Outcome-Driven
Automated Treatment Planning
Medical College of Wisconsin
AAPM, July 28, 2008, MO-D-AUD A-3
BGRT
Additional Biological
Assays and Imaging
(during treatment)
Outcomes, Imaging,
Assays (after
(after treatment)
Refined biological
parameters
Outcomes, Imaging,
Assays (after
(after treatment)
PopulationPopulation-based
prescription
(Phase I)
Stewart & Li, MP, 2007
TCP (1 – NTCP)
Normal Tissue Complication
Probability (NTCP)
Radiation Dose
Why use outcome models?
Outcomes, Imaging,
Assays (after
(after treatment)
Analysis
(assays, images, outcomes)
Biological Assays and
Additional Imaging
(before treatment)
Tumor Control
Probability (TCP)
Individualized
adaptive BGRT
(Phase III)
Refined biological
parameters
Boost dose to selected
tumor regions
(Phase II)
Probability of Outcome
X. Allen Li
Goal of Radiation Therapy
Cell culture
and animal
experiments
Refined biological
parameters
New and refined
biological models
• To fully describe responses as a function of
any dose to any volume
• To predict responses based historical data
• To supplement or replace dose-volume
criteria for plan optimization and
evaluation.
Define tumor
targets and organs
at risk
Anatomical and
Biological Imaging
(before treatment)
1
Model parameterization based on clinical data
Breast cancer
Outcome modeling for treatment planning
•
•
•
•
•
Survival probability (LQ)
TCP (Poisson model)
NTCP (LKB, Serial, Parallel)
EUD for both tumors and normal tissues
Clinical Response Models (Maximum likelihood analysis)
Problems:
Three Clinical studies:
•
Resch et al (2002): BCS then
48Gy + 20Gy LDR or 52 Gy +
9.7Gy HDR. Same TCP.
•
Fourquet et al (1995):
(1995): RT alone
58Gy + 20Gy boost using 192Ir
LDR or 60Co EBRT
TCPIr=76% vs TCPCo=61%.
•
Mazeron et al (1991):
(1991): RT alone
45Gy + 37Gy 192Ir LDR
R=0.32R=0.32-0.49Gy/h TCP=60%
R=0.5R=0.5-0.59Gy/h TCP=72%
R=0.6TCP=84%
R=0.6-0.9Gy/h
Still phenomenological rather than predictive
Unreliable model parameters (QUANTEC mission)
Prostate cancer
α = 0.3 Gy -1
α / β = 10 Gy
T rep = 1 hour
Guerrero & Li, PMB 3307,2003
Malignant
gliomas
• α = 0.15±
±0.05 Gy-1
MG
• α/β
β = 3.1 ± 2.0 Gy
• Clonogenic cell
number: 106~107
= 0.06 ± 0.05 Gy-1
/ = 10.0 ± 15.1 Gy
Grade 1&2
= 0.35 ± 0.07 Gy-1
/ = 4.3 ± 5 Gy
Grade 3
= 0.11 ± 0.10 Gy-1
/ = 5.8 ± 11.8 Gy
Grade 4
= 0.04 ± 0.06 Gy-1
/ = 5.6 ± 9.4 Gy
Wang, Guerrero & Li, IJROBP 2003
Qi, Schultz, Li , IJROBP, 2006.
2
LKB NTCP: Lung
Liver Cancer
= 0.029±0.004 Gy-1
/ = 9.9±1.8 Gy
Td = 100±18 days
Patient
no.
Median
Dose
(Gy)
Fraction
scheme
(Gy/fx)
Reference
Liang
128
53.6
4.88
Cancer
Vol103,218
(2005)
Dawson
128
61.5
1.5
J. Clin Onco
Vol18, 2210
(2000)
Int. J. Rad.
Onco Biol.
Phys. 55
329 (2003)
Reference
Lung
Burman et al. 1991
Martel et al. 1994
Kwa et al. 1998
n
m
TD50 (Gy)
0.87
0.87
1
0.18
0.18
0.30
24.5
28
30.5
Pneumonitis
SWOG grade
SWOG grade
Seppenwoolde et al.
2003
Moiseenko et al. 2003
0.99
1
1.02
0.37
0.28
0.26
30.8
43
21.0
SWOG grade 2 RP
SWOG grade 3 RP
Symptomatic pneumonitis
0.80
0.37
21.9
Radiographic and symptomatic
pneumonitis
Observed RP Rate
0.7
0.6
0.5
55
1.8
Seong M
51
45
1.8
0.2
1.8
0.1
32.5
1 RP
2 RP
Graham et al. 1999 (Washington U) - RTOG grade >=2
Seppenwoolde et al. 2003 (Netherlands) - SWOG grade >=2
Moiseenko et al. 2003 (Canada) - RTOG grade >=2
Willner et al. 2003 (Germany) - NCI CTC grade >=2
Kim et al. 2005 (Korea) - RTOG grade >=3
Yorke et al. 2005 (MSKCC) - RTOG grade >=3
Chang et al. 2006 (U of Florida) - NCI CTC grade >=2
Maximum likelihood fit - all RP
0.8
83
24
Fractionation Scheme
1.8-2 Gy q.d.
1.8-2 Gy q.d.
1-2.7 Gy q.d.; normalized to 2
Gy/fr using / of 2.5 or 3 Gy
1-2.7 Gy q.d.; normalized to 2
Gy/fr using / of 2.5 or 3 Gy
1-2 Gy q.d.; normalized to 2
Gy/fr using / of 3 Gy
1
0.9
Seong H
Seong L
Endpoint
0.4
0.3
n=1
m = 0.39
TD50 = 28.6 Gy
0
0
Tai et al, IJROBP, 2008
Liver NTCP: BED=D*(1+d/ / +f*N)
D: total dose, d: fraction dose, N: # of fractions
10
20
30
40
50
Mean Lung Dose (Gy)
Semenenko & Li 2007
Use of outcome models in
computerized treatment planning
• Plan evaluation
• Plan optimization
Tai et al, 2008
3
Equivalent Uniform Dose
Problems to evaluate complex plans with DVH
•
•
•
•
•
Complicated anatomy, multiple OARs
Complicated/crossing DVHs
Difficult for visual inspection
Plan merit not quantified
DVH failure for spatial tumor
heterogeneity
EUD: the dose that, if distributed uniformly, will lead
to the same biological effect as the actual non-uniform
dose distribution. ………………..Niemierko. MP. 1997
S =
Vo
i
Vi
S (Di )
V0
S = exp( − (α ⋅ EUD + β d ⋅ EUD − 1 . 4 γ
Vi: a volume
element
Quantitative evaluation and comparison of
complicated plans based on biological
effectiveness are desirable.
Alternatively:
EUD =
EUD
))
d
− ln( S )
α + β d − 1 .4 γ / d
EUD =
a
i
vi D
1
a
,
i
EUD-based Figure-of-merit index (fEUD)
fEUD
=
Plan Optimization
1 .0
n
1 .0 + k ⋅ i = 1
m
j =1
ω ⋅ EUD i
i
OAR
•
•
•
•
ω j ⋅ EUD j
Tumor
n, m : number of OARs and targets;
ωi, ωj : weighting factors for each OAR and target;
k : the relative importance factor between tumor
and OAR.
•
Mathematical forms of treatment goals
Increase if goal is not met
Good if value less than or equal to 0
Physical (dose-based) cost functions
•
•
•
• Condensing complex DVHs into one # (range: 0-1)
Cost Functions
Overdose/underdose volume constrains
Maximum/minimum doses
Biological (dose-response model based) cost function.
•
•
Target/OAR EUDs
TCP/NTCP.
• The larger fEUD, the superior the plan
4
Models used in Monaco
Two commercial biological TPS
Model
Name /
description
CMS Monaco
Phillips Pinnacle
Tumor Poisson cell
kill model
OAR
Parameters required
Comments
1. Cell sensitivity (0.1-1.0)
2. EUD prescription (Gy)
Mandatory cost
function for
targets; no
penalty for hot
spots
Serial
1. Power law exponent a ( 1) Penalizes for hot
complication 2. EUD (Gy)
spots
model
Parallel
1. Reference dose (Gy)
Effective for
complication 2. Power law exponent a ( 1) reducing mean
model
3. Mean organ damage (%)
organ dose
Monaco
Models used for optimization in Pinnacle
Structure
Target
OAR
model
Parameters
Objectives/
constraints
Comments
Min EUD
Volume
parameter (a<1)
EUD
Penalizes for too low EUD
Target
EUD
Volume
parameter (a<1)
EUD
Penalizes for any deviation
from the desired EUD
Max EUD
Volume
parameter (a 1)
EUD
Penalized for too high EUD;
can be used with both serial
and parallel structures
5
Models used for plan evaluation in Pinnacle
∏ ∏(
PTCPNTCP
TCPTCP
NTCPNTCP
11
+ =−= =−−
i
∏
TCPTCP
=
i
i
Case 1 – H&N (XiO
(XiO IMRT vs Monaco)
)
i
i
i
Case 1: H & N
70
60
90
80
Monaco
XiO
Monaco
XiO
70
Max Dose (Gy)
50
Min Dose (Gy)
Case 2: H&N (Tomo
(Tomo vs Monaco)
40
30
20
60
50
40
30
20
10
10
0
0
le
ti d
tid
aro paro andib
L
M
Rp
0
4
0
rd
Co PTV5 PT V5 PTV7
Rp
80
Monaco
XiO
tid rotid dible
a
n
Lp
Ma
0
0
4
rd
Co PTV5 PTV5 PTV7
60
50
50
40
30
20
40
30
20
10
10
0
le
tid
ti d
aro paro andib
L
M
Rp
Monaco
XiO
70
60
gEUD (Gy)
Mean Dose (Gy)
70
aro
80
0
4
0
0
rd
Co PTV5 PT V5 PTV7
Rp
aro
tid rotid dible
a
n
Lp
Ma
0
0
4
rd
Co PTV5 PTV5 PTV7
6
Case 3 – Chest Wall (Tomo
(Tomo vs Monaco)
Pinnacle vs Monaco: Prostate
100
Xio
Monaco
80
Percent volume (%)
Rectum
Pinn_phy
Pinn_bio
60
PTV73.8
Bladder
40
20
0
0
20
40
60
80
100
120
Percent dose (%)
Monaco vs Pinnacle: H & N
Pinnacle vs. Monaco: fEUD
100
XiO Monaco Pinn_Phy Pinn_Bio
Xio
Monaco
80
Pinn_phy
Percent Volume (%)
Pinn_bio
R-Partoid
PTV
60
fEUD
(H&N)
40
Cord+1cm
20
0.38
0.42
0.37
0.40
fEUD
0.26
(Prostate)
0.31
0.27
0.29
0
0
20
40
60
80
100
120
Dose (Gy)
7
Monaco vs Tomo vs XiO
H&N site 1 – oropharynx
H&N Monaco vs Tomo vs XiO: fEUD
Site
Comparison of DVHs in Monaco (solid lines), Tomo (dashed lines) and XiO (dotted lines) plans.
Monaco
Tomo
XiO
1. Oropharynx
0.51
0.49
0.44
2. Paranasal sinus
0.12
0.13
0.08
3. Nasal cavity
0.12
0.19
0.16
4. Oral cavity
0.52
0.50
0.47
5. Larynx
0.771
0.766
0.73
6. Nasopharynx
0.27
0.25
---
Greatest fEUD
Monaco vs Tomo vs XiO
Heterogeneity Index (HI) for primary target
Site
Monaco
Tomo
XiO
1. Oropharynx
1.16
1.03
1.13
2. Paranasal sinus
1.19
1.08
1.22
3. Nasal cavity
1.45
1.06
1.31
4. Oral cavity
1.23
1.04
1.14
5. Larynx
1.17
1.09
1.15
6. Nasopharynx
1.23
1.15
---
Greatest HI
Smallest HI
Smallest fEUD
Dose escalation to GTV using Monaco
DVHs corresponding to a conventional (GTV dose = 70 Gy; solid lines) and escalated dose (GTV
dose = 84 Gy; dashed lines) plans.
8
Why biologically-based is better
The standard formulation of biological
cost functions
All cost functions can be expressed as
• Since, by definition, there are an infinite # of
DVHs that leads to an EUD for a given organ,
biological cost functions can lead to the desired
EUD directly.
• Can find solutions that may not be apparent to the
user
• Can get the best possible result (not just any
acceptable result) and will get it more quickly and
easily
F=
1
N
N
f(Di )
i=1
where
Di
Dose in volume element i of an organ/tumour
N
f(D)
objective density = badness factor
Number of volume elements
The properties of f fully define the behaviour of the cost function !
Courtesy M. Alber
Bloemfontein 2006
Serial structure (spinal cord, rectum)
How to make a biological cost function:
serial models
badness f
FSU
FSU
FSU
FSU
No way!
Parallel structure
FSU
This as maximum
is where we
get nervous
OK for the
whole volume
FSU
lung, liver
This is the dose
we can accept
in a reasonably
small subvolume
FSU
FSU
Dose
Courtesy M. Alber
Bloemfontein 2006
9
How does a serial complication model
control the DVH ?
In contrast, a quadratic penalty:
The length of
the weight arrow
grows as
or similar functions
Volume
Volume
D k −1
exp(αD)
DVH control
only for doses
greater than threshold
Dose
Dose
Courtesy M. Alber
Courtesy M. Alber
Bloemfontein 2006
Bloemfontein 2006
Not all organs are serial:
parallel complication models
How does a parallel complication model
control the DVH ?
badness f
The length of
the weight arrow
grows as
Here, the
tissue has
lost function
Volume
At this dose,
we begin to see
changes
exp( − D)
2
D))
(1+exp( −
or similar functions
OK for the
whole volume
Dose
Dose
Courtesy M. Alber
Bloemfontein 2006
Courtesy M. Alber
Bloemfontein 2006
10
In contrast, a DVH constraint :
targets
badness f
dose too low
The constraint
controls only a
single point
Volume
This dose we
aim to deliver
exponential
law of
cell inactivation
Dose
If we increase
the dose further,
we do not gain
much, but some
patients may benefit
Dose
Courtesy M. Alber
Courtesy M. Alber
Bloemfontein 2006
Bloemfontein 2006
Dependence of model parameter
Cautions for using BBTP
Phys
cord
80
Pinnacle
a=1.0
a=5.0
a=20
60
• Cold and hot spots (evaluating with DVHs
and 3D dose distribution)
• Sensitivity of model parameters
• Extrapolation/interpolation between
fractionations (EUD, DVH)
• ……
40
20
100
0
0
10
20
30
Dose (cGy)
40
50
Phys
80
Percent volume (%)
Percent volume (%)
100
a=0.1
a=1.0
a=15
60
40
mandible
20
0
0
10
20
30
40
50
60
70
Dose (Gy)
11
AAPM Task Group 166:
Evolution of biologically based TPS
Evolution
stage
Optimization strategy
Evaluation strategy
THE USE AND QA OF BIOLOGICALLY
RELATED MODELS FOR TREATMENT
PLANNING
Example
0
Dose-volume-based
objectives/constraints
DVHs
EUD/TCP/NTCP
The majority of
TPS currently
in use
1
EUD for OARs,
EUD+dose-constraints
for targets
DVHs and/or
relative values of
TCP/NTCP/P+
CMS Monaco
Philips Pinnacle
2
EUD
Absolute values of
TCP/NTCP/P+
Future
developments
3
Absolute values of
TCP/NTCP/P+
Absolute values of
TCP/NTCP/P+
Future
developments
X. Allen Li (Chair)
Markus Alber
Joseph O. Deasy
Andrew Jackson
Kyung-Wook Ken Jee
Lawrence B. Marks
Mary K. Martel
Alan E. Nahum
Andrzej Niemierko
Vladimir Semenenko
Ellen D. Yorke
Summary:
Acknowledgement
Biologically based treatment planning
• Is more effective to generate plans with
better normal tissue sparing
• Needs to be implemented with cautions
• Requires more data/work for outcome
modelling
•
•
•
•
Vladimir Semenenko, Ph.D •
An Tai, Ph.D
•
Jian Wang, Ph.D
•
Mariana Guerrero, Ph.D
•
Sharon Qi, Ph.D
Guangpei Chen, Ph.D
J. Frank Wilson, MD
Chris Schultz, MD
• Kathleen Schmainda, Ph.D • Beth Gore, MD
• Beth Erickson, MD
• Rob Stewart, Ph.D
Members of AAPM TG-166
• Is coming into clinic and is here to stay !
12