Upload
truongquynh
View
215
Download
0
Embed Size (px)
Citation preview
7/29/2017
1
Making Better IMRT Plans Using a New Direct Aperture Optimization Approach
Dan Nguyen, Ph.D.
Division of Medical Physics and Engineering
Department of Radiation Oncology
UT Southwestern
AAPM Annual Meeting
Aim of Radiotherapy Research
• Overall goal of radiotherapy research is to find practical and efficient ways to maximize radiation dose to the tumor while minimizing dose to normal tissues and organs.
2
Aim of Radiotherapy Research
• Overall goal of radiotherapy research is to find practical and efficient ways to maximize radiation dose to the tumor while minimizing dose to normal tissues and organs.
3
Planning Target Volume(PTV)
7/29/2017
2
Aim of Radiotherapy Research
• Overall goal of radiotherapy research is to find practical and efficient ways to maximize radiation dose to the tumor while minimizing dose to normal tissues and organs.
4
Organ at Risk (OAR)
Intensity Modulated Radiation Therapy
• Intensity Modulated Radiation Therapy (IMRT) is a widely accepted and effective technique to radiotherapy.
• Inverse planning approach
• Planner defines prescription dose to the PTV and structure importance
• Fluence map optimization finds beamlet intensities that best satisfies planner criteria
5
12
OAR
OAR
OARPTV
Fluence mapFluence map optimization
(FMO)
Problem of conventional IMRT
• Conventional FMO does not take into account hardware delivery constraints
• Once the fluences are optimized, they must be converted into an approximation deliverable by multileaf collimators (MLC).
• MLC Segmentation
6
Mulitleaf Collimator (MLC)OptimizedFluence Map
MLC DeliverableFluence Map
Image courtesy of Varian Medical Systems, Inc. All rights reserved
Stratification MLC
Sequencing
Deliver!
7/29/2017
3
MLC Segmentation
• Aim of segmentation is to limit the number of deliverable apertures that the MLCs can form.
• Why? Delivering large number of small fields leads to complications with small field dosimetry error buildup and low output (longer treatment time).
• This segmentation step, particularly the stratification process, degrades the fluence map and the resulting dose distribution.
7
Stratification MLC
Sequencing
Stratification MLC
Sequencing
Direct Aperture Optimization
• Direct Aperture Optimization (DAO) attempts to solve the problem by reformulating the optimization to account for machine constraints.
• Creates fluence maps that are directly deliverable.
• Removes need for MLC segmentation step.
• Commercial DAO uses a simulated annealing algorithm
• Stochastic algorithm
8
Simulated annealing DAO
9
The routine randomly selects either to change the intensity or a leaf position for modification.
A random number based on a Gaussian determines the size and direction of the change. The standard deviation of the Gaussian is
1 11
1 /
Changes that lower the cost objective are always accepted. Otherwise, the change is accepted with a probability
1
1 /
Shepard DM, Earl MA, Li XA, Naqvi S, Yu C. Direct aperture optimization: A turnkey solution for step-and-shoot IMRT. Medical Physics. 2002;29:1007-18
7/29/2017
4
Simulated Annealing DAO for VMAT
• From the seminal paper by Otto1, VMAT creates an arc by progressively sampling new beams (figure below), and using simulated annealing DAO to
• Solve for exactly one deliverable aperture for each beam
10
1Otto, Karl. "Volumetric modulated arc therapy: IMRT in a single gantry arc." Medical physics 35.1 (2008): 310-317.
[1]
Simulated annealing DAO
• Changes 1 MLC leaf or aperture intensity at a time
• Worked well for VMAT because of progressive sampling of beams
• Added beams would adopt an aperture shape and intensity value from their neighbor
• In general has difficulty scaling to large problems with lots of variables to optimize
• DAO for static beam IMRT
• Multiple apertures per beam
• No progessive sampling
• Stochastic algorithm.
• Probabilistic nature of algorithm cannot guarantee reproducibility of results.
11
Rethinking DAO Problem
• Need to come up with a way to find a fluence map that can be MLC sequenced without any modification to the fluence map.
• We want to optimize a fluence map that has
• Piecewise-constant regions.
• Limited number of discrete intensity levels
12
Can be delivered by MLCs without further modification
7/29/2017
5
Rethinking DAO Problem
• Other goals
• Fast to solve
• Piecewise-constant regions can take any shape (no library of aperture to choose from)
• Large regions
13
Image Segmentation
• In the domain of mathematics imaging and vision, effective image segmentation formulations and algorithms have been developed.
• Piecewise-constant Mumford-Shah formulation solved with Primal Dual Hybrid Gradient Algorithm1.
14
1Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[1]
Image Segmentation
15
Regions can be arbitrarily shaped, and does not require a finite sized library of shapes for optimizer to search from
1Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[1]
7/29/2017
6
Multiphase piecewise-constant Mumford-Shah function
16
Mumford-Shah
• The original Mumford-Shah functional
• Proposed by Mumford and Shah in 19891
• Segments an image into piecewise smooth sub-regions
• Piecewise constant version proposed in 20012
• Also known as Chan-Vese model
• Used level-set functions to solve for piecewise constant regions
17
1Mumford, David, and Jayant Shah. "Optimal approximations by piecewise smooth functions and associated variational problems." Communications on pure and applied mathematics 42.5 (1989): 577-685.2Chan, Tony F., and Luminita A. Vese. "Active contours without edges."Image processing, IEEE transactions on 10.2 (2001): 266-277.
Mumford-Shah
• Chan et al. proposed a convex relaxation method in 20061.
• Used a labeling array instead of level set functions.
• Easy scaling for any number of segments.
• Allowed for the problem to be efficiently evaluated with a proximal algorithm called primal dual hybrid gradient (PDHG)2 = fast.
18
1Chan, Tony F., Selim Esedoglu, and Mila Nikolova. "Algorithms for finding global minimizers of image segmentation and denoising models." SIAM journal on applied mathematics 66.5 (2006): 1632-1648.2Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[2]
7/29/2017
7
Mumford-Shah
• Multiphase Piecewise-Constant Formulation
• x is the original image
• u defines the shape of the segments
• c assigns a value to each segment
19
, 12
u 0,
1
1Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[1]
[1]
Mumford-Shah
• Multiphase Piecewise-Constant Formulation
• x is the original image
• u defines the shape of the segments
• c assigns a value to each segment
20
, 12
u 0,
1
1Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[1]
[1]
Mumford-Shah
• Multiphase Piecewise-Constant Formulation
• x is the original image
• u defines the shape of the segments
• c assigns a value to each segment
21
, 12
u 0,
1
1Chambolle A, Pock T. A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging. Journal of Mathematical Imaging and Vision. 2011;40:120-45.
[1]
[1]
7/29/2017
8
Mumford-Shah Summary
• The multiphase piecewise constant Mumford-Shah function
• Can use a labeling array for segmentation
• Can be solved quickly with PDHG algorithm
• Solves for pairwise disjoint segments
• Solves all segments simultaneously
22
The FMO formulation
23
FMO formulation
• A common convex FMO formulation incorporates both a dose fidelity term and a total variation (TV) regularization term
• x is the fluence map
• A is the dose calculation matrix
• d is the prescription dose
• W weights the structures of interest
24
7/29/2017
9
FMO formulation
• TV regularization helps piecewise smooth the fluence maps
25
Without TV regularization With TV regularization
DAO: Combining FMO with Mumford-Shah function
26
Unifying FMO and image segmentation into a single formulation
27
FMO with anisotropic TV Multiphase piecewise constant Mumford Shah1
12
, 12
1Chambolle, Antonin, and Thomas Pock. "A first-order primal-dual algorithm for convex problems with applications to imaging." Journal of Mathematical Imaging and Vision 40.1 (2011): 120-145.
7/29/2017
10
Unifying FMO and image segmentation into a single formulation
28
FMO with anisotropic TV Multiphase piecewise constant Mumford Shah1
, 12
1Chambolle, Antonin, and Thomas Pock. "A first-order primal-dual algorithm for convex problems with applications to imaging." Journal of Mathematical Imaging and Vision 40.1 (2011): 120-145.
12
Unifying FMO and image segmentation into a single formulation
29
FMO with anisotropic TV
0, 0, 1 ,
Multiphase piecewise constant Mumford Shah1
1Chambolle, Antonin, and Thomas Pock. "A first-order primal-dual algorithm for convex problems with applications to imaging." Journal of Mathematical Imaging and Vision 40.1 (2011): 120-145.
, 12
12
, , 12
12
0, 0, 1 ,
, , 12
12
12
Unifying FMO and image segmentation into a single formulation
30
, 12
FMO with anisotropic TV Multiphase piecewise constant Mumford Shah1
1Chambolle, Antonin, and Thomas Pock. "A first-order primal-dual algorithm for convex problems with applications to imaging." Journal of Mathematical Imaging and Vision 40.1 (2011): 120-145.
7/29/2017
11
0, 0, 1 ,
, , 12
12
12
Unifying FMO and image segmentation into a single formulation
31
, 12
FMO with anisotropic TV Multiphase piecewise constant Mumford Shah1
1Chambolle, Antonin, and Thomas Pock. "A first-order primal-dual algorithm for convex problems with applications to imaging." Journal of Mathematical Imaging and Vision 40.1 (2011): 120-145.
Deterministic Direct Aperture Optimization
32
0,
u 0, 1 , ,
, , 12
12
0,
u 0, 1 , ,
, , 12
12
Deterministic Direct Aperture Optimization
33
Because of the third term, this formulation is not convex.
However, if we update one variable (x,c, or u) while holding the other 2 constant, we obtain a convex module.
7/29/2017
12
Deterministic Direct Aperture Optimization
34
This can be locally solved by updating the variables in a alternating block fashion.
Module 1: Solve for x, while holding c and u constantModule 2: Solve for c, while holding x and u constantModule 3: Solve for u, while holding x and c constant
0,
u 0, 1 , ,
, , 12
12
Deterministic Direct Aperture Optimization
35
All 3 modules can be efficiently solved with PDHG
0,
u 0, 1 , ,
, , 12
12
Evaluation
36
7/29/2017
13
Evaluation
• 3 planning cases were tested
37
Patient
Number of allowed
segments per beam
( )
Prescription dose
(Gy)PTV volume (cc)
Glioblastoma Multiforme
(GBM)10 30 57.77
Lung (LNG) 10 50 47.84
Head & Neck (H&N) 20
54 197.54
59.4 432.56
69.96 254.98
Evaluation
• Shell and skin structures added around PTV•
• Dose was calculated for each patient for 1162 beams
• Convolution/Superposition (CVSP)
• Dose array resolution: 0.25 cm x 0.25 cm x 0.25 cm
• Beamlet resolution: 0.5 cm x 0.5 cm
• 20 beam angles were selected for each patient
• Column generation algorithm
38
Evaluation
• 20 beam plans created using the DAO with Mumford-Shah (DAOMS) and the simulated annealing DAO (DAOSA) methods.
• DAOMS: is initialized to 0
• DAOSA: is initialized to have conformal beams around PTV
• Computer
• Intel Core i7-3960X CPU
• 6 physical cores overclocked to 4.00 GHz
• GeForce GTX 690 GPU
• 32 GB RAM
39
7/29/2017
14
Results
40
Size of problem
Length of Number of non-zeros in dose
array
GBM 2,175 37,411,365
LNG 2,011 27,910,230
H&N 18,934 243,728,264
41
Convergence
42
0 200 400 600 800 1000 1200 1400 1600 1800 200010
7
108
109
1010
Iteration
Obj
ectiv
e va
lue
GBM: Convergence plot for DAOMS
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
107
108
109
1010
GBM: Convergence plot for DAOSA
Iteration
Obj
ectiv
e va
lue
Modules 1&2
Module 3
7/29/2017
15
Total solve time
43
0
5000
10000
15000
20000
25000
30000
35000
40000
GBM LNG H&N
time
(s)
Solve time
DAO_MS DAO_SA
DAOMS is 9.5 to 40 times faster
Lower is better
GBM: Dose Wash
44
GBM: DVH
45
0 5 10 15 20 25 30 350
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
11
Fra
ctio
nal v
olum
e
GBM: DAOMS
(solid) vs DAOSA
(dotted)
Dose (Gy)
PTVBrainBrainstemChiasmSpinalCordR Opt NrvL Opt NrvR EyeL EyeR LensL LensR CochleaL Cochlea
7/29/2017
16
LNG: Dose Wash
46
LNG: DVH
47
0 5 10 15 20 25 30 35 40 45 50 55 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
11
Fra
ctio
nal v
olum
e
LNG: DAOMS
(solid) vs DAOSA
(dotted)
Dose (Gy)
PTVSpinalCordTracheaProximalBronchusHeartEsophagusLung
H&N: Dose Wash
48
7/29/2017
17
H&N: DVH part 1
49
0 10 20 30 40 50 60 70 800
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
11
Fra
ctio
nal
vol
ume
H&N part 1: DAOMS
(solid) vs DAOSA
(dotted)
Dose (Gy)
PTV6996PTV5940PTV5400BrainstemChiasm
H&N: DVH part 2
50
0 10 20 30 40 50 60 70 800
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
11
Fra
ctio
nal
vol
ume
H&N part 2: DAOMS
(solid) vs DAOSA
(dotted)
Dose (Gy)
CordR Opt NrvL Opt NrvR CochleaL CochleaL Parotid
H&N: DVH part 3
51
0 10 20 30 40 50 60 70 800
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
11
Fra
ctio
nal v
olu
me
H&N part 3: DAOMS
(solid) vs DAOSA
(dotted)
Dose (Gy)
MandibleLipsOral CavityLarynxPharynxEsophagus
7/29/2017
18
Statistics
• On average, DAOMS reduced Dmax and Dmean by (% of prescription dose):
• (GBM) 0.01% and 0.001%
• (LNG) 3.67% and 1.08%
• (H&N) 10.91% and 10.81%
• The average dose coverage, D98 and D99, was increased by 1.66% and 2.21% of the prescription dose.
52
Aperture Comparison
53
Segment evaluation
54
02468
101214161820
GBM LNG H&N
Num
ber
of s
egm
ents
Average number of segments per beam
DAO_MS DAO_SAMaximum allowed number of apertures Lower is better
7/29/2017
19
Segment Evaluation
55
0
10
20
30
40
50
60
GBM LNG H&N
Num
ber
of b
eam
lets
Mean number of beamlets in a segment
DAO_MS DAO_SAHigher is better
Discussion
56
Discussion
• DAOMS results in perfectly piecewise constant fluence maps that are equivalent to apertures without additional stratification.
• In terms of computation speed, DAOMS is far superior
• 9.5 to 40 fold increase in speed to converge
57
7/29/2017
20
Discussion
• For simpler cases (GBM and H&N), DAOMS and DAOSA are very comparable in dosimetry.
• Similar OAR sparing (DAOMS does slightly better)
• DAOSA still competitive in PTV coverage and homogeneity.
• For complicated cases (H&N), DAOMS is clearly superior in all aspects
• DAOSA has difficulty reaching a pareto optimum
58
Limitation of DAOMS
• A major limitation of DAOMS is that the piecewise constant segmentation only solves for pairwise disjoint regions
• This property causes DAOMS to have:
• More apertures on average (to increase complexity)
• Smaller apertures (each aperture competes for space)
59
DAOSA: overlappingDAOMS: Pairwise disjoint
Conclusion
60
7/29/2017
21
Conclusion
• Novel DAO formulated by combining FMO with a multiphase piecewise constant Mumford-Shah segmentation.
• Can generate any shaped MLC segment on the fly.
• Dosimetrically competitive to commercial simulated annealing method for simple cases, and superior for complex cases.
61
Thank you!
62
63