PaperlayerView exact revision on arXiv

Source-linked AI summary

Multicriteria VMAT optimization

David Craft, Dualta McQuaid, Jeremiah Wala, Wei Chen, Ehsan Salari, Thomas Bortfeld

arXiv:1105.4109v3physics.med-ph

TL;DR

VMAT planning is challenging because delivery constraints couple adjacent angles and create a large-scale nonconvex optimization problem. VMERGE selects a plan from a 180-beam multicriteria solution and makes it deliverable through validated fluence-map merging. The method provides a user-specified epsilon-optimality guarantee for single-arc coplanar VMAT plans.

  • Problem

    VMAT planning couples neighboring-angle leaf motion with dose delivery, creating a large-scale nonconvex optimization problem.

    Source: §1 · p. 2

  • Method

    VMERGE selects a preferred 180-beam multicriteria plan, then merges neighboring fluence maps while maintaining dose quality to improve delivery efficiency.

    Source: §1 · pp. 2–3

  • Results

    For single-arc coplanar VMAT, the method guarantees a dose distribution within a user-specified tolerance of optimality.

    Source: §4 · p. 18

  • Takeaways & Limitations

    Planners can navigate planning-objective tradeoffs and explore the tradeoff between dose quality and delivery time.

    Source: §1 · p. 2, §4 · p. 17

    The merging strategy is greedy, and whether other strategies or dose-distribution-based similarity could be more successful remains to be investigated.

    Source: §4 · p. 17

Abstract

from arXiv · show

Purpose: To make the planning of volumetric modulated arc therapy (VMAT) faster and to explore the tradeoffs between planning objectives and delivery efficiency. Methods: A convex multicriteria dose optimization problem is solved for an angular grid of 180 equi-spaced beams. This allows the planner to navigate the ideal dose distribution Pareto surface and select a plan of desired target coverage versus organ at risk sparing. The selected plan is then made VMAT deliverable by a fluence map merging and sequencing algorithm, which combines neighboring fluence maps based on a similarity score and then delivers the merged maps together, simplifying delivery. Successive merges are made as long as the dose distribution quality is maintained. The complete algorithm is called VMERGE. Results: VMERGE is applied to three cases: a prostate, a pancreas, and a brain. In each case, the selected Pareto-optimal plan is matched almost exactly with the VMAT merging routine, resulting in a high quality plan delivered with a single arc in less than five minutes on average. VMERGE offers significant improvements over existing VMAT algorithms. The first is the multicriteria planning aspect, which greatly speeds up planning time and allows the user to select the plan which represents the most desirable compromise between target coverage and organ at risk sparing. The second is the user-chosen epsilon-optimality guarantee of the final VMAT plan. Finally, the user can explore the tradeoff between delivery time and plan quality, which is a fundamental aspect of VMAT that cannot be easily investigated with current commercial planning systems.

1 Introduction

VMAT planning is challenging because delivery constraints couple neighboring angles, making delivery-aware optimization large and nonconvex. VMERGE starts from a finely sampled ideal plan and makes it deliverable through dose-validated fluence-map merging, while letting planners explore planning and delivery tradeoffs.

1 Introduction

  • VMAT planning is larger than IMRT optimization because it considers every angle, and neighboring-angle leaf motion creates a large-scale nonconvex problem.The paper notes that these challenges can make VMAT planning much more time-consuming than IMRT planning.Source: §1 · p. 2
  • Interactive multicriteria optimization lets planners navigate Pareto-optimal plans to select a preferred tradeoff between target coverage and healthy-organ sparing.The paper applies this approach to explore both planning-objective tradeoffs and the tradeoff between dose quality and delivery efficiency.Source: §1 · p. 2
  • A 2 Gy fraction over one arc at maximum dose rate and gantry speed implies average segments about one-third the size of a 10 cm×10 cm field.At half gantry speed, the example predicts segments with a further 50% reduction in size.Source: §1 · p. 3
  • Gantry slowdown may be needed when fluence exceeds the dose deliverable at top speed or when leaf travel cannot complete the required modulation in time.The proposed merging routine validates dose quality to remove unnecessary slowdowns caused by over-delivery of fluence maps.Source: §1 · p. 3
  • Prior VMAT approaches include Otto et al.’s nonconvex direct-parameter search and Bzdusek et al.’s convex fluence optimization followed by sequencing.The paper uses “IMRT” for fixed-angle delivery from a small number of gantry positions, while noting that the term also includes arc therapies.Source: §1 · pp. 1, 4
  • VMERGE starts with fine beam spacing and unidirectional leaf sequencing, then merges neighboring maps while maintaining dose quality to improve delivery efficiency.The user-chosen epsilon tolerance controls how close the deliverable plan remains to the ideal dose solution.Source: §1 · p. 4

2 Methods

The method first selects an ideal dosimetric plan from a 180-beam multicriteria IMRT solution. It then successively coarsens the fluence maps to speed delivery while keeping dose quality within user-selected bounds.

2 Methods

  • A multicriteria problem over 180 equi-spaced beams lets the planner select a best-compromise plan for target coverage and healthy-organ sparing.This selected plan is treated as ideal because treatment time is not considered at this stage.Source: §2 · p. 4
  • The sequencing step successively coarsens the fluence maps to make delivery faster while keeping dose quality within user-selected bounds.This lets the user explore the tradeoff between dose quality and delivery time.Source: §2 · p. 5
  • The following sections describe the components of the VMAT planning approach.The overview identifies the ideal-plan optimization and delivery coarsening as its two stages.Source: §2 · pp. 4–5

2.1 180 beam IMRT solution and Pareto surface plan selection

The 180-beam stage approximates a Pareto surface of feasible plans, then lets the planner select a preferred convex combination. Fluence-map smoothing is also used to make similarity comparisons more meaningful and improve delivery time.

2.1 180 beam IMRT solution and Pareto surface plan selection

  • The dose model uses voxel doses d, dose-influence matrix D, and concatenated beamlet fluences f, with a convex constraint set of simultaneously achievable dose bounds.Conflicting or harder dosimetric goals are treated as objectives rather than constraints to avoid infeasibility.Source: §2.1 · p. 5
  • The objectives include minimizing maximum or mean structure dose and maximizing minimum or mean structure dose, using convex functions suited to a linear solver.The authors state these objectives are typically sufficient for high-quality plans in multicriteria planning.Source: §2.1 · p. 5
  • The solver approximates the Pareto surface by tightening objective bounds, after which the planner selects a preferred convex combination of calculated plans.For the studied cases, selections emphasized conformal target dose and mean-dose sparing of important critical structures.Source: §2.1 · pp. 5–6

2.1.1 Fluence map smoothing

  • Max beamlet smoothing caps fluence levels, while SPG smoothing periodically reduces the largest-SPG row’s peaks and redistributes fluence to adjacent beamlets.Both smoothing methods are described as convex; smoothing can make map-similarity comparisons more meaningful and reduce final delivery time.Source: §2.1.1 · p. 6

2.2 Unidirectional leaf sequencing

Unidirectional sequencing produces deliverable leaf trajectories while limiting leaf speed across and between fluence maps. Alternating sweep directions also hide closed-leaf gaps beneath a collimator jaw, though minimum leaf gaps and leakage are not explicitly modeled.

2.2 Unidirectional leaf sequencing

  • Alternating left-to-right and right-to-left dynamic leaf sweeps keep leaves aligned at opposite field edges between arc portions, hiding closed-leaf gaps beneath a collimator jaw.This reduces undesirable leakage radiation.Source: §2.2 · p. 6
  • Leaf trajectories must respect a maximum leaf velocity both within a fluence map and between successive maps.The sliding-window trajectories use local fluence gradients and maximum leaf velocity, with constant dose rate over each arc portion.Source: §2.2 · pp. 6–7
  • The traversal time for all leaf pairs depends on field width and the fluence map's SPG, divided by dose rate; gantry speed or dose rate is adjusted when leaf and gantry times mismatch.Minimum leaf gaps and leaf leakage are not explicitly considered.Source: §2.2 · pp. 6–7

2.3 Merging neighboring fluence maps

Neighboring fluence maps are greedily merged to shorten beam-on time while preserving the optimized dose distribution within a user-selected tolerance. The planner can stop merging to choose a desired balance between delivery time and plan quality.

2.3 Merging neighboring fluence maps

  • Merging neighboring fluence maps reduces leaf travel distance, allowing the gantry to move faster and lowering beam-on treatment time.Each distinct map otherwise requires a full unidirectional leaf sweep across its assigned arc portion.Source: §2.3 · p. 8
  • The similarity score favors merging maps that are alike and span smaller arc portions, which is expected to have less effect on dose distribution.The score uses the Frobenius norm of the difference between maps, normalized by their arc lengths and scaled by their combined arc length.Source: §2.3 · p. 8
  • At each greedy iteration, the neighboring pair with the lowest similarity score is merged into a summed fluence map covering the union of their arc portions.After n iterations, 180 initial maps become 180 − n maps; each merge averages the maps over their combined arc length.Source: §2.3 · p. 8
  • The stopping criterion lets the planner trade treatment time against plan quality, with the final dose distribution intended to remain nearly identical to the original 180-beam plan and within a user-chosen epsilon of optimality.Merged maps are sequenced and their delivered fluence is assigned to the nearest original 2° map for simulation.Source: §2.3 · pp. 8–9

2.4 VMAT settings

The delivery settings specify maximum gantry, leaf, and dose rates for 2 Gy fraction plans; fraction dose affects delivery because it is linked to these limits.

2.4 VMAT settings

  • The delivery limits are 1 rotation/min maximum gantry speed, 2.5 cm/sec maximum leaf speed, and 600 MU/min maximum dose rate.Each case is designed as a 2 Gy fraction plan.Source: §2.4 · p. 9
  • Fraction dose matters for VMAT because it is linked to dose rate, gantry speed, and leaf speed.This contrasts with step-and-shoot IMRT optimization, where fraction dose scaling does not fundamentally affect the plan.Source: §2.4 · p. 9
  • Displayed DVHs are scaled to total dose across all fractions, and optimization formulations are also shown for total dose.Source: §2.4 · p. 9

2.5 Clinical Cases

VMERGE is evaluated on prostate, pancreas, and brain cases with distinct prescriptions and voxel and beamlet resolutions. The formulations use site-specific dose objectives, with target coverage fixed in the prostate case.

2.5 Clinical Cases

  • Three clinical cases use distinct prescriptions and spatial resolutions: prostate at 77 Gy, pancreas at 50.4 Gy, and brain at 32 Gy.The cases use 1 cm × 1 cm beamlets for prostate and pancreas and 0.5 cm × 0.5 cm beamlets for brain.Source: §2.5 · p. 9

Prostate formulation

  • With prostate coverage fixed at 77 Gy, the formulation minimizes mean rectum, bladder, and unclassified-tissue doses.Unclassified tissue comprises voxels outside all other structures.Source: §2.5 · p. 9, Prostate formulation · p. 10

Pancreas formulation

  • The pancreas formulation minimizes mean shell, kidney, liver, and stomach doses while maximizing minimum target dose.The shell is a 0.7 cm band around the target used to promote dose conformality.Source: Pancreas formulation · p. 10

Brain formulation

  • The brain formulation minimizes mean chiasm, brainstem, and unclassified-tissue doses while maximizing minimum target dose.The two adjacent lesions, including one inside the brainstem, are treated as a single combined target structure.Source: §2.5 · p. 9, Brain formulation · p. 10

2.6 Sensitivity to algorithm settings

The prostate sensitivity tests use a 2° angular grid as the baseline, then examine a coarser 4° grid and smaller beamlets. These variations test whether the VMAT solution depends on discretization choices.

2.6 Sensitivity to algorithm settings

  • The sensitivity study compares a 2° baseline grid with a 4° grid, then halves beamlet width on the coarser grid.The 2° spacing reflects the final-dose calculation used by most commercial VMAT solutions; the smaller beamlets are 0.5 cm × 1 cm.Source: §2.6 · pp. 10–11
  • The tests assess whether the resulting VMAT plan remains fundamentally correct rather than sensitive to these parameter choices.Source: §2.6 · p. 10

3 Results

Across prostate, pancreas, and brain cases, iterative fluence-map merging reduces delivery time while retaining plans close to the original dose distributions. The prostate tests also show that smoothing and discretization affect delivery time or dose shaping without undermining the method’s stability.

3.1.1 Sensitivity: fluence map smoothing

  • 413, 339, and 311 seconds are the prostate delivery times for no smoothing, max-beamlet smoothing, and SPG smoothing, respectively.The final plans are highly similar; SPG smoothing is used for the nominal cases because it improves delivery time.Source: §3.1.1 · p. 11

3.1 Prostate case

  • 311 s with 35 arc portions is selected for the prostate case, versus 906 s initially, before target coverage begins to fall after 145 iterations.The merged plan maintains 95.3% of prostate volume at 77 Gy; mean rectum and bladder doses are each 0.50 × 77 Gy.Source: §3.1 · p. 11, Figure 1 · p. 12

3.1.2 Sensitivity: discretization settings

  • Reducing beamlet size lowers mean anterior-rectum dose from 50% to 44% of prescription, while final-plan treatment times and target coverage remain similar.The study reports that algorithm stability and usefulness are unchanged by beamlet size.Source: §3.1.2 · pp. 12–13, Figure 2 · p. 13

3.2 Pancreas case

  • 202 s with 20 fluence maps retains nearly the original 180-beam DVH in the pancreas case, with 98.3% of tumor volume receiving 50.4 Gy.Plan quality begins to degrade after 160 merges; the initial delivery time was 859 s.Source: §3.2 · p. 13, Figure 3 · p. 14, §3.3 · p. 14

3.3 Brain case

  • 228 s with 25 fluence maps retains nearly the original brain-case DVH, with 97.6% of target volume receiving 32 Gy.Quality begins to degrade after 155 merges, when mean optic-chiasm dose starts to rise.Source: §3.3 · pp. 13–14

4 Discussion and Conclusions

VMERGE starts from a finely sampled, relaxed solution and coarsens it toward a deliverable plan while preserving dose quality within a user-chosen tolerance. Its discussion highlights the role of finite leaf speed, the quality–time tradeoff, and extensions and limitations of the approach.

4 Discussion and Conclusions

  • Finite leaf speed makes single-arc coplanar VMAT challenging because even low-SPG fields may require slower gantry motion to give leaves time to travel.With infinite leaf speed, SPG alone governs plan complexity and delivery time and can be minimized in a convex setting.Source: §4 · p. 15
  • Unlike coarse-to-fine VMAT methods, VMERGE begins at the global minimum of a relaxed convex problem and modifies it toward deliverability, while the compared approaches offer no optimality guarantee.The authors note that this distinction matters for treatment-modality comparisons, which can be muddled without optimality guarantees.Source: §4 · p. 16
  • Max beamlet smoothing reduces treatment time from 413 seconds to 339 seconds (an 18% reduction), and SPG smoothing reduces it by another 28 seconds.Neither smoothing approach dramatically changes the final dose distribution; SPG smoothing is more useful for higher dose-per-fraction plans.Source: §4 · p. 16
  • Adjacent fluence maps can be similar because their dose-influence matrices are similar, and successive greedy merges can preserve dose quality; large arc portions tend to avoid critical structures.The discussion relates larger arc portions to target coverage and notes that evenly distributed pancreatic OARs correspond to more uniform arc portions; other merging strategies remain to be investigated.Source: §4 · pp. 16–17
  • A complete VMERGE planning session is estimated at 35 minutes, or about 15 minutes if dose-influence matrix calculation and Pareto-surface computation are parallelized.Merging and sequencing takes about five minutes when evaluating the quality–time curve, or about 10 seconds when the desired number of merges is known in advance.Source: §4 · p. 17
  • The initial 180-beam solution uses simple formulations, and naive pencil-beam dose calculations may degrade after final calculations include output-factor corrections and leaf transmissions.The discussion also identifies collimator rotation, couch rotation, and non-unidirectional sequencing as possible extensions; couch rotation expands the optimization space substantially.Source: §4 · pp. 17–18
  • Because VMERGE allows the gantry to slow as needed, multiple arcs offer no dosimetric advantage; a double arc may save time for many highly modulated maps that benefit from resetting leaf positions.The authors speculate that such maps will not often arise clinically, while partial arcs could save time by focusing delivery on more beneficial angles.Source: §4 · pp. 17–18
  • For single-arc coplanar VMAT, VMERGE guarantees a dose distribution within a user-specified tolerance of the provably optimal 180-beam solution, a guarantee the authors say commercial systems do not provide.The conclusion notes that non-coplanar arcs, dynamic collimator rotation, optimal partial-arc creation, and integrating MCO exploration with delivery-efficiency tradeoffs remain open challenges.Source: §4 · p. 18
Loading 1105.4109v3…