28
1 1 Introducing Interactive Planning on the CyberKnife ® System: The New VOLO Optimizer © Accuray and/or its affiliates. All rights reserved. Accuray confidential. |

Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

11

Introducing Interactive Planning on the CyberKnife®

System:The New VOLO™

Optimizer

© Accuray and/or its affiliates. All rights reserved. Accuray confidential. |

Page 2: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

2

• Quality• Excellent dosimetric plan quality for simpler cases• Excellent dosimetric plan quality for difficult and larger cases• Efficiently delivered plans

• Performance• Fast solution times• Predictably responsive to inputs and changes to inputs• Final solution minimally different than the optimized solution

• Ease of Use• Provide an optimization approach is familiar to many in the treatment

planning community

CyberKnife® VOLO™ Optimizer

Page 3: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

3

• User Interface• Familiar optimization approach; weighted cost function• Easy to setup for simple cases• Easy to modify for challenging cases

• Framework allows extension to other developments• Interactive optimization• Automatic optimization• Biological optimization goals

CyberKnife® VOLO™ OptimizerGoals for new optimization

Page 4: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

4

• Very fast convergence for all CyberKnife® collimation options, especially compared to Sequential

• For MLC, ability to explore tradeoffs in “Live” mode – OAR sparing versus target coverage

• For Iris™/ Fixed, fast convergence still provides rapid review of tradeoffs• All plans are deliverable and final in the optimization step• Ability to control delivery efficiency (MU/ Treatment Time) via simple

parameters• Total MU Penalty and Min Beam MU per Fraction

• No manual post-processing for removal of low MU beams (no explicit Beam, Segment, Node, Time Reduction)

• Sampling has minimal impact on performance; more intuitive control of sampling on a per VOI basis

• Ability to optimize at High resolution for optimal plan review

CyberKnife® VOLO™ OptimizerBenefits

Page 5: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

5

• Segment shape adaptation improves plan quality and obviates needs for manual selection of MLC shape properties

• Multiple targets and multiple dose levels handled transparently• Monte Carlo optimization is integrated into the workflow for MLC and Iris

i.e. perform MC optimization in a single step• MC calculation post non-MC optimization is also permitted

• Ability to choose number of nodes for both Iris and MLC • For MLC, support for 3D conformal like mode with a single segment per

node that can then be adapted. • Simplifies or obviates need for QA

• Ability to stop the optimizer during MLC segmentation or Iris/ Fixed collimator optimization with a valid, deliverable plan

CyberKnife® VOLO™ OptimizerBenefits

Page 6: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

6

CyberKnife® VOLO™ OptimizerIris™/Fixed Collimator optimization workflow

VOI Sampling Node Selection Candidate Beam Selection Beam Dose

Beam Weight Optimization with low MU pruning

(stop early if MC)

MC Beam Dose Final Beam Weight Optimization with

pruningFinal Dose Calculation

If final dose calculation algorithm is Monte Carlo

Page 7: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

7

CyberKnife® VOLO™ OptimizerInCise™ MLC optimization workflow

VOI Sampling Node selection Beamportselection

Beamlet selection Beamlet Dose

Fluence Map Optimization

including Smoothing

Segment Generation Segment Dose

Segment Weight Optimization with low MU pruning

Segment Shape Adaptation

Segment Dose Calculation –

FSPB, FSPB-LS or Monte Carlo

Final Segment Weight

Optimization with low MU pruning

Final Dose Calculation

Page 8: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

8

CyberKnife® VOLO™ OptimizerInCise™ MLC optimization workflow

• Choose number of nodes• Perform node selection and targeting• Optimize fluence map• Optimize segments

• Includes shape adaptation• Includes automatic pruning to respect minimum

MU deliverable segments• Compute final dose

• Automatically on completion of all optimization phases

• Final IS Final – plan can be saved and approved in this state

Page 9: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

9

• Smoothness Penalty (MLC)• Controls the degree of smoothness of the fluence map.

• Total MU Penalty• Replaces Segment/ Beam, Node and Time Reduction in

one a priori objective• # of optimization iterations• # of adaptation iterations

• Zero and One (3D Conformal) are permissible• Max Segments per Beam

• Change for larger targets• Max Segments• Max Segment MU

• Pruning steps will remove 0 weighted apertures and those between 0 and the min. Remaining will be re-optimized

• Max Beamlet MU• Min Segment MU per Fraction

CyberKnife® VOLO™ OptimizerInCise™ MLC optimization workflow parameters

Page 10: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

10

Parameter Range, Default

Phase Effect

Fluence Smoothness Penalty

[0, 10], 1 Fluence Optimization Controls the degree of modulation. Lower modulation gives better correspondence between fluence and segment DVH.

Total MU Penalty [0, 10], 1 Segment Weight Optimization

Controls plan efficiency – MU and segments. Higher efficiency results in possible greater DVH deviation going from fluence to segments.

# of Optimization Iterations

[50, 500], 50 Segment Weight Optimization and Segment Shape Adaptation

Higher means overall slower result. Too few and solution may not have converged. Fluence optimization continues to convergences in a Live mode or until Stopped by the user.

# of Adaptation Iterations

[0, 5], 3 Number of shape adaptation rounds

To improve plan quality, each adaptation allows leaf positions to vary within one beamlet’s extents. A leaf can potentially be pushed by a distance of ‘n’ beamlets to the left or right, where n is the number of adaptation iterations.

Max Segments Per Beam

[1, 50], 5 Segment Generation Limits the number of segments that can be generated from each fluence map. Indirectly effects plan efficiency.

Max Segments [20, 500], 300 Pruning Controls plan efficiency . Pruning is performed iteratively after each segment weight optimization. Higher the value of this parameter, better is the correspondence between fluence and segment DVH but lower the efficiency.

Max Segment MU [100, 5000], 500

Segment Weight Optimization

MU weight limit, used to spread out dose and limit potential for dose. fingers

Max Beamlet MU [120, 6000], 600

Fluence Optimization Beamlet MU is applied during fluence optimization.May need to increase if DVH shows lack of coverage after fluence optimization (MU starvation).

Min Segment MU per Fraction

[2, 100], 2 Pruning Controls plan efficiency . Pruning is performed iteratively after each segment weight optimization. Segments with MU below this value will be removed. Increasing this value to create more efficient plans.

CyberKnife® VOLO™ OptimizerInCise™ MLC optimization workflow parameters

Page 11: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

11

CyberKnife® VOLO™ OptimizerIris™ / Fixed optimization workflow

• Choose collimators• Choose number of nodes• Perform node selection and targeting• Either:

• Optimize to completion -• Stop and manually compute final dose

• Final IS Final • Automatically on completion of all

optimization phases• Final IS Final – plan can be saved

and approved in this state

Page 12: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

12

• Total MU Penalty• Replaces Segment/ Beam, Node and Time

Reduction in one a priori objective

• # of optimization iterations• Max Beams

• Used to determine the starting point for Monte Carlo optimization

• Max Beam MU• Pruning steps will remove 0 weighted apertures

and those between 0 and the min. Remaining will be re-optimize

• Min Segment MU per Fraction

CyberKnife® VOLO™ OptimizerIris™ parameters

Page 13: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

13

CyberKnife® VOLO™ OptimizerOptimization for efficiency - total MU penalty

Observations:• The number of segments and total MU are lower for higher total MU

penalties• The deviation between final and fluence DVH is greater for higher total MU

penalties

Total MU Penalty=0 Total MU Penalty=0.5 Total MU Penalty=1.0

Page 14: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

14

• Accuray customers submitted patient de-identified treatment plans delivered on CyberKnife® or on linacs

• Plans included clinical goals (prescription), dose distribution, treatment planning and delivery times

• Accuray subject matter experts (SMEs) re-planned the cases with VOLO™

• Plans were compared using a third party DICOM viewer to compare clinical results

CyberKnife® VOLO™ Challenge

Page 15: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

15© Accuray and/or its affiliates. All rights reserved. Accuray confidential. |

CyberKnife® VOLO™ OptimizerClinical comparisons patients summary

Patient Region Specific Site Target Volume(cc) Prescription Fraction

1 Prostate Prostate 55.2 40Gy + 50Gy 5

2 Intracranial Skull Base 59.5 and 9.8 24Gy 3

3 Lung Peripheral 18.5 60Gy 3

4 Intracranial Brain mets 11.2 35Gy 5

5 Lung Peripheral 12.5 54Gy 1

6 Intracranial Brain mets 4.4 21Gy 1

7 Spine Lumbar Spine 185.8 45Gy 5

8 Intracranial Brain mets 0.34, 0.56 and 0.33cc 15Gy 1

Page 16: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

16© Accuray and/or its affiliates. All rights reserved. Accuray confidential. |

CyberKnife® VOLO™ OptimizerClinical comparisons outcomes summary

1. Created MLC (or IRIS) plans to show value of alternate collimator

Patient Specific Site

Coll-imation

Nodes/Beams

MU Planning Time(min)

Est. Tx Time (min)

Nodes/Beams

MU Planning Time (min)

Est. Tx Time(min)

1 Prostate IRIS 43, 244 80,000 120 49 70, 160 39,982 120 36

2 Skull Base

MLC 156, 184 19,337 420 40 60, 93 6,852 60 24

3 LungPeriph.

FixedMLC1

60, 134 35,601 240 42 41, 9652, 54

32,24920,765

240120

3323

4 Brain mets

FixedMLC1

105, 230 51,836 60-120 50 85, 16554, 60

10,8254,202

3015

3616

5 LungPeriph.

MLC 5 Beams Linac

12,611 200 33 31, 36 13,752 60 24

6 Brain mets

MLC IRIS1

4 BeamsLinac

5,362 200 25 27,4127, 121

3,9205962

6060

2026

7 Lumbar Spine

IRIS 86, 199 73,399 720 55 60, 83 31,211 180 21

8 Brain mets

FixedMLC1

na/91 18,089 n/a 35 40, 9547,72

1706610560

6015

3430

Reference Plan CyberKnife® VOLO™

Page 17: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

17

• Patient prescribed to UCLA protocol (A021501)

• 40Gy in 5 fractions to the PTV with a boost to 50Gy to an intra-prostatic lesion (50Gy)

• Rx isodose preferred in the range 50 to 62%

• Block testes and limit doses to urethra, rectum, rectal mucosa, bladder, bowel and femoral heads

• Customer plan is Iris with several sizes used. VOLO™ plans is repeated with similar size and an alternate MLC plan was also created

• Target size: 55.2cc

CyberKnife® VOLO™ OptimizerGenesis Healthcare HDR-like SBRT prostate

Reference (Iris™) VOLO™ MLCVOLO™ Iris™

Page 18: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

18

• VOLO™ plans are equivalent or slightly improved in most metrics

• VOLO™ plans are substantially more efficient

CyberKnife® VOLO™ OptimizerProstate case - plan metrics comparison

Reference (Iris) VOLO Iris VOLO MLC

PTV V40Gy > 95%(%) 95.1 96.2 96.0PZ D50% > 50Gy (Gy) 53.7 54.0 53.2SV V25Gy >95% (%) 97.2 95.8 96.8Urethra Max < 44Gy (Gy) 44.7 44.5 43.3

Rectal Mucosa V30Gy < 1% (%) 1.6 0.9 0.6

Rectum V20Gy < 50% 14.3 12.6 12.6Rectum V32Gy < 20% 4.0 3.3 2.9

Rectum V36Gy < 10% 2.2 1.3 1.3

Rectum V40Gy < 5% 0.9 0.1 0.2

Bladder V20y < 40% 17.9 24.9 19.9

Bladder V40Gy < 10% 1.7 2.4 2.7

MU 80,000 39982 22851

Nodes, Beams 43, 244 70, 160 53, 83

Treatment Time 49 36 24

Planning Time ~2 hours ~2 hours ~ 1 hour

Page 19: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

19

CyberKnife® VOLO™ OptimizerComparison with squential from Accuray demo database

VOLO Sequential

Nodes/Beams

MU (k)Tr Time (mins)

Nodes/Beams

MU (k)Tr Time (mins)

Prostate, PACE (MLC) 28/42 11.8k 12 76 19k 17

Prostate, RTOG0938 (MLC) 28/50 13.9 13 28/49 18.8 14

Lung, Central (MLC) 31/35 14.3 14 31/52 18.2 17

Liver (MLC) 23/36 12.0k 15 31/48 19.8k 20

Meningioma (MLC) 23/41 5.5k 13 20/83 9.8k 18

Brain Mets (MLC) 30/85 11.5k 30 27/60 14.5k 30

Pituitary (Iris) 43/70 9.1k 16 75/83 14.6k 21

Adrenal (Iris) 50/113 37.5k 33 90/176 73k 42

Acoustic (Iris) 28/45 5.1k 16 36/110 8.0k 24

Page 20: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

20

Sequential

CyberKnife® VOLO™ OptimizerAdrenal (Iris™)

Optimization Time: 40 minutesTime Reduction: 4 minutesFinal Dose Calc: ~1 minute

Optimization Time: 3 minutesTime Reduction: Not neededFinal Dose Calc: Not needed

VOLO™

Page 21: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

21

Sequential

CyberKnife® VOLO™ OptimizerMeningioma (MLC)

Opt. Time: ~30 minutes per iterationTime Reduction: 12 minutesFinal Dose Calc: ~1 minute

Optimization Time: 2-3 minutesTime Reduction: Not neededFinal Dose Calc: Not needed

VOLO™

Page 22: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

2222

CyberKnife®

VOLO™ OptimizerAlgorithm Details

Application to MLC

Page 23: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

23

• Dose calculation algorithm is Finite Size Pencil Beam (FSPB)• Dose per unit intensity stored for each beamlet at each sample point

• Stored as a sparse matrix (𝑑𝑑𝑖𝑖𝑖𝑖): Rows::Sample pts, Columns::Beamlets• Let 𝑤𝑤𝑖𝑖 be the jth beamlet intensity• Let 𝐷𝐷𝑖𝑖 be the total dose at the ith sample point• 𝐷𝐷𝑖𝑖 = ∑𝑖𝑖 𝑤𝑤𝑖𝑖𝑑𝑑𝑖𝑖𝑖𝑖

• Values < 0.001 cGy are not stored in [d] (cut-off is 0.005 for Iris/Fixed)• Compare this with 0.01 for Sequential, Iris™ / Fixed and MLC• Lower values possible because we are not computing beam/segment volumes

before optimization

• GPU implementation ensures high performance• Isodose lines are not displayed since dose is available only at sample pts• DVH calculation is approximated well as noted before

CyberKnife® VOLO™ OptimizerBeamlet dose calculation

Page 24: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

24

• Intensity modulation• Optimize beamlet intensities (𝑤𝑤𝑖𝑖)• Objective is a weighted sum of

one-sided quadratic dose penalties

• Smoothness penalty based on the spatial gradient of beamlet intensities

• 0 ≤ 𝑤𝑤𝑖𝑖≤ Max beamlet MU

CyberKnife® VOLO™ OptimizerFluence map optimization (FMO): overview

Page 25: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

25

• Fluence maps can be noisy• Since each beamlet intensity is

independent

• Add a penalty for non-smoothness

• L1-norm of spatial gradient of intensities

• 𝜌𝜌∑𝑖𝑖𝜕𝜕𝑤𝑤𝑖𝑖𝜕𝜕𝜕𝜕

+ 𝜕𝜕𝑤𝑤𝑖𝑖𝜕𝜕𝜕𝜕

• Well-defined boundaries between intensity levels

• Easier to generate segments• User-specified weight controls level

of smoothness• Heuristic dynamic weight balances

this with dose cost

CyberKnife® VOLO™ OptimizerFMO: smoothness penalty

Page 26: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

26

• Reducing Level Algorithm*• For each fluence map, get the max intensity• Divide max intensity by 2 and choose as first intensity level for slicing• Select the fluence map region above this intensity level• Process this region and extract the largest deliverable segment• If no deliverable segment, select next level as half of previous and try again• Subtract the selected intensity level from the generated segment region of the fluence

map• Update fluence map with the residual intensities• Select next intensity level as half of previous and repeat• Stop when residual intensity drops below a threshold

• Sort all generated segments by MU*area and select the requested max number per beamport

CyberKnife® VOLO™ OptimizerSegment generation

*P Xia, LI Verhey, “Multileaf collimator leaf sequencing algorithm for intensity modulated beams with multiple static segments”, Med. Physics 25 (1998), p. 1424-1434

Page 27: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

27

• Optimize segment MUs (𝑤𝑤𝑖𝑖)• During optimization 0 ≤ 𝑤𝑤𝑖𝑖≤ Max Segment MU• Objective is a weighted sum of one-sided quadratic dose penalties• Total MU penalty is added for treatment efficiency• User-specified weight to control the effect of total MU penalty• Heuristic dynamic weight to balance Total MU and Dose/ DV

penalties• Pruning segments below min MU ensures deliverability

• No more than 2% of total MU is pruned at each iteration• Re-optimize weights if any segment with non-zero MU is pruned• L-BFGS-B is used for optimization

CyberKnife® VOLO™ OptimizerSegment weight optimization

Page 28: Introducing Interactive Planning on the CyberKnife System · 9 • Smoothness Penalty (MLC) • Controls the degree of smoothness of the fluence map. • Total MU Penalty • Replaces

28

• Initially segments have leaves positioned at beamlet boundaries

• By allowing leaves to be positioned anywhere within a beamlet• OARs nearby targets can be spared better• Target homogeneity can be improved

• In every segment, for each open leaf, identify a beamlet adjacent to it

• Optimize the open fraction of these boundary beamlets• Bounded between 0 and 1

• Objective is the weighted sum of one-sided quadratic dose penalties

• In order to approximate dose for each adapted segment shape• Reference segment dose for initial segment shape is adjusted using fractional beamlet dose • Heuristic scaling factors are used to adjust beamlet dose based on the containing segment

• Number of adaptations can be controlled by the user

• Adapted shapes are processed to form deliverable segments

CyberKnife® VOLO™ OptimizerSegment shape adaptation*

*Cassioli & Unkelbach, “Aperture shape Optimization for IMRT treatment planning”, Phys. Med. Biol. 58 (2013) 301-318