Medical Image Computing

ModuleStep1:Results

After the scene loads, you willhave the data sets that are neededfor this tutorial and may continueon to step 2. Confirm that you havethe data sets listed in the scenedisplay on the left.

H. Jeremy Bockholt and Mark Scully-13-National Alliance for Medical Image Computing

ModuleStep2:Setup

The next step is to select the IntensityStandardization module: Select and expandthe Module menu, First select Segmentation,next select Lesion Segmentation, finally selectJoin Intensity Standardization.

H. Jeremy Bockholt and Mark Scully-14-National Alliance for Medical Image Computing

ModuleStep2:Setup

In order to run the joint intensity standardizestep, confirm the input parameters to themodule that are listed on the left, you willalso need to scroll down and confirm theoutput stats parameters shown on the nextslide.

H. Jeremy Bockholt and Mark Scully-15-National Alliance for Medical Image Computing

ModuleStep2:Running

Finishing configuring the intensitystandardize procedure to run by confirmingthe output stats parameters as shown on theleft.

Next, Click Apply to run this step, fortypical computers, expect 2-5 minutes ofprocessing time.

H. Jeremy Bockholt and Mark Scully-16-National Alliance for Medical Image Computing

ImportanceofIntensityStandardization

• At this stage, intensity standardization is nowcomplete. This preprocessing is needed becausethe scale and range of intensities within each ofthe T1, T2, and FLAIR sequences vary acrossindividuals, the intensity standardization attemptsto normalize the intensities within eachsequence. The normalized intensity values arethe input for the next phase, step 3 whichclassifies the input images from step 2 to predictwhere white matter lesions are present in thegiven data set.

H. Jeremy Bockholt and Mark Scully-17-National Alliance for Medical Image Computing

ModuleStep3:Setup

The next step is to select the Predict Lesionsmodule: Select and expand the Modulemenu, First select Segmentation, next selectLesion Segmentation, finally select PredictLesions

H. Jeremy Bockholt and Mark Scully-18-National Alliance for Medical Image Computing

ModuleStep3:Running

Finishing configuring the Predict Lesionsprocedure to run by confirming the input andoutput parameters as shown on the left.

Next, Click Apply to run this step, fortypical computers, expect 90-120 minutes ofprocessing time.

H. Jeremy Bockholt and Mark Scully-19-National Alliance for Medical Image Computing

ModuleStep3:Results

Once complete the results will load as alesion mask volume, as well as, a heat map,or lesion probability volume

H. Jeremy Bockholt and Mark Scully-20-National Alliance for Medical Image Computing

PredictedLesionsFollowing Stage 3, you now have both volumes and

probability images that represent where whitematter lesions are predicted for the given dataset. The remainder of this tutorial will show youfirst how you may use the volume data tomeasure the lesion load, and also how the datacan be rendered to visualize the location andextent of lesions for the given data set.

H. Jeremy Bockholt and Mark Scully-21-National Alliance for Medical Image Computing

ExampleMeasurement

As a use case of the results, you may use theLabel Statistics Module to summarize thelesion load. As shown select Statistics andthen select LabelStatistics.

H. Jeremy Bockholt and Mark Scully-22-National Alliance for Medical Image Computing

ExampleMeasurement

After selecting Flair as input GrayscaleVolume and Predict Lesions Volume as theInput Labelmap, a summary of the volumeis provided in the Label Statistics. In thisexample, the total lesion volume is 384mm^3 for lupus003 subject.

H. Jeremy Bockholt and Mark Scully-23-National Alliance for Medical Image Computing

SetupExampleModel

You may also create a model of theclassified lesions by using the Model Makermodule. To do this, select Surface Modelsand then select the Grayscale ModelMaker.

H. Jeremy Bockholt and Mark Scully-24-National Alliance for Medical Image Computing

ResultsExampleModel

To the left shows how the Predict LesionsVolume can be rendered using the modelmaker and above shows the rendered result.

H. Jeremy Bockholt and Mark Scully-25-National Alliance for Medical Image Computing

Discussion• This concludes the objective of the tutorial; however, since

the tool has produced a label map, as shown in theexample, you may now measure the volumes of theautomatically labeled lesion tissue or summarize theanatomical location of lesions for other cases in the tutorialdata set. Since, the lesion load is associated with symptomseverity and can be used to guide treatment and care.

• You may use the lesion label maps as input to the changetracker capability in Slicer to assess time course of theillness (change in lesion size, number over time).

• You may use the label maps to assess either perfusion ordiffusion deficits through co-registration of the lesion mapswith pMR, ASL, or DTI.

H. Jeremy Bockholt and Mark Scully-26-National Alliance for Medical Image Computing

Conclusion• This capability provides an intuitive graphical

user interface to interact with the data• The tool has been built in an open-source

environment and is readily available to thescientific community

H. Jeremy Bockholt and Mark Scully-27-National Alliance for Medical Image Computing

ForMoreInformation• Register as a user of this 3dSlicer Module using

the NITRC resource to keep updated on anychanges or additions to either the capability ortutorial– http://www.nitrc.org/projects/lupuslesion/

• You may also send e-mail message with anyquestions or concerns to Jeremy Bockholt(jbockholt@mrn.org)

H. Jeremy Bockholt and Mark Scully-28-National Alliance for Medical Image Computing

Acknowledgments

National Alliance for Medical Image ComputingNIH U54EB005149

And other support:

DOE DE-FG02-99ER6274NIH 5R01HL077422-02NIH P41 RR13218NIH U24-RR021992