ISSSMA 2013 - June 3rd 2013 - 1

Advanced quantification in oncology PET

Irène BuvatIMNC – UMR 8165 CNRS – Paris 11 University

Orsay, France buvat@imnc.in2p3.fr

http://www.guillemet.org/irene

ISSSMA 2013 - June 3rd 2013 - 2

Two steps

Radiotracer concentration

(kBq/mL)

• Glucose metabolism• Metabolically active tumor volume• etc…

Static imagingDynamic imaging

ISSSMA 2013 - June 3rd 2013 - 3

• Tumor segmentation

• Identification of indices that best characterize the tumor in a specific context

• Interpretation of tumor changes during therapy

• Understanding the relationship between macroscopic parameters (from PET images) and microscopic tumor features

Quantification issues in oncology PET

ISSSMA 2013 - June 3rd 2013 - 4

• Tumor segmentation

• Identification of indices that best characterize the tumor in a specific context

• Interpretation of tumor changes during therapy

• Understanding the relationship between macroscopic parameters (from PET images) and microscopic tumor features

Quantification issues in oncology PET

ISSSMA 2013 - June 3rd 2013 - 5

Current quantification in oncology PET

Tracer uptake (kBq/mL)

SUV (Standardized Uptake Value)=

Tracer uptake

Injected activity / patient weight

SUV ~ metabolic activity of tumor cells

ISSSMA 2013 - June 3rd 2013 - 6

Comparing 2 PET scans : current approach

• Need to identify and possibly delineate the tumors

• Each tumor = 1 single SUV

• Change compared to an empitical threshold (provided in recommendations such as EORTC, PERCIST)

• Tedious when there are many tumor sites

~12 weeks

PET1 PET2

SUV=10SUV=5

SUV=5SUV=2

ISSSMA 2013 - June 3rd 2013 - 7

A novel parametric imaging approach

Goal : Get an objective voxel-based comparison of 2 PET/CT scans

~ 12 weeks

PET1 PET2

ISSSMA 2013 - June 3rd 2013 - 8

1. PET image registration based on the CT associated with the PET scans

PET2PET1

3. Identification of voxels in which SUV significantly changed between the 2 scans using a biparametric analysis

Main steps

2. Voxel-based subtraction of the 2 image volumes

PET2PET1 PET1-PET2’

- T21 =

ISSSMA 2013 - June 3rd 2013 - 9

VOI selection

Step 1

Registration of the PET volumes using the T21 transformation

Identification of the transformation needed to realign the 2 CT

T21

(rigid transform using Block Matching)

CT2 CT2

CT1 CT2 PET1 PET2

PET1PET2

T21 realigned with

ISSSMA 2013 - June 3rd 2013 - 10

Subtraction of the 2 realigned PET scans

Step 2

PET1/CT1T21{PET2/CT2} - = T21{PET2} – PET1/CT1

Each point corresponds to a voxel

ISSSMA 2013 - June 3rd 2013 - 11

Identification of the significant tumor changes in the 2D-space by solving a Gaussian mixture model

Step 3

xiPET1(i)-PET2(i)

PET1(i)

f(xi|) = pk (xi|k,k)k=1

K

mixture densitymixture parameters

2D Gaussian density

vector of parameters (p1, … pK, 1, ….K, 1, … K)

ISSSMA 2013 - June 3rd 2013 - 12

Step 3: results

0

-3

-6

SUVParametric image

V: volume with a significant change

SUV: change magnitude

Nb of voxels[P

ET

1 -

T21{P

ET

2}] S

UV

[PET1] SUV

Background

Physiological changes

Tumor voxels

ISSSMA 2013 - June 3rd 2013 - 13

Example

Identification of small tumor changes (lung cancer)

Progressive*

PET3T21{PET2} - PET1after solving the GMM

+1.3

PET3PET1 T21{PET2} T21{PET2} - PET1after solving the GMM

+1.6

PET1 T21{PET2}

ISSSMA 2013 - June 3rd 2013 - 14

Clinical validation : 28 patients with metastatic colorectal cancer

• All tumors identified as progressive tumors at D14 were confirmed as such 6 to 8 weeks after based on CT (RECIST criteria)

• Among the 14 tumors identified as progressive tumors by RECIST criteria, 12 were identified as such at D14 using PI while only 2 were identified using EORTC criteria (SUVmax)

Necib et al, J Nucl Med 2011

78 tumors with 2 PET/CT (baseline and 14 days after starting treatment)

NPV PPV Sensitivity* SpecificityEORTC 91% 38% 85% 52%PI 100% 43% 100% 53%

* for detecting lesions

ISSSMA 2013 - June 3rd 2013 - 15

time (weeks)

0 12 23 35 46

Longitudinal study

Problem : characterize the tumor changesNo method, each scan is usually compared

only to the previous one

Comparing more than 2 PET/CT scans

ISSSMA 2013 - June 3rd 2013 - 16

First step: PET image registration based on the associated CT

A parametric imaging solution

…

T21

T31

…

TT2121

TT3131

0 12 23 35 46

Rigid transformationRigid transformationBlock MatchingBlock Matching

Use of the transformation identified based on the CT to register the PET scans

Time (weeks)

ISSSMA 2013 - June 3rd 2013 - 17

Model : a factor analysis model

SUV units

Decreasing uptake over time

f2(t)

Increasing uptake over time

f3(t)

time

time t

I1(i). + I

2(i). + I

3(i).

tStable uptake

over timef1(t)

SUV

voxel i

SUV (i, t) Ik (i).k1

K

fk (t)k (t)

t

ISSSMA 2013 - June 3rd 2013 - 18

Priors :- Non-negative Ik(i) coefficients- Non-negative fk(t) values- In each voxel, the variance of the voxel value is roughly proportional to the mean

Solving the model

Iterative identification of Ik(i) et fk(t) (Buvat et al Phys Med Biol 1998) using a Correspondence Analysis followed by an oblique rotation of the orthogonal eigenvectors

SUV (i, t) Ik (i).k1

K

fk (t)k (t)

ISSSMA 2013 - June 3rd 2013 - 19

Sample results

Lung cancer patient with 5 PET/CT scans

0 12 23 35 46 weeks

Normalized SUV

ISSSMA 2013 - June 3rd 2013 - 20

Why is such an approach useful?

Heterogeneous tumor responses can be easily identified

0 12 23 35 46

weeks

Normalized SUV

ISSSMA 2013 - June 3rd 2013 - 21

Sample results: early detection of tumor recurrence (1)

12 23 35 46 weeks

T1

T2 2 cyc-4 cyc-3 cyc-no cycle

12 weeks

2 cycles

T1

T2

12 23 35 weeks

2 cyc - 4 cyc - 3 cycles

T1

T2

12 23 weeks

2 cycles - 4 cycles

T1

T2

ISSSMA 2013 - June 3rd 2013 - 22

12 23

weeks

2 cycles - 4

cycles

BT2

T3

EORTC : no tumor recurrence detected at PET3

SUVmean

0

2

4

6

8

10

12

14

1 2 3 4 5

Examen

SUVmoy T2

SUVmoy T3

lol

PET scan

PI : tumor recurrence detected at PET3

Sample results: early detection of tumor recurrence (2)

ISSSMA 2013 - June 3rd 2013 - 23

Discussion / conclusion

- No need to precisely delineate the tumors

- Makes it possible to detect small changes in metabolic activity

- Summarizes changes between two or more scans in a single image

- Shows heterogeneous tumor response within a tumor or between tumors

ISSSMA 2013 - June 3rd 2013 - 24

Thanks to

Hatem Necib, PhDJacques Antoine Maisonobe, PhD student

Michaël Soussan, MD

Camilo Garcia, MD, Institut Jules Bordet, BruxellesPatrick Flamen, MD, Institut Jules Bordet, Bruxelles

Bruno Vanderlinden, MSc, Institut Jules Bordet, BruxellesAlain Hendlisz, MD, Insitut Jules Bordet, Bruxelles