Real-time Re-planning for Online
Adaptive Radiotherapy
Steve Jiang, Xuejun Gu , Chunhua Men , Xun
Jia, Oliver Fluck, DJ Choi, Amit Majumdar
Conventional Radiotherapy
 Treatment simulation
 Build a virtual patient model
 Treatment planning
 Perform virtual treatment using virtual machine on virtual patient
 Treatment delivery
 Same treatment is repeated for many fractions
Repeat
 Basic assumption: human body is a static system
Days
Simulation
Days
Planning
2
Treatment
Human Body Is A Dynamic System
Week 1
Week 3
Tumor
Van de Bunt et al. ‘06
 Tumor volume shrinkage in response to the treatment
 Tumor shape deformation due to filling state change of neighboring organs
 Relative position change between tumor and normal organs
3
Solution
 Develop a new treatment plan that is optimal to
patient’s new geometry
 Adaptive radiation therapy (ART)
4
Offline ART
Yes
Re-Sim?
No
Days
Simulation
Repeat
Days
Planning
Treatment
 Significant tumor response and/or significant weight loss
→ Plan modification in the middle of treatment
 Low efficiency → Limited to 1 modification per patient
and very few patients
 Only works for gradual change of patient anatomy
5
Online ART
Repeat
Days
Days
Simulation
Planning
5-8 min
On-board Imaging
Re-planning
Treatment
 On-board volumetric imaging has recently become available
 Major technical obstacle for clinical realization of online ART
 Real-time re-planning
 Imaging dose
 Clinical workflow
6
Online Re-planning Process
CBCT
Reconstruction
Planning CT
w/ Contours
Deformable
Image Regis
Deformed
pCT and
Contours
Treatment Planning
System
Dose
Calculation
Beam Setup
Dose
Deposition
Coefficients
Dose
Distribution
Plan
Re-optimization
New Plan
7
Initial Plan
Development of GPU-based Realtime Deformable Image Registration
Gu et al Phys Med Biol 55(1): 207-219, 2010
8
Deformable Image Registration
 Morphing one image
into another with correct
correspondence
9
Deformable Image Registration with ‘Demons’
CPU  GPU
Start
Moving
Image Im(r)
Static
Image
I m (r n )
I s (r n )
No
Active
Force?
No
Passive
Force?
Yes
Yes
n
Gradient ∇I m (r )
n
Gradient ∇I s (r )
No
Stopping
Criteria
Moving vector dr n
Yes
GPU  CPU
End
10
Updating r n +1 = r n + dr n
n +1
n +1
Compare I s (r ), I m (r )
10
Gu et al Phys Med Biol
55(1): 207-219, 2010
Results for GPU-based Demons Algorithms
Method
Case 1
Case 2
Case 3
Case 4
Case 5
Average
PF
1.11/6.80 1.04 /7.18
1.36/7.39
2.51/6.49
1.84/7.24
1.57/7.02
ePF
1.10/6.82 1.00/7.20
1.32/7.42
2.42/6.56
1.82/7.08
1.53/7.02
AF
1.15/8.29 1.05/9.24
1.39/8.79
2.34/7.75
1.81/8.44
1.55/8.50
DF
1.19/7.71 1.16/8.65
1.48/8.02
2.59/8.30
1.91/8.44
1.66/8.22
aDF
1.11/8.36 1.02/8.69
1.35/8.97
2.27/7.77
1.80/8.70
1.51/8.50
IC
1.24/11.07 1.28/11.47 1.42/11.54 3.27/10.46 1.67/10.98 1.78/11.10
3D spatial error (mm) / GPU time (s), image size 256×256×100
~100x speedup compared to an Intel Xeon 2.27 GHz CPU
11
Development of GPU-based Realtime Dose Calculation
Gu et al Phys Med Biol 54(20) 6287-97, 2009
Jia et al Phys Med Biol 55(11): 3077–3086, 2010
12
Finite-size Pencil Beam (FSPB) Model
  a

 a

 f + x' 
 f − x' 

A
d
(
)
Ei
2

 + erf  2
DFSPB
( x, d , z ) = ∑ i
erf 
4
(
)
(
)
d
d
2
σ
2
σ




i =1
i
i
 






 
  b

 b

 f + z ' 
  f 2 − z' 

 + erf  2
erf 
 2σ i (d ) 
  2σ i (d ) 


 



3
13
Results for GPU-based FSPB Algorithm
Voxel size (cm3)
0.50x0.50x0.50
Beamlet size # Voxels
#
CPU Time GPU Time
Speedup
(cm2)
(× 106 ) Beamlets
(sec)
(sec)
0.22
2500
21.22
0.06
0.20x0.20
373
0.37x0.37x0.37
0.20x0.20
0.51
2500
42.80
0.10
409
0.30x0..30x0.30
0.20x0.20
1.00
2500
78.27
0.18
419
0.25x0.25x0.25
0.20x0.20
1.73
2500
124.54
0.30
421
0.25x0.25x0.25
0.25x0.25
1.73
1600
120.14
0.29
415
0.25x0.25x0.25
0.33x0.33
1.73
900
112.78
0.27
416
0.25x0.25x0.25
0.50x0.50
1.73
400
100.77
0.24
417
~400x speedup compared to an Intel Xeon 2.27 GHz CPU
< 1 sec for a 9-field prostate IMRT plan
14
FSPB with
3D
Density
Correction
Jelen and
Alber, Phys
Med Biol,
52(3) : 617633, 2007
MC
FSPB
15
Monte Carlo Dose Calculation on GPU
Start
Transfer data to GPU including random # seeds, cross sections, and pre-generated e- tracks etc.
a). Clean local counter
b). Simulate one MC history on
thread #1
c). Put dose to global counter
a). Clean local counter
b). Simulate one MC history on
thread #1
c). Put dose to global counter
……
a). Clean local counter
b). Simulate one MC history on
thread #1
c). Put dose to global counter
Yes
No
Reach a preset # of
histories ?
Transfer data from GPU to CPU
End
 Directly map DPM code on GPU
 Treat a GPU card as a CPU cluster
16
Results for GPU-based MC Dose Calculation
Case
#
Source
type
# of
Histories
Stan Dev
CPU
(%)
Stan Dev
GPU
(%)
TCPU
(min)
TGPU
(min)
TCPU/TGPU
1
Electron
107
0.66
0.65
8.3
1.8
4.5
2
Photon
109
0.41
0.41
94
17
5.5
~5x speedup compared to an Intel Xeon 2.27 GHz CPU
< 3 min for 1% sigma for photon beams
17
Development of GPU-based Realtime Plan Re-optimization
Men et al Phys Med Biol 54(21):6565-6573, 2009
Men et al Phys Med Biol 2010 (in print)
Men et al Med Phys 2010 (under review)
18
Plan Re-optimization: A FMO Model
min ∑ F (z l )
j j
j ∈V
Subject to
z l = ∑ Dl x
j
ij i
i∈N
x ≥0
i
j ∈V
i∈N
Where
Fs − (z ) =
∑ (max(0, z
j
− z lj )) 2
s ∈T
Fs + (z ) =
∑ (max(0, z
l
j
− z j )) 2
s∈S
j∈Vs
j∈Vs
19
Results for GPU-based FMO Algorithm
100
Rectum
PTV
80
Bladder
Volume [%]
60
Femoral Head
40
20
Body
0
0
20
40
Dose [Gy]
60
80
# beamlet
#
voxel
Case
size
beamlets
size
(mm2)
(mm3)
1
2,055
10×10
4×4×4
2
6,433
5×5
4×4×4
3
6,433
5×5
2.5×2.5×2.5
#
voxels
(×104)
3.6
3.6
14.0
#
non-zero Dij's
(×106)
3.1
10.6
43.3
~40x speedup compared to an Intel Xeon 2.27 GHz CPU
~0.5 sec for re-optimizing a 9-field prostate IMRT plan
Men et al Phys Med Biol 54(21):6565-6573, 2009
20
GPU
time
(s)
0.2
0.5
2.8
GPU-based DAO Algorithm
Start
Transfer data from CPU to GPU
Solve the sub-problem
Add one aperture to the master problem
Solve the master problem
No
Satisfy stop
criterion?
Yes
Transfer data from GPU to CPU
End
Men et al Phys Med Biol 2010 (in print)
21
Results for GPU-based DAO Algorithm
C a se # be am le ts
7,196
P1
P2
7,137
P3
5,796
P4
7,422
P5
8,640
H1
5,816
H2
8,645
H3
9,034
H4
6,292
H5
5,952
# v oxe ls # no n-ze ro D ij ’s R un nin g time (sec )
2 ,763, 243
45,9 12
1.7
48,6 42
2 ,280, 076
0.7
28,9 31
1 ,765, 294
0.8
39,8 22
2 ,717, 424
2.3
49,2 10
3 ,086, 884
1.6
33,2 52
1 ,576, 418
1.0
59,6 15
3 ,162, 752
2.4
74,4 38
3 ,500, 188
1.8
31,5 63
1 ,596, 168
1.8
42,3 30
2 ,215, 202
2.5
22
Results for GPU-based DAO Algorithm
23
VMAT Plan Optimization: Difficulties
 Large optimization problem
 Highly constrained
 Direct aperture optimization
24
VMAT Plan Optimization: Our Solution
 Solve it as a large-scale convex programming
problem using a column generation approach
 MLC apertures are generated one by one by
solving a master problem and a subproblem
iteratively
 At each iteration the subproblem generates
the most promising and deliverable aperture
 And the master problem minimizes a cost
function (dose and dose rate)
25
VMAT Plan Optimization: Our Results
Case
P1
P2
P3
P4
P5
H1
H2
H3
H4
H5
# non-zero Dij’s CPU time GPU time
# beamlets # voxels
(×107)
(sec)
(sec)
40,620
45,912
2.3
340
22
59,400
48,642
3.2
265
18
38,880
28,931
1.8
276
20
43,360
39,822
2.6
410
26
51,840
49,210
3.0
348
23
51,709
33,252
2.5
290
21
78,874
59,615
5.0
468
27
90,978
74,438
5.5
342
25
71,280
31,563
2.6
363
25
53,776
42,330
3.5
512
31
26
VMAT Plan Optimization: A Prostate Case
IMRT
VMAT
27
VMAT Plan Optimization: A H/N Case
IMRT
VMAT
28
Summary
 We have developed GPU-based computational
tools for real-time treatment re-planning
 For a typical 9-field prostate case
 The deformable registration can be done in 7 seconds
 The dose calculation takes less than 2 seconds
 The plan re-optimization takes less than 1 second (FMO), 2
seconds (DAP), or 30 seconds (VMAT)
 A new plan can be developed in about 10-40 seconds
 Next step
 Deformable registration issues
 Clinical tests (codes in public domain, research platform)
29
Acknowledgement
http://radonc.ucsd.edu/Research/CART
30
Consequence of Patient Anatomical Variation
 An optimal treatment plan may become less optimal
or not optimal at all
 Dose to tumor ↓
 Dose to normal tissues ↑
 Dose to tumor ↓ → Tumor control ↓
 Dose to normal tissues ↑ → Toxicity ↑
 Toxicity ↑ → Prescribed tumor dose ↓ → Tumor
control ↓
31
Initial Plan Guided Re-optimization (IPGRO)
 The initial plan has satisfied physician's clinical
requirements by incorporating dose-volume or biological
constraints into the optimization model
 We try to minimize the necessity for the physician's reapproval of the re-optimized plan by complying with the
physician's preference (e.g., the locations of hot/cold
spots).
 While keeping the dose to each voxel inside critical
structures lower than that in the initial plan, we tried to
constrain the dose to each voxel inside the target in
between the initial plan dose and the prescription dose,
with a mild force to pull it towards the prescription dose
32
Prescription Dose Guided Re-optimization
(PDGRO)
 It is based on the cumulative dose distribution
(CDD) and dose volume histogram (DVH)
constraints
 This model, compared to the IPGRO model, has
larger solution space and thus may lead to a
new plan of higher quality
 However, the new plan might be very different,
e.g., in terms of the locations of hot/cold spots,
from the initial plan and the physician's reapproval may be needed
33