optimization of nonuniformly fractionated radiotherapy ... · melissa r. gaddy (ncsu) nonuniform...
TRANSCRIPT
Optimization of Nonuniformly Fractionated RadiotherapyTreatments
Melissa R. Gaddy1, Sercan Yıldız2, Jan Unkelbach3, David Papp1
1. NC State University, 2. UNC Chapel Hill, 3. University Hospital Zurich
May 2, 2017
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 1 / 31
Outline
1 Optimization in Radiotherapy
2 The Fractionation Problem
3 The SDP Relaxation
4 Future Directions
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 2 / 31
Optimization in Radiotherapy
Outline
1 Optimization in Radiotherapy
2 The Fractionation Problem
3 The SDP Relaxation
4 Future Directions
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 3 / 31
Optimization in Radiotherapy
External-Beam Radiotherapy
Main objective: Irradiate the tumor while avoiding the organs andsurrounding tissue
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 4 / 31
Optimization in Radiotherapy
External-Beam Radiotherapy
Main objective: Irradiate the tumor while avoiding the organs andsurrounding tissue
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 5 / 31
Optimization in Radiotherapy
External-Beam Radiotherapy
Linear accelerator Multi-leaf collimator
https://www.varian.com/oncology/products/treatment-delivery/clinac-ix-system, accessed: 2017-11-04.http://newsroom.varian.com/imagegallery?mode=gallery&cat=2473, accessed: 2017-18-04.
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 6 / 31
Optimization in Radiotherapy
Intensity Modulated Radiotherapy (IMRT)
Deliver multiple unmodulated beamsfrom the same angle to superimposetheir effect
Reemsten and Alber. Handbook of Optimization in Medicine, Pardalosand Romeijn (eds.), Springer. 2009.
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 7 / 31
Optimization in Radiotherapy
Intensity Modulated Radiotherapy (IMRT)
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 8 / 31
Optimization in Radiotherapy
Optimization in Radiotherapy
Reemsten and Alber. Handbook of Optimization inMedicine, Pardalos and Romeijn (eds.), Springer.2009.
Discretize patient into a 3-D grid ofvoxels
Decision variables:
x : vector of beamlet weightsd : vector of doses delivered to eachvoxel of the patient
Dose-influence matrix D relatesbeamlet weights to dose absorbed bythe patient Dx = d
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 9 / 31
Optimization in Radiotherapy
Optimization in Radiotherapy
Dose-based IMRT model:
minx ,d
∑i∈I
wiFi (d)
s.t. Dx = d
x ≥ 0
Each Fi is a piecewise quadratic penalty function to penalize underdose,overdose, or mean dose.
E.g. Fi (d) =∑v∈V
(dv − dpresv )2+ to penalize overdose
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 10 / 31
The Fractionation Problem
Outline
1 Optimization in Radiotherapy
2 The Fractionation Problem
3 The SDP Relaxation
4 Future Directions
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 11 / 31
The Fractionation Problem
What is Fractionation?
Total treatment is delivered over a series of days or weeks (eachtreatment day is called a fraction)
Currently in the clinic, the same treatment is delivered each day.
Want to evaluate potential improvement of delivering different doseseach day
Uniform Nonuniform
• Same treatment every day • Different treatment each day• Beamlet weights: x , doses: d • Beamlet weights: xt , doses: dt
(t = 1, . . . ,N)• Convex optimization problem • Nonconvex optimization problem• Based on physical dose d • Based on biologically effective
dose b
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 12 / 31
The Fractionation Problem
Biologically Effective Dose (BED) Model
Incorporate the effect of fractionation with biologically effective dose:
b =N∑t=1
(dt +
d2t
(α/β)
)
dt is dose absorbed in fraction t, (t = 1, . . . ,N)
α and β are tissue-specific parameters
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 13 / 31
The Fractionation Problem
Nonuniform Model
Dose-based IMRT model:
minx ,d
∑i∈I
wiFi (d)
s.t. Dx = d
x ≥ 0
(Nonconvex) Nonuniformlyfractionated IMRT model:
minx ,d ,b
∑i∈I
wiFi (b)
s.t. bv =N∑t=1
(dvt + d2vt
(α/β)v) ∀v
Dxt = dt t = 1, . . . ,N
xt ≥ 0 t = 1, . . . ,N
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 14 / 31
The Fractionation Problem
Evaluating Plan Quality
How can we fairly evaluate the benefit of nonuniform fractionation?
Two challenges:
Cannot fairly compare a dose-based plan with a BED-based plan
Compute a BED-based uniform reference plan
Cannot interpret the objective function
Prioritize a single clinical objective
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 15 / 31
The Fractionation Problem
Uniform Model
Assuming x1 = · · · = xN , we rewrite the model to solve for a uniformreference plan.
minx ,d ,b
∑i∈I
wiFi (b)
s.t. bv = Ndv
(1 +
dv(α/β)v
)∀ voxels v
Dx = d
x ≥ 0.
Despite having the same quadratic equality constraints, this is convex.
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 16 / 31
The Fractionation Problem
Constrained Nonuniform Model
Prioritize the clinical objective F1
Let b∗ be the BED obtained in the uniform reference plan.
minx ,d ,b
F1(b)
s.t. Fi (b) ≤ Fi (b∗) i 6= 1
bv =N∑t=1
(dvt + d2vt
(α/β)v) ∀ voxels v
Dxt = dt t = 1, . . . ,N
xt ≥ 0 t = 1, . . . ,N
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 17 / 31
The Fractionation Problem
Experimental Setup
Five liver cases with various geometries (2-D slice)
Five-fraction treatments with 21 equispaced beams
Clinical objectives:
Minimize the mean BED in the healthy liver tissueLower and upper BED bounds on all voxels
Computed globally optimal uniform reference plan
Computed locally optimal nonuniform plan
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 18 / 31
The Fractionation Problem
Locally Optimal Solution of Nonuniform FractionatedModel
High single-fraction dose to subregions of the tumor
Consistent, low dose to healthy tissue
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 19 / 31
The Fractionation Problem
Locally Optimal Solutions of Nonuniform Model
These are qualitatively about the same, but each partitions the tumordifferently
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 20 / 31
The Fractionation Problem
Results
The nonuniformly fractionated plans achieve a 13% - 35% reduction inmean liver BED
Case Description Mean liver BEDin uniformreference plan
Mean liver BEDin nonuniformplan
Mean liver BEDreduction
1 Large central lesion 84.54 75.87 12.75%2 Small lesion 26.14 19.47 34.26%3 Two small lesions 59.54 50.24 18.51%4 Lesion abutting chest wall 47.51 38.65 22.92%5 Lesion abutting GI tract 88.67 77.38 14.59%
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 21 / 31
The Fractionation Problem
Comparing Uniform and Nonuniform Plans
Uniform reference plan Nonuniform DEQ5
Difference
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 22 / 31
The SDP Relaxation
Outline
1 Optimization in Radiotherapy
2 The Fractionation Problem
3 The SDP Relaxation
4 Future Directions
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 23 / 31
The SDP Relaxation
Bound on the Maximum Achievable Benefit
Recall the constrained nonuniform fractionation model:
minx ,d ,b
F1(b)
s.t. Fi (b) ≤ Fi (b∗) i 6= 1
bv =N∑t=1
(dvt + d2vt
(α/β)v) ∀ voxels v
Dxt = dt t = 1, . . . ,N
xt ≥ 0 t = 1, . . . ,N
where Fi is of the form∑v∈V
(bpresv − bv )2+ or similar
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 24 / 31
The SDP Relaxation
Bound on the Maximum Achievable Benefit
Reformulate the model into a QCQP using auxiliary variables
Introduce Xt = xtxTt and write the quadratic inequalities as linear
inequalities in xt and Xt .E.g. Incorporating Dxt = dt , the first constraint
bv =N∑t=1
(dvt + d2vt
(α/β)v)
becomes
bv =N∑t=1
⟨[1 xTtxt Xt
],
[0 eTv D
2DT ev
21
(α/β)vDT eve
Tv D
]⟩
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 25 / 31
The SDP Relaxation
Bound on the Maximum Achievable Benefit
Replace Xt = xtxTt with the convex relaxation
[1 xTtxt Xt
]< 0.
Add the componentwise inequality Xt ≥ 0 to tighten the bound.
Since the formulation is convex and symmetric in the fractions, wecan assume x1 = · · · = xN and X1 = · · · = XN .
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 26 / 31
The SDP Relaxation
SDP Relaxation for Constrained Nonuniform Modelmin
x,X ,p,q,rr1
s.t.∑v∈Vi
p2iv ≤ Fi (b∗) ∀i ∈ I+
∑v∈Vi
q2iv ≤ Fi (b∗) ∀i ∈ I−
ri ≤ Fi (b∗) ∀i ∈ Im, i 6= 1
piv ≥ N
⟨[1 xT
x X
],
[− bhiiv
N
eTv D
2DT ev
2Cv
]⟩∀v ∈ Vi , ∀i ∈ I+
qiv ≥ −N⟨[
1 xT
x X
],
[− bloiv
N
eTv D
2DT ev
2Cv
]⟩∀v ∈ Vi , ∀i ∈ I−
ri ≥N
|Vi |∑v∈Vi
⟨[1 xT
x X
],
[−mhi
iN
eTv D
2DT ev
2Cv
]⟩∀i ∈ Im
piv ≥ 0, qiv ≥ 0, ri ≥ 0 ∀i , v[1 xT
x X
]< 0, x ≥ 0, X ≥ 0.
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 27 / 31
The SDP Relaxation
SDP Lower Bounds
Solved the SDP for the five liver cases
Obtained a lower bound m∗ on the mean BED to healthy liver tissue
Compared the reduction achieved by the nonuniform plans to thebound on the maximum possible reduction
gap closed =munif −mnonunif
munif −m∗
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 28 / 31
The SDP Relaxation
Results
The nonuniform plans closed 78%-96% of the gap
Case Description Mean liver BEDin uniformreference plan
Mean liver BEDin nonuniformplan
SDPLowerBound
Gapclosed
1 Large central lesion 84.54 75.87 73.38 77.69%2 Small lesion 26.14 19.47 18.58 88.23%3 Two small lesions 59.54 50.24 48.03 80.80%4 Lesion abutting chest wall 47.51 38.65 37.65 89.86%5 Lesion abutting GI tract 88.67 77.38 77.02 96.91%
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 29 / 31
Future Directions
Ideas for Future Research
Develop a method to solve the large-scale SDP for athree-dimensional patient
A variable for each beamlet (hundreds)A constraint for each voxel (tens of millions)
Find even better local solutions, or tighten the lower bound, or both
Incorporate uncertainty into the model
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 30 / 31
Future Directions
Conclusion
Locally optimal nonuniformly fractionated treatment plans
Maintained treatment effectiveness in the tumor
Reduced Biologically Effective Dose (BED) in healthy liver tissue by13-35%
Closed 78-96% of the bound on maximum achievable benefit(solutions are near the globally optimal treatment plan)
Melissa R. Gaddy (NCSU) Nonuniform Fractionation in Radiotherapy May 2, 2017 31 / 31