WE-B-AUD C CE-Therapy

July 30, 2008

Dose calculation
algorithms in
3DCRT and IMRT
Tom Knöös, Lund,
Sweden
Brendan McClean, Dublin,
Ireland 1

Objectives
1. To provide an educational review of the physics and
techniques behind convolution algorithms

2. To review the methods used to improve the simulation

efficiency i.e. pencil beam and collapsed cone convolutions

3. To briefly review the performance of codes currently used

for clinical treatment planning.

4. To discuss the issues associated with experimental

verification of dose calculation algorithms.

5. To briefly review the potential clinical implications of

accurate calculated dose distributions.

1
1
The Problem e-

• Modelling the linac

– Energy fluence
• Source models
• Monte Carlo
• Modelling of dose in
patients
– Interpolation and correction
of measured data
– Fluence to dose modelling
– Monte Carlo

Fluence to dose 1
Convolution

= ⊗ [1D convolution]

D( x) = ∫ T ( x′) ⋅ K ( x − x′)dx′
D (r ) = T (r ) ⊗ K (r )

This idea was explored by several papers at the ICCR 1984

4

2
Modelling primary photon 1
energy fluence and loss
Source
• Ray-tracing Total Energy
Released in Mass (TERMA)
• Similar to determining
effective or radiological
depth

μE μ Fan ray lines

−μ E ⋅ z eq
T (E, z) = ⋅ Φ ( E , z ) ⋅ E = E Φ 0 ( E ,0 ) ⋅ e ⋅E
ρ ρ
z
1
zeq =
ρ water ∫ ρ ( z ' ) ⋅ dz '
0
5

1
Modelling dose deposition
• Dose distribution around a
single interaction point
– Point dose kernel
• Separate primary, 1st scatter,
2nd scatter, multiple residual
scatter dose kernels

3
0.5 cm Primary
1

1.25 MeV

First scatter
10 cm

10 cm Total

Data available in
spherical coordinates
along 48 directions

Mackie et al 1988, PMB 7

1
Convolve!
• Apply the dose kernel to each TERMA
point
• Integrate over the whole volume i.e.
a convolution

= ⊗ [1D convolution]

D( x) = ∫ T ( x′) ⋅ K ( x − x′)dx′
D (r ) = T (r ) ⊗ K (r )
8

4
1
Convolution in 2D

⊗ =

Dose
TERMA ⊗ Deposition = Absorbed Dose
Kernel

[ ]

Convolution is efficiently solved by Fast Fourier Transform techniques

9

Example: Point kernel 1

convolution - CMS
• Re-sampling of Mackie’s kernels to
Cartesian coordinates
• FFT solution
• Two separate calculations:
– A primary kernel for which the calculation is
performed at high-resolution but over a small
region – high gradient – short range
– A scatter kernel, where the calculation is
performed at a lower resolution but over a
larger area – low gradient – long range
– Time saving of about 65 % by this technique

10

5
1
Limitations of convolution
• Kernels are not invariant in space
– Energy distribution varies with position in
beam
• Beam softening laterally
• Beam hardening longitudinally
• Kernels vary with density
• Divergence leading to tilted kernels
• Pre-calculated kernels won’t make it!!!
• FFT not suitable – analytical methods
must used – time consuming

• Approximate methods required

11

1st approximation 2
Pencil Beam
• Reduce the dimensionality of the problem by pre-
convolving in the depth dimension

1.0
0.9
0.8
Relative fluence

0.7
0.6
0.5
0.4
⊗ =
0.3
0.2
0.1
0.0
0 10 20 30 40 50
Depth (cm)
• => Pencil beams (PB)
• Superposition of pencil beams in 2D => Faster

12

6
2
Illustration of Pencil Beam
superpositioning (convolution)

⊗ =

Dose
Energy fluence ⊗ Deposition = Absorbed Dose
Kernel

13

Construction of Pencil Kernels

De-convolved from
measurements

From measurements
by differentiating

Calculated by
Monte Carlo
methods

14

7
Example: Pencil beam 2
model – Nucletron
• Pencil beams based on MC calculated
point kernels, integrated and fitted
to a limited number of depth doses
• Separates “primary” and “scatter”
dose
• Heterogeneities handled via effective
path length – only longitudinal
scaling
• Extensive beam modelling
Nucletron (former MDS Nordion, Helax-TMS) 15

Example: Pencil beams 2

model - Eclipse
• Uses pencil beams extracted
from measurements (SPB) or
from Monte Carlo calculation
(AAA)
• Heterogeneities handled via
effective path length –
longitudinal
• AAA adds a scaling of the
spread of the pencil based on
density – lateral
• AAA also have an extensive
beam modelling
Analytical Anisotropic Algorithm
16

8
2nd approximation 2
Collapsed cone convolution

Collapsed Cones
Continuous Dose Spread Kernel Discrete Dose Spread Directions

Quantization in cones

Kernels are discretised

Collapsing removes the inverse
square law – only exponential
attenuation is left
17

Number of collapsed cones 2

or directions
• Sufficient density of cones to
distribute energy to all voxels
– Not possible but at least while the
energy is significant
– ~100 (Mackie et al, 1996 Summer
school)
– Voxels will be missed at large distances
– very low energy contribution
• 128 CC are used in CMS (48 for the
fast version)
• 106 CC are standard in OMP
18

9
2
Implementation issues
Accounts for
-Heterogeneities
Kernels scaled for different tissues
-Lateral energy transport θ
δ
-Beam Hardening and Off-axis spectrum
softening
Included in Ray Trace process θ
- Tilt of kernels
Included in Transport

HVL(0)/HVL Lund / Dublin group, 2008

1.20
Polyenergetic Spectrum accounted for by
Regular beam -
Monte Carlo data
Regular beam -

weighted sum of monoenergetic kernels

1.15 Tailor et al.
Regular beam-
Measurements

calculated by Monte Carlo 1.10 FFF - Monte

Carlo data
FFF -
1.05 Measurements

Weights determined by comparison 1.00

with measured data 0.95

Angle (deg.)
0.0 2.5 5.0 7.5 10.0 12.5

19

2
Examples: Collapsed cone
• Pinnacle These are
– Polyenergetic weighted kernels, total energy ‘iso-scatter’
– Off-axis/tilting considered during TERMA lines.
– Collecting dose or dose point of view They link
• CMS points
– Two kernels, Primary electron dose and producing
scattered photon dose equal scatter
– No Off-axis/tilting to here.
– Collecting dose or dose point of view
• Nucletron
Primary
– Two Kernels are used:
interaction
• One for Collision Kerma into Primary Dose
• One for ‘Scerma’ into Phantom Scatter Dose point
– Kernels parameterised and fitting parameters These are
stored for run time ‘isodose’
– Off-axis/tilting lines
From Deshpande, Philips

20

10
2
Further approximation
• Multigrid solution (CMS)
– Only calculate dose using superposition at points where it is
necessary, and at all other points use interpolation to get a reasonable
estimate of dose
• Adaptive CCC (Pinnacle)
– Only performs convolution at every 4th point in the TERMA array
– Gradient search performed on TERMA array
– Dose in between is interpolated if gradient low (i.e TERMA doesn’t
change much)
– Convolution performed at every point if TERMA gradient high

Example from CMS

21

3
Conclusions
• Inhomogeneities are handled by scaling
the kernels rectilinearly with electron
density according to the theorem by
O’Connor 1957
• Type a – Models primarily based on EPL
scaling for inhomogeneity corrections.
– Eclipse/SPB, OMP/PB, PPLAN, XiO/Convolution
– LONGITUDINAL scaling
• Type b – Models that in an approximate
way consider changes in lateral electron
transport
– Pinnacle/CC, Eclipse/AAA, OMP/CC,
XiO/Superpositioning.
– LONGITUDINAL and LATERAL scaling
22

11
Performance of
convolution
models

23

Comparison in homogeneous 3
water phantoms

All systems are expected to work excellent in homogenous water

Knöös et al, 1994, PMB 24

12
Pencil beam
calculations in a 3
blocked fields

From van’t Weld, 1997,

Radioth Oncol

From Storchi and Woudstra, 1996, PMB

From Van Esch et al, 2006, Med Phys

25

AAA-PB model in Ecplise 3

Elekta
Carefully implemented
algorithms together with
accurate beam models
works for most linacs

•Gamma-analysis, calc-meas
•Inside field after buildup
Siemens •Less than 0.5 % of points
outside 3 mm/1 %
• One implementation 0.7 %

Varian

Cozzi et al, 2008, Z Med Physik

26

13
A problem using pencil beams 2
Irregular geometries

The same dose to in all geometries since the PB is

pre-integrated to a certain depth/length

See also Hurkmans et al, 1986, RO 27

Convolution methods in 3
homogeneous water
• Differences in beam modelling
– Head scatter
– Electron contamination
– Wedges/Blocks
– MLC
• May lead to slightly different accuracy
• Basically all models perform well in water
– Point, pencil or collapsed cone
implementations

28

14
Single beam
Comparison in 3
inhomogeneous phantoms

6 MV 15 MV

6 MV 15 MV

From Fogliata et al 2007, PMB Density 0.2 g/cm3 29

Two beams
PB with longitudinal (PB) 3
plus lateral scaling (AAA)
Pencil beam - NC-No Correction, MB-Modified Batho: Both without lateral scaling
AAA- with lateral scaling

6 MV 10 MV

Med Phys 2007

30

15
Multiple beams
PB w/wo lateral
scaling and CC vs MC

1.0 0.2 1.0 0.2

0.4 0.1 0.4 0.1

6 MV 18 MV
From Lasse Rye Aarup, Copenhagen 31

Tangential treatment of breast 3

XiO/Conv Eclipse/AAA
Knöös et al, 2006, PMB

Eclipse/ModBatho Pinnacle/CC

OMP/PB XiO/Super

32

16
Tangential treatment of breast 3
XiO/Conv Eclipse/AAA

Average Average
Knöös et al, 2006, PMB

values for values for

6 MV type a type b
PTV Mean 100 99.3
Eclipse/ModBatho Pinnacle/CC
PTV D95 91.0 90.4

PTV D5 108.8 108.9

PTV D5-D95 17.8 18.5

Pulm sin D5 92.6 83.1 High dose area

Pulm sin
OMP/PB D50 3.3XiO/Super 4.0

DXX is the dose level that encompasses XX % of the volume

33

5 field 18 MV – lung 3
XiO/Conv Eclipse/AAA
Knöös et al, 2006, PMB

OMP/PB Pinnacle/CC

34

17
5 field 18 MV – lung 3
XiO/Conv Eclipse/AAA

6 MV 18 MV
Average Average Average Average
values for values for values for values for
Knöös et al, 2006, PMB

type a type b type a type b

PTV Mean 100 97.5 100 96.3

PTV D95 ~max 92.7 91.3 95.2 91.5
PTV D5 ~min 106.2 102.8 104.4 99.8

PTV D -D 13.5 11.6 9.2 8.3

OMP/PB5 95 Pinnacle/CC
Pulm Sin D50 14.2 19.7 15.7 20.6
Pulm Sin D5 103.7 96.3 101.3 91.9
Pulm Sin D1 107.9 100.0 104.3 95.9

35

Results from RPC

thorax phantom

• 15 cases planned with type a

– 84% ± 16% of the pixels met the
criteria (5%/5mm)
• 30 cases planned with type b
– 99% ±4% of the pixels met the
criteria (5%/5mm)

AAPM 2008 TU-C-AUD B-3 P Alvarez et al

36

18
3
Conclusions – Dose changes
• Prostate • Lung - PTV
– non-significant – 2-4 % lower average dose
• H&N – Wider penumbra
– none (depending on • Lung (treated side)
accuracy of scatter – 10 % lower dose to the
integration) and air highest irradiated parts of
cavities (air or low dense the lung
water)
– 5 % higher dose (15 =>
• Breast 20 %) to the lung (D50)
– slightly lower dose to • Lung (healthy side)
breast and especially in
– Average dose identical
lung in proximity to the
(9.8-10.7 %)
target however larger
irradiated lung volume

Knöös et al, 2006, PMB

37

How to verify dose 4

calculation algorithms?

Med Phys 1998

38

19
4
Process of acceptance

User 3) Specific
3) Specific performance
performance in
in
users environment
users environment

Vendor/user 2) Generic
2) Generic performance
performance in
in
users environment
users environment

Vendor 1) Generic
Generic performance
performance
1)

ESTRO booklet no 7, 2004 39

4
Commissioning
STEP 1
STEP 2
STEP 3
STEP 4
Verification
STEP 6 STEP 5

Verification of dose calculation

Wieslander 2007 40

20
MC methods facilitate verification

➋➌➍➎➏➐➑
➋➌
➎
➏➐➑

41

Implications of introducing new 5

and more accurate algorithms
• Significant changes in • Careful analysis of
dose to target changes is required
volumes and before adopting new
surrounding tissues algorithms
especially when lung – Retrospectively re-
is involved calculate plans when
– Consequences for clinical outcome is
assessment of dose- known?
effect relationships – Construct new plans
with older algorithms
and re-calculate?
– New plans with old
prescriptions and new
algorithms?
– Optimize plans to the
same biological effect
on PTV and/or OAR?

Morgan et al 2008 42

21
Implications of introducing new 5
and more accurate algorithms
• Significant changes in • Careful analysis of
dose to target changes is required s
volumes and before adopting
i c istnew
e
surrounding tissues h y s d th
algorithms
especially when lung n p stan cere-
– Retrospectively
e e s
is involved e r
w nd outcome
calculate e n
plans when
– Consequences for d bety clinical
u se qu is
e l l n
known?
assessment of dose-ed co
e to futia–l Construct
i n
effect relationships
s ists en
new plans
with older algorithms
n t
u s siocolog d po and re-calculate?
c n n
Dis nd oects a
– New plans with old
prescriptions and new
a e ff algorithms?
– Optimize plans to the
same biological effect
on PTV and/or OAR?

Morgan et al 2008 43

Conclusion
• Convolution methods are accurate
– For low density regions – use models
with lateral scaling
• Verification
– Also Vendor’s responsibility!
• Be careful when transferring to more
accurate models but…
• Important to do this!
44

22