Development and Validation of Qualitative and Quantitative

Descriptors in Gliomas

David A Gutman MD PhD

Department of Biomedical Informatics

Emory University

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Quick overview of Glioblastoma (GBM)

• Most common primary brain

tumor in adults

• Median survival 50 weeks

• ISBTRC Goals:

– To leverage rich datasets to understand the mechanisms of glioma progression through In Silico analysis

– To manage, explore and share semantically complex data among researchers

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

The Cancer Genome Atlas (TCGA)

• Characterize 500 tumors for each of a variety of cancers

• Clinical records

• Genomics: gene, miRNA expression, copy number, sequence, DNA methylation

• Imaging: pathology and radiology

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

TCGA and Imaging Data: Radiology and Pathology

• The Cancer Imaging Archive (TCIA) now contains radiology data on ~ 150 patients from the TCGA GBM data set

• Pathology data is also available on ~ 200 patients

• Our extended group’s goal is to “mine” radiology and pathology data for phenotypes that correlate with genetic and clinical characteristics of the patients

• Dr. Cooper presented some of our work correlating pathology with genomics and outcomes

• Parallel effort has been underway for radiology data sets

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Overall question…

• Do tumors that “look” different behave differently?

– e.g. different outcome

– Different genetic profiles

– Need for a standardized method to describe what the tumors look like…

Problems…

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Genetic signatures can define tumor subtypes

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Clustering identifies three morphological groups

• Analyzed 200 million nuclei from 162 TCGA GBMs (462 slides)

• Named for functions of associated genes:

Cell Cycle (CC), Chromatin Modification (CM),

Protein Biosynthesis (PB)

• Prognostically-significant (logrank p=4.5e-4)

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Representative nuclei

Large,

hyperchromatic

nuclei

Small light nuclei,

Eosinophilic cyoplasm

Intermediate

L Cooper

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

How Does One Effectively Marry Imaging Findings of a Tumor to its Genomics?

X

Genetic Microarray

A Flanders

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

VASARI Feature Set

• A set of 30 imaging characteristics to describe high grade gliomas (GBM) using standardized vocabulary that is reproducible and understandable by neuroradiologists

• Effort led by Adam Flanders and Carl Jaffe involving coordinating “reads” and feature set development by ~ 8 neuroradiologists

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Defining a Rich Set of Qualitative and Quantitative Image Biomarkers

• This has been a community-driven ontology development project to create a comprehensive set of imaging observations for GBM – Collaboration with ASNR

• Collaborators were asked to provide a list of clinical or literature observations that could be used to describe MRI features of GBM

• Imaging features (26 features / 4 categories)

– Location of lesion

– Morphology of lesion margin (definition, thickness, enhancement, diffusion)

– Morphology of lesion substance (enhancement, PS characteristics, focality/multicentricity, necrosis, cysts, midline invasion, cortical involvement, T1/FLAIR ratio)

– Alterations in vicinity of lesion (edema, edema crossing midline, hemorrhage, pial invasion, ependymal invasion, satellites, deep WM invasion, calvarial remodeling)

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

F5 – Proportion Enhancing

Visually, when scanning through the entire tumor volume, what proportion of the

entire tumor would you estimate is enhancing? (Assuming that the entire

abnormality may be comprised of: (1) an enhancing component, (2) a non-enhancing

component, (3) a necrotic component and (4) a edema component.)

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

F7 – Proportion Necrosis

Visually, when scanning through the entire tumor volume, what proportion of the tumor is estimated to represent necrosis?

Necrosis is defined as a region within the tumor that does not enhance or shows markedly diminished enhancement, is

high on T2W and proton density images, is low on T1W images, and has an irregular border). (Assuming that the entire

abnormality may be comprised of: (1) an enhancing component, (2) a non-enhancing component, (3) a necrotic component

and (4) a edema component.)

(6) 68-95%

(7) >95%

(8) All 100%

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Capturing structured annotations and markups/AIM Data Service

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

For validation, focused on semi-quantitative features

• Compared various outcome and genomic measures with these features

• Also did comparisons between qualitative and quantitative volumetric measurements performed at MGH by Colen et. al using 3D slicer, and measurements done at Emory using the Velocity Platform

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Correlating between quantitative and qualitative features: Man vs Machine

Results of univariate linear regression for agreement between VASARI

measurements and measurements derived from quantitative

volumetric analyses.

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Agreement between qualitative and quantitative feature set

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Inter-rater agreement of relevant imaging features between radiologists scores according

to VASARI standard

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

3d Slicer Volume Segmentation (R. Colen/MGH)

Visualization of quantitative volumetric segmentation methodology. Region

corresponding to edema/tumor infiltration (blue) was segmented from

FLAIR sequences whereas contrast enhancement (yellow) and necrosis

(orange) have been segmented from T1 post contrast weighted images

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Machine vs Machine?

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Cleaning up the raw data from TCIA

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Developed some tooling to help with image validation & QA

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Slicer Volumes vs Velocity Derived Volumes

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Do image features predict outcome?

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Combination of clinical and imaging features

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Are imaging features equally distributed across Verhaak classification subtypes?

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Correlation of Volumetric Data with Outcome

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

Future Work

• Working on extracting features from volumetric images and doing pathway analysis

• Also Rajan Jain (TJU) and Scott Hwang (Emory) have begun doing feature extraction/markups of perfusion and DTI data

• Continue to collect imaging data from TCGA GBM contributors (as we track them down)

• Continue to revise/simplify feature set

• Consider extending feature set to lower grade cases

C

en

ter f

or C

om

preh

en

siv

e I

nfo

rm

ati

cs

In Silico Brain Tumor Research Center Team

• Emory University

– Lee Cooper

– Joel Saltz

– Daniel Brat

– Carlos Moreno

– Chad Holder

– Scott Hwang

– Doris Gao

– William Dunn

– Tarun Aurora

• NCI

– Eric Huang

– Carl Jaffe

• MGH

– Rivka Colen

• Henry Ford Hospital

– Tom Mikkelsen

– Lisa Scarpace

• Thomas Jefferson University

– Adam Flanders

• SAIC Frederick

– John Freymann

– Justin Kirby