IMRT – the inverse

problem and inverse

planning
Laurence Court, PhD
University of Texas MD Anderson Cancer Center
[email protected]

Prepared for: School on Medical Physics for Radiation Therapy: Dosimetry and Treatment
Planning for Basic and Advanced Applications, April 2017

1
Conflicts of interest
• Court receives funding from NIH, CPIRT, Varian and
Elekta

2
IMRT is 35 years old this year!

Brahme A, Roos JE, Lax I. Solution of an integral equation encountered in

rotation therapy. Phys Med Biol1982;27:1221–9.

3
Introduction to IMRT and
the inverse problem

4
 6
Slide from Charlie Ma
 7
Slide from Charlie Ma
Simple Example of Optimization
Assume that intensity's add and no attenuation

Beam 1

0 0 0 100 100 100 0 0 0

Based on slides by Peter Balter

0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0
100 100 100 100 200 200 200 100 100 100
Beam 2 100 100 100 100 200 200 200 100 100 100
100 100 100 100 200 200 200 100 100 100
0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0 8
Simple Example of Optimization
If we have a critical structure we want to avoid we can lower the
intensity of one or more of the beamlets that that cross that
structure
Beam 1

0 0 0 100 100 100 0 0 0

0 0 0 0 100 100 100 0 0 0

Based on slides by Peter Balter

0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0
100 100 100 100 200 200 200 100 100 100
Beam 2 100 100 100 100 200 200 200 100 100 100
100 100 100 100 200 200 200 100 100 100
0 0 0 0 100 100 100 0 0 0
0 0 0 0 100 100 100 0 0 0
9
0 0 0 0 100 100 100 0 0 0
Simple Example of Optimization
This results in a decrease in dose to the critical structure but also
to other parts of the dose distribution.

Beam 1

0 0 0 100 50 100 0 0 0

0 0 0 0 100 50 100 0 0 0

Based on slides by Peter Balter

0 0 0 0 100 50 100 0 0 0
0 0 0 0 100 50 100 0 0 0
100 100 100 100 200 150 200 100 100 100
Beam 2 100 100 100 100 200 150 200 100 100 100
50 50 50 50 150 100 150 50 50 50
0 0 0 0 100 50 100 0 0 0
0 0 0 0 100 50 100 0 0 0
10
0 0 0 0 100 50 100 0 0 0
Simple Example of Optimization
This underdose can be made up from other beamlets in other beams restoring
dose to the target but resulting in dose inhomogeneity in the target, the more
beam angles to more opportunity to achieve an optimal plan.

Beam 1

0 0 0 100 50 100 0 0 0

Based on slides by Peter Balter

0 0 0 0 100 50 100 0 0 0
0 0 0 0 100 50 100 0 0 0
0 0 0 0 100 50 100 0 0 0
150 150 150 150 250 200 250 150 150 150
Beam 2 150 150 150 150 250 200 250 150 150 150
50 50 50 50 150 100 150 50 50 50
0 0 0 0 100 50 100 0 0 0
0 0 0 0 100 50 100 0 0 0
11
0 0 0 0 100 50 100 0 0 0
Dose
calculation

Multiple fields: Desired dose:

Di  
j 1 n
CijW j D0  CW0
Simplified: Beamlet weight:
D  CW W0  D0 /C  D0C
12
1
 Can we solve this?
• No
• Huge problem
• Degenerate problem – many solutions
• Ideal dose may not be achievable
• Many unknowns (>1000s beamlet weights)
• Conflicting requirements…..not all of which are clear
• Lots of structures……
• Etc….

13
 What is meant by optimization?
• Not necessarily looking for the true optimum plan
• Many constraints such as deliverability, type of radiation, beam geometry,
planning time….
• Many a priori choices (reduce search space) – constrained optimization
• Beam energy, gantry and collimator angles
A simple objective function:

O  {D0 (1)  Db (1)}2  {D0 (2)  Db (2)}2 ...

0.3
0.25
Objective
function, 0.2
O
0.15
0.1
0.05
0
Iteration step 14
0 10 20 30 40 50 60 70 80 based
Partially 90 100
on slides from Charlie Ma
 15
From Webb, The British Journal of Radiology, 76 (2003), 678–689
 16
Partially based on slides from Charlie Ma
What needs to be in the cost function?
Coverage Good coverage of PTV
Look at 100% and 98% coverage
Hot Spots < 5%
Cord < 46 Gy
Exp Cord 50Gy isodose line shouldn’t cross
Parotid Mean dose ~ 26Gy
Uninvolved < 60 Gy
Larynx / post (attempt to approach 50Gy)
cricoid
Oral cavity No hot spots outside volumes (>60 Gy)
and not hot spots in the mandible
17
Defining the prescription
(and cost functions)

18
The prescription
• The prescription defines the goals of the treatment.
• Target DVH
• Sensitive structure DVH
• Set goals, priorities, penalties

• The plan quality can be scored using either physical or biological criteria.

• It is difficult to reduce all of our treatment planning goals into a set of

equations or a single scoring function

• Warning: no consistency expected in terminology used by different

vendors!

19
 Types of Cost Functions
target organ at risk

2
wl (D-Pl ) (D-Dc)2

2
wu(D-Pu )

Pl Pu Dc
Lower constraint Upper constraints
20
Based on a slide from Yakov Pipman
 Constraint (Pinnacle)
Objective (Pinnacle)
21
The Cost Function
• Cost functions are built based on objectives, there
are a number of objective types possible.
• Minimum Dose
• Maximum Dose
• DVH constraint no more than “x” % of the
structure can exceed a dose of “y”.
• Equivalent Uniform Dose
• …
• Each objective can have a weighting factor
• If the weighing Factor is very high (infinite) that
objective becomes a “Constraint” (in Pinnacle, at
least)

22
 Minimum/Maximum Dose
Advantages
• Constraints can be used guarantee adequate
dose uniformity in the tumor.
• Useful for serial structures such as the spinal
cord.
Disadvantages
• Allowing small hot and/or cold spots are
often provide a significant improvement in
dose conformity.
• One point can dominate the optimization.
• If target and RAR are in close proximity,
these constraints often cannot be satisfied. 23
 Mean Dose
Advantages

• Easy to formulate.

Disadvantages

• Of limited value for most sensitive

structures.
• Dramatically different dose distributions can
have the same mean dose.

24
 Setting constraints
Eclipse screen shot

25
Biological Objective
Functions and
Constraints

26
 Biological Objectives/Constraints

• Biological objective functions and constraints

are outcome related.
• Biological models are used to predict
treatment outcome.
• Tumor Control Probability (TCP).
• Normal Tissue Complication Probability
(NTCP).
• Uncomplicated TCP (UTCP or P+).
• Equivalent Uniform Dose (EUD).

27
 Equivalent Uniform Dose (EUD)

• Two dose distributions are equivalent if the corresponding

biological/clinical outcomes are equivalent
• Normal structures and targets.

1
 a
EUD   vi Di 
a

i1 

*Niemierko A. Med Phys, 26(6), 1999. 28

 Equivalent Uniform Dose (EUD)
Structure (Source) End-point a
Chordoma base of skull (MGH) Local control -13
Squamous cc (Brenner) Local control -13
Melanoma (Brenner) Local control -10
Breast (Brenner) Local control -7.2
Parotids (Eisbruch) Salivary function (<25%) <0.5
Parotids (Chao) Salivary function (<25%) 0.5
Liver (Lawrence) Liver failure 0.6
Liver (Dawson) Liver failure 0.9
Lung (Kwa) Pneumonitis 1.0
Lung (Emami) Pneumonitis 1.2
Kidney (Emami) Nephritis 1.3
Liver (Emami) Liver failure 2.9
Heart (Emami) Pericarditis 3.1
Bladder (Emami) Symptomatic contracture 3.8
Brain (Emami) Necrosis 4.6
Colon (Emami) Obstruction/perforation 6.3
Spinal cord (Powers) White matter necrosis 13
Esophagus (Emami) Perforation 18
Spinal cord (Schultheiss) Paralysis 20

29
Example values – no guarantees!
 Biological
Objectives/Constraints

Advantages
• Our goal is to improve
patient outcome, and this
is precisely what is
modeled with these
techniques.

Disadvantages
• Because of uncertainties in
the parameters included in
the models, the accuracy of
the models is often called
into question.

30
Based on a slide from David Shephard
 Plan Optimization
Fixed Field IMRT

• Beamlet based optimization

• Direct aperture optimization

31
 The Beamlet Model
Before an IMRT optimization, each beam is divided into
a number of smaller beamlets (pencil beams), and the
corresponding dose distributions are computed.

32
Slide from David Shephard
Beamlet-Based Inverse Planning
Beamlet weights are optimized to produce an
optimized fluence map for each beam direction.

1 1 2 2 1 1
1 1 2 2 2 1 1
1 1 1 2 2 1 1
1 2 1 2 3 2 1
2 3 3 2 1 2 1
2 3 3 1 1 1
1 1 1 1 1 1
1 1 1 1 2 2
1 1 1 1 2 2

33
Eclipse’s IMRT dashboard
34
Leaf sequencing

35
 Intensity Modulation
• Step and shoot MLC
• The intensity pattern developed by
the TPS is converted into a finite
number of segments
• For each segment the MLCs leaves
are set and the beam is on for a
determined amount of time
• The summation of all the
segments is equal to the planned
intensity
• Pinnacle

36
Slide from Peter Balter
 Intensity Modulation
• Sliding Window MLC
• MLC leaves move continuously
while the treatment machine is
on
• The field is divided into a
number of control points that
have target positions for each
leaf at each fraction amount of
dose delivered
• The linac modulates leaf speed,
then dose rate to ensure the
targets for each control point
are within tolerance values.

37
Slide from Peter Balter
How Can We Make Any Intensity
Shape with an MLC?

Dose
10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm

39
Slide from Chen Chui
Dose
10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm

40
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

41
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

42
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

43
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

44
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

45
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

46
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

47
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

48
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

49
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

50
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

51
 10

Position
-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

52
 10

-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm
Leaf A Leaf B

53
 Done!

10

-5cm -4 -3 -2 -1 0 +1 +2 +3 +4 +5cm

54
From Optimized Intensity Map to Treatment
Leaf Sequencing
• The optimized treatment plan is not immediately ready
for delivery.
• A leaf sequencing algorithm needs to be applied to
translate the each optimized (theoretical) fluence map
into a set of deliverable aperture shapes.
• The constraints imposed by the multileaf collimator are
accounted for in the leaf sequencing step.
• Final plan dose distribution changes
• This is the approach taken by Eclipse for dynamic IMRT.
• It was the approach used by Pinnacle for step-and-shoot
IMRT (older versions)

55
Based on a slide from David Shephard
Direct aperture
optimization
(DAO)

56
Direct Aperture Optimization (DAO)

1. Inverse planning technique where the aperture

shapes and weights are optimized
simultaneously.
2. All of the MLC delivery constraints are included
in the optimization
3. The number of aperture per beam angle is
specified in the prescription.

57
58
 Simulated Annealing
• DAO uses simulated annealing, an optimization technique using
random sampling techniques.
• The term simulated annealing derives from the roughly analogous
physical process of heating and then slowly cooling a substance to
obtain a strong crystalline structure.
• In each simulation, a minima of the cost function corresponds to this
ground state of the substance.
• The basic principle is that by allowing occasional ascent in the search
process, we might be able to escape the trap of local minima.

Figure from Webb – the first

person to introduce SA to
radiotherapy in the late5980’s
DAO Optimization via Simulated Annealing

1) Pick a parameter (leaf position, aperture weight)

randomly
2) Change the parameter by a random amount
3) Calculate objective function based on the new
dose distribution
4) Objective function lower: accept change
5) Objective function higher: accept change with
certain probability

60
Prescription: 3 apertures per angle
Begin with 3 identical copies

61
Pick an Parameter and Make a Change

Aperture 1
Leaf pair 6
Left leaf position
Move leaf in 2cm

62
Keep or Reject the Change
Based on:

1. MLC constraints.
2. Cost function & Annealing Rules.

63
 MLC Constraints
Some sample Elekta constraints:
1) Opposed leaves 2) Opposed-adjacent
cannot come closer leaves cannot come
than 1-cm from one- closer than 1-cm from
another one-another
< 1cm < 1cm

Not allowed Not allowed

64
After numerous iterations...

Add them up along with their weights…

65
Final intensity map from DAO

66
Small number of apertures can produce large
number of intensity levels

Example: 3 apertures/angle

3 separate
weights

1 2 3 4 5 6 7

67
Small number of apertures can produce large
number of intensity levels

Nn  2 1 n

N = Number of intensity levels

n = Number of apertures

For 3 apertures, 7 intensities

For 4 apertures, 15 intensities
For 5 apertures, 31 intensities
For 6 apertures, 63 intensities

68
Volume-Modulated Arc
Therapy
VMAT

69
70
 Eclipse VMAT

• In Otto’s paper, he used DAO to

produced IMAT plans.
• Two key innovations:
1. Focused on a single arc approach with more
control points in the single arc. Termed “VMAT”.
2. Progressive sampling was used to improve the
speed of the algorithm.
• This is the approach utilized in Eclipse

71
 Dynamic Source Model

Sample
Spacing
Sampling Flexibility Accuracy

Coarse  X
Gantry
Arc

72
Courtesy of Karl Otto
 Dynamic Source Model

Sample
Spacing
Sampling Flexibility Accuracy

Coarse  X
Gantry
Arc
Fine X 

73
Courtesy of Karl Otto
 Progressive Sampling
13 4
Sample 7
12
Spacing
Sampling Flexibility Accuracy 3
11
Coarse  X 6
Gantry
10 Arc
Fine X  2

 
9
Progressive
5
8
1

74
Courtesy of Karl Otto
75
 Progressive Sampling
Maximum MLC Leaf Sample Spacing (cm)
0.5 1 2 3 4 5 6 8 10
10000

Fixed Sampling

1000
Final Cost Value

100

10

Progressive Sampling
1

1 2 4 6 8 10 12 16 20
Beam Sample Spacing (deg)

76
Courtesy of Karl Otto
 Varian Eclipse

• Planning is performed using Direct Aperture Optimization.

• Typical plan uses 1 arc with 177 control points.
• For some cases, multiple arcs are use to improve the plan
quality or provide adequate coverage of large targets.

77
 SmartArc Optimization (Philips)
1. Beams are generated at the start and the stop angles and at 24
increments from the start angle.
2. A fluence map optimization is performed.
3. The fluence maps are sequenced and filtered so that there are
only 2 control points per initial beam angle.
4. These control points are distributed to adjacent gantry angles
and additional control points are added to achieve the desired
final gantry spacing.
5. All control points are processed to comply with the motion
constraints of VMAT.
6. The DMPO algorithm is applied with an aperture based
optimization that takes into account all of the VMAT delivery
constraints.
7. The jaws are conformed to the segments based on the
characteristics of the linac.

78
Courtesy of Philips Medical
Treatment planning is an art

79
Figure from Hunt et al, IJROBP 54(3), 953-962, 2002
Multi-criteria
optimization (MCO)

80
 IMRT planning process is complex
• Long planning time
• Not clear which knobs to turn
• Tradeoffs unclear
Time for IMRT planning for a
• Clinician’s judgment indirect (the complex case (excluding
process does not encourage physician contouring)
participation

Normal tissue sparing Target coverage

Efficiency

N=167, ASTRO 2004

81
Based on slides by Thomas Bortfeld
Pareto surface (or the Possibility Frontier)

Utility curves =
equivalent plans
(determined by the MD
– these are not well
determined)

Pareto surface

82
Craft et al, IJROBP 82, e83-e90, 2012
• PC1: Liver and stomach vs. left and right kidneys
• PC2: Right kidney and stomach vs. left kidney and liver
83
Spalke et al, PMB 54, 3741-3754, 2009
MCO PLANNING (PARETO OPTIMIZATION) - RAYSEARCH
Vilfredo Pareto, born 1848
Pareto-optimality, “efficient”: (Paris) – died 1923 (Geneva)
Industrialist, Sociologist,
“You cannot make anybody better off Economist, Philosopher
without making someone else worse off” Taught in Lausanne, lived in
Céligny near Geneva

RaySearch
MCO PLANNING (PARETO OPTIMIZATION) - RAYSEARCH

Pareto-optimality, “efficient”:
“You cannot make anybody better off
without making someone else worse off”

RaySearch
Increased physician involvement
Reduced planning time

86
Craft et al, IJROBP 82, e83-e90, 2012
Technique comparison

87
Technique comparison: MU/cGy
Pinnacle, DAO Eclipse

• Dynamic IMRT is less MU-efficient than step-and-shoot or VMAT

88
McCarroll et al, Journal of Global Oncology 2017
Technique comparison: Treatment time

Pinnacle, DAO
Eclipse

89
McCarroll et al, Journal of Global Oncology 2017
End on a happy thought:

• The combination of
• IMRT
• IGRT
• 4DCT IMRT/4DCT
• Increased
• Local control
• Overall Survival 3D CRT/CT

• Decreased
• Pneumonitis

90
 The Radiotherapy Process - IMRT
Treatment Delivery
Patient selection Inverse optimization
plan (dMLC, S&S, etc)

Dose distribution
Imaging studies Prescription goals calculation
Treatment Delivery

Verification of Patient
Immobilization Planning Treatment Plan evaluation and
Position and Beam
devices and at-risk Volumes approval
Placement

Target definition
Treatment parameter
(anatomy, physiology Organs at risk Plan test and
transfer to R&V and to
and the natural delineation verification
treatment unit control
history of the disease)

91
Slide from Yakov Pipman