1

Department of Radiation Oncology��� University of Michigan

The

Radiomics and the Coming ���Pan-Omics Revolution���

Issam El Naqa, MA, PhD

07/14/2015 2

The “-Omics” World I• Definition:

– A field of study in biology ending in –omics(genomics, transcriptomics, proteomics ormetabolomics)

• Objective:– Collective characterization and quantification

of pools of biological molecules that translateinto the structure, function, and dynamics ofan organism(s)

Wikipedia.org

07/14/2015 3

The “-Omics” World II

www.iupui.edu 07/14/2015 4

Omics-Based Test Development Process

• According to the Institute of Medicine (IOM):

2

07/14/2015 5

Radiogenomic Modeling of ���Rectal Bleeding in Prostate cancer

Coates et al, RO, 2015

CNV

V10, V20, D50 and XRCC1 CNV 07/14/2015 6

Radiogenomics assessment via PCA visualization

07/14/2015 7

Identification of robust biomarkers for RP using Proteomics with limited samples (n= 3x3)

Oh, Craft et al., JPR, 2011

p>>n Use current

knowledge as regularizer

07/14/2015 8

Radioproteomics in Lung Cancer

Lee et al, Med Phys, 2015

Serum biomarkers in Radiation Pneumonitis

3

07/14/2015 9

Radiation Response as Pan-Omics

Treatment Technique

DNA damage detection and repair genes

Fibrotic and Inflammatory

Cytokines

Anti-oxidant Enzymes

Dose Prescription

Anatomical Imaging

Dose-Volume Metrics

Tumor site

Tumor Stage

Histology

Demographics Biomarkers Clinical Factors

Physical Factors

Planning Data

Functional imaging

Panomics

Integration of physics (radiomics) and biology (genomics) 07/14/2015 10

Lung Cancer Jamboree El Naqa 2014

07/14/2015 11

Lung Cancer Jamboree

Imaging Tumor

Pneumonitis

Fibrosis

Biomarkers

07/14/2015 12

Radiomics• A ‘new’ form of –omics

– Quantitative information from multi-imaging modalities (PET, CT, MRI, etc)could be related to biological andclinical endpoints (Lambin et al, 2012)

– In oncology, it is decoding the TumorPhenotype with Non-Invasive Imaging(radiomics.org)

4

07/14/2015 13

An image worth thousand(s) words

07/14/2015 14

Our early radiomics work

El Naqa, Galatsanos, et al. IEEE TMI, 2002 El Naqa, Galatsanos, et al. IEEE TMI, 2004

07/14/2015 15

More recently: HNC PET

El Naqa, Deasy, et al. PMB, 2009 07/14/2015 16

Common radiomics features

El Naqa, 2014

5

07/14/2015 17

PET/CT from NSCLC local tumor

SUVmax=13.7 07/14/2015 18

IVH

Texture map

07/14/2015 19

1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Patient Group

Pred

ictio

n

OriginalPrediction model

Radiomics PET/CT model

rs=0.5908 (p=0.0013)

( ) 29.2* _ 80 3.11* _ 70 0.826g PET V CT V= − + −x

Vaidya et al., RO ‘11 07/14/2015 20

Textures VS CTCv4 fibrosis score

Co-occurrence

Run-length

Global

Texture change in low dose region agrees better with radiologists’ assessments

6

07/14/2015 21

Textures VS dose/time

Co-occurrence

Run-length

Global

!"#$

!%$

#$

%$

"#$

"%$

&#$

&%$

%$ "%$ &%$ '%$ (%$!"#$%

&'(")*+,

"-%.(/0&("

"1+.("2345"

3678"0&(*9%" )*+,-$

.$)/0123$

!"#$

!%&$

!%#$

!'&$

!'#$

!&$

#$

&$ '&$ %&$ "&$ (&$

!"#$%

&'(")*+,

"-%.(/0&("

1+.("2345"

,(%&"67" )*+,-$

.)$/012,-$

07/14/2015 22

Textures VS biomarkers

CTCv4 fibrosis group (N=6) No fibrosis group (N=14)

07/14/2015 23

PET/MR fusion for Sarcoma mets to the lungs No MetsLungs

With MetsLungs

Carrier-Vallieres et al., PMB, 2015

PREDICTION (on BOOTSTRAP

SAMPLES) AUC = 0.984 ± 0.002

Sensitivity = 0.955 ± 0.007 Specificity = 0.926 ± 0.004

A total of 27,405 feature combinations were extracted from 51 patients

07/14/2015 24

Methodology

7

07/14/2015 25

Fused versus Separate Scans

Do texture features extracted from FUSED scans provide better assessment of

tumor aggressiveness

than those extracted from SEPARATE

scans ??

QUESTION FUSION EXAMPLE FDG-PET

MRI T2FS

FDG-PET/T2FS

07/14/2015 26

MULTIVARIABLE ANALYSIS���(Identifying optimal parsimonious model)

FEATURE SET

REDUCTION FEATURE

SELECTION PREDICTION

PERFORMANCE ANALYSIS

~ 10 000 FEATURES

FINAL MODEL COMPUTATION (LR coefficients)

Set balanced between: - Predictive power - Maximum info

- Varying initialization - Maximization of AUC

- ROC analysis - Correction for small sample size effect - Choice of best model

07/14/2015 27

FEATURE SET REDUCTION41 textures * 240 extraction parameter combinations:

~10 000 texture-parameter features

~ INITIAL FEATURE

SET

Part 2 Interdependence of

feature j with already chosen

features

Part 3 Interdependence of feature j with features not yet

removed

Goal: Allow the creation of a feature set balance between predictive power and maximal information (Gain)

Part 1 Prognostic value of

feature j

07/14/2015 28

FEATURE SET REDUCTION���Part 1 – Prognostic value

rs(xj,y) : Spearman’s rank correlation between

feature j and outcome y (Lung Mets)

For each bootstrap sample, calculate a new rs(xj*b,y) from the training set. Repeat for 1000 bootstrap

samples and record the mean.

8

07/14/2015 29

FEATURE SET REDUCTION���Part 2 – Interdependence with selected features

PIC = 1-MIC Potential information

coefficient

MIC : Maximal Information coefficient (Reshef et al., Science 334, 2011). Has the ability to capture any simple of complex relationship types of association between two variables (independent to the modeled outcome)

07/14/2015 30

FEATURE SET REDUCTION���Part 3 – Interdependence with unselected features

PIC = 1-MIC Potential information

coefficient

MIC : Maximal Information coefficient (Reshef et al., Science 334, 2011). Has the ability to capture any simple of complex relationship types of association between two variables (independent to the modeled outcome)

07/14/2015 31

Model Order Selection

Most parsimonious model (set of

features)

07/14/2015 32

COMPLETE PREDICTION MODEL

1. PET/T2FS -- SZE : Small Zone Emphasis texture extracted on fused PET/T2FS scans2. PET/T1 -- ZSV: Zone Size Variance texture extracted on fused PET/T1 scans 3. PET/T1 -- HGZE: High Gray-Level Zone Emphasis extracted on fused PET/T1 scans 4. PET/T2FS -- HGRE: High Gray-Level Run Emphasis extracted on PET/T2FS scans

IDENTIFIED SET OF FEATURES

9

07/14/2015 33

PREDICTION MODEL RESPONSE

PREDICTION (on BOOTSTRAP SAMPLES)

AUC = 0.984 ± 0.002 Sensitivity = 0.955 ± 0.007 Specificity = 0.926 ± 0.004

Carrier-Vallieres et al., PMB, 2015 https://github.com/mvallieres/radiomics 07/14/2015 34

STAMP (Simulator for Texture Analysis in MRI/PET)

Vallieres, Laberge et al., AAPM, 2014

07/14/2015 35

Conclusions

• Treatment outcomes are multifactorial (Pan-Omics)– Combination of physical (radiomics) and biological

(radiogenomics) factors

• Radiomics is an essential element of the Pan-Omicsworld and constitute a powerful tool to interrogatewealthy imaging imaging information– Single and multiple modalities

• Separate and fused

• Radiomics involves two main steps– Robust feature extraction– Robust modeling

07/14/2015 36

ACKNOWLEDGMENTS

• Joseph Deasy, PhD• Carolyn R. Freeman, MBBS • Jan Seuntjens, PhD

• Sonia R. Skamene, MD

• Sangkyu Lee, PhD candidate

• Martin Carrier-Vallieres, PhD Candidate