acceptodds
Under review as a conference paper at ICLR 2027

FROM ANATOMY TO DELIVERY: PHYSICS-ALIGNED CREDIT ASSIGNMENT FOR VMAT RADIOTHERAPY PLANNING

Abstract

Radiation therapy planning translates a precise spatial balance between curative target coverage and healthy organ sparing into executable machine instructions. Volumetric modulated arc therapy (VMAT) realizes this delivery through a rotational trajectory of tightly coupled multileaf collimator apertures and beam intensities. Existing automated approaches struggle to connect clinical objectives directly to these machine parameters. A common approach to addressing this challenge is to use two-stage pipelines that optimize intermediate dose representations without physical constraints, causing the clinical trade-offs to degrade when eventually converted into deliverable actions. Conversely, methods that act directly on machine variables typically rely on a single plan-level clinical reward. This global scalar creates a severe credit assignment bottleneck, failing to isolate which specific leaf or intensity adjustment improved the target or harmed an organ. Differentiating the clinical objective provides local dose sensitivities for every machine variable. We find that translating these independent local gradients into a continuous, deliverable trajectory requires separating physical credit construction from policy learning. We introduce Physics-Aligned Credit for VMAT (PAC-VMAT) to realize this separation. The framework evaluates the evolving plan through a dose engine to construct explicit, fixed credit targets. Parameterized Leaf and MU policy agents then learn to coordinate these local clinical sensitivities into mechanically feasible aperture and intensity updates. Evaluated on lung and head-and-neck benchmarks, PAC-VMAT substantially improves the clinical target-organ balance over scalar-reward reinforcement learning and direct machine optimization, achieving plan quality comparable to closed-source commercial engines. This direct translation from patient anatomy to executable trajectories reduces optimization time to minutes, safely lowering collateral organ exposure while strictly satisfying mechanical delivery constraints.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
radiotherapy planning, VMAT, reinforcement learning, credit assignment, physics-informed learning, medical AI, treatment planning.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uLbPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PAC-VMAT, a framework for optimizing radiotherapy plans directly through machine parameters. It uses a differentiable dose engine to compute clinical sensitivities and converts them into fixed credit targets for separate leaf and MU policies. The method aims to produce favorable dose distributions while satisfying mechanical delivery constraints. Experiments on 70 lung and 67 head-and-neck cases show improvements over direct optimization and adapted RL baselines, with additional comparisons to commercial treatment plans.

Critical strengths
  1. The paper addresses a practical problem in radiotherapy planning. Optimizing machine parameters directly avoids the mismatch between predicted dose distributions and executable trajectories. I also find the use of physical sensitivities for credit assignment interesting, especially compared with relying on a single plan-level reward.

  2. The experimental results are encouraging. The evaluation covers 137 cases across two anatomical sites, and Figure 5 shows improvements over direct machine optimization, including a reduction in lung target hotspot from 67.43 to 64.12 Gy. The attention to mechanical constraints and implementation details also strengthens the work.

Critical weaknesses
  1. I have some concerns about dose calculation consistency. Table 1 shows a large discrepancy between Eclipse and PyDoseRT, with the lung target hotspot changing from 63.38 to 112.77 Gy for the commercial reference plans. This may involve normalization or calibration differences, but it makes the comparison with commercial plans difficult to interpret. Independent dose recalculation would be helpful.

  2. The contribution of the proposed credit assignment is not fully isolated. The direct optimization baseline differs in update schedules, weighting, and regularization, so the gains cannot be attributed entirely to the fixed credit targets. The RL comparisons are also somewhat limited because the evaluated versions use patient-specific adaptations rather than the original cross-patient training procedures.

Questions for authors
  1. What explains the large differences between Eclipse and PyDoseRT dose calculations in Table 1? Are these mainly caused by normalization or machine calibration settings?

  2. Have the PAC-optimized plans been recalculated using an independent clinical treatment planning system? If so, do the improvements in target coverage and organ sparing remain?

  3. Can the authors compare PAC against direct gradient optimization with matched update schedules, normalization, regularization, and computational budget? This would help isolate the benefit of the proposed credit targets.

  4. Are the improvements in organ doses evaluated at comparable target coverage? It would be useful to know whether the reductions remain when the target coverage requirements are held fixed.

  5. Could the authors clarify how the RL baselines were adapted, particularly which parts of their original training procedures were removed or changed? How should these results be interpreted relative to the original methods?

  6. What is the actual runtime for the 70-step configuration used in the main results? The reported 4.90-minute estimate appears to correspond to only 20 steps.

7 comments
AuthorsOpenReview ↗

Author Response (W1, Q1, Q2), Part 1: Dose-Engine Consistency and Independent Recalculation

We thank the reviewer for the constructive comments and respond to each point below.

W1 & Q1 & Q2: Dose-Engine Consistency and Independent Recalculation

1. Cross-engine transfer and coverage rescaling. The two values come from different evaluation conditions. The Eclipse row in Table 1 evaluates the GDP-HMM reference plans in their original commercial environment [1], whereas the Proprietary plans row recalculates the same Leaf and MU trajectories in PyDoseRT without re-optimization. This cross-engine transfer lowers unscaled target coverage, and rescaling the dose by 60/D9560/D_{95} to match target coverage amplifies the hotspot to 112.77 Gy. PAC-VMAT and all primary baselines are optimized and evaluated within PyDoseRT, a commissioned dose engine validated against physical measurements and 181 clinical VMAT plans [2].

2. Independent dose recalculation. We further recalculate the GDP-HMM and PAC-VMAT trajectories with three independent open-source dose engines, using the same CT volumes, structure masks, and machine parameters, following independent calculation-based dose verification practice for IMRT and VMAT [3].

Two engines use Monte Carlo transport, the reference approach for heterogeneous anatomy such as the lung [4]: ompMC, an OpenMP-parallel re-implementation of EGSnrc coupled photon–electron transport [5], and EGSnrc DOSXYZnrc [6, 7], with each arc simulated through its synchronized phase-space source [8]. The third engine, pyRadPlan, is the Python successor of the open-source planning toolkit matRad [9] and computes photon dose with the singular-value-decomposition pencil-beam algorithm [10]. Tab. R1 reports unscaled target coverage alongside dosimetric endpoints evaluated at matched target coverage.

Tab. R1: Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(a) GDP-HMM reference delivery

MetricLimitEclipsePyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.10 ± 0.2334.24 ± 6.4346.75 ± 1.0946.26 ± 1.5447.60 ± 1.37
PTV D95 after scaling≥ 6060.00 ± 0.2360.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7263.38 ± 0.99112.77 ± 19.0867.15 ± 1.2268.51 ± 2.7281.24 ± 6.81
Lung–CTV V20 ↓≤ 3720.64 ± 10.6426.48 ± 15.2420.16 ± 10.6220.31 ± 10.5921.33 ± 11.12
Lung–CTV mean ↓≤ 2011.79 ± 5.0114.79 ± 7.3611.53 ± 5.0511.59 ± 5.0312.21 ± 5.41
Spinal cord D0.03cc ↓≤ 50.541.61 ± 26.0937.08 ± 20.6327.88 ± 12.6628.31 ± 12.6730.01 ± 13.19
Esophagus mean ↓< 3419.73 ± 4.2022.42 ± 14.9314.67 ± 8.9214.85 ± 8.9915.52 ± 9.25
Heart D66.7% ↓< 452.44 ± 2.752.81 ± 2.742.27 ± 2.702.22 ± 2.712.69 ± 2.95
AuthorsOpenReview ↗

Author Response (W1, Q1, Q2), Part 2: Dose-Engine Consistency and Independent Recalculation

W1 & Q1 & Q2 (continued): Dose-Engine Consistency and Independent Recalculation

Tab. R1 (continued): Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(b) PAC-VMAT delivery

MetricLimitPyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.00 ± 0.0052.93 ± 4.0252.49 ± 4.6155.64 ± 3.42
PTV D95 after scaling≥ 6060.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7264.12 ± 2.7372.42 ± 3.6373.35 ± 4.3475.00 ± 5.56
Lung–CTV V20 ↓≤ 3719.20 ± 9.0520.19 ± 8.6920.27 ± 8.6620.21 ± 8.70
Lung–CTV mean ↓≤ 2010.39 ± 4.6312.05 ± 5.2512.07 ± 5.2012.27 ± 5.51
Spinal cord D0.03cc ↓≤ 50.536.37 ± 14.1344.81 ± 19.2845.30 ± 19.3344.32 ± 18.95
Esophagus mean ↓< 3415.16 ± 9.4915.53 ± 9.8915.69 ± 9.9815.93 ± 10.00
Heart D66.7% ↓< 451.79 ± 2.812.26 ± 3.902.18 ± 3.862.64 ± 4.03

Independent recalculation with ompMC, EGSnrc, and pyRadPlan shows that the 112.77 Gy hotspot is specific to transferring and rescaling the reference trajectories in PyDoseRT. PAC-VMAT maintains competitive target homogeneity and organ sparing across the independent dose calculations.

Takeaway: The Eclipse–PyDoseRT discrepancy reflects cross-engine transfer combined with coverage rescaling, while independent recalculations confirm that PAC-VMAT's plan quality is not specific to PyDoseRT.

[1] Gao et al., “Automating RT Planning at Scale: High Quality Data for AI Training,” 2025.
[2] Simkó et al., “A Physics-Informed, Plug-and-Play Dose Engine for Gradient-Based Radiotherapy Treatment Planning,” 2025.
[3] Zhu et al., “Report of AAPM Task Group 219 on Independent Calculation-Based Dose/MU Verification for IMRT,” 2021.
[4] Chetty et al., “Report of the AAPM Task Group No. 105: Issues Associated with Clinical Implementation of Monte Carlo-Based Photon and Electron External Beam Treatment Planning,” 2007.
[5] Doerner and Caprile, “Technical Note: Parallel Implementation of the EGSnrc Monte Carlo Simulation of Ionizing Radiation Transport Using OpenMP,” 2017.
[6] Kawrakow, “Accurate Condensed History Monte Carlo Simulation of Electron Transport. I. EGSnrc, the New EGS4 Version,” 2000.
[7] Kawrakow and Walters, “Efficient Photon Beam Dose Calculations Using DOSXYZnrc with BEAMnrc,” 2006.
[8] Lobo and Popescu, “Two New DOSXYZnrc Sources for 4D Monte Carlo Simulations of Continuously Variable Beam Configurations, with Applications to RapidArc, VMAT, TomoTherapy and CyberKnife,” 2010.
[9] Wieser et al., “Development of the Open-Source Dose Calculation and Optimization Toolkit matRad,” 2017.
[10] Bortfeld et al., “Decomposition of Pencil Beam Kernels for Fast Dose Calculations in Three-Dimensional Treatment Planning,” 1993.

AuthorsOpenReview ↗

Author Response (W2, Q3): Comparison with Direct Gradient Optimizers

W2 & Q3: Comparison with Direct Gradient Optimizers

1. Ablations on update schedules and credit targets. Appendix D.8 and Tables 10 and 11 control the update schedule and intermediate credit refresh, showing that neither factor accounts for the main improvement. Table 3 further holds the policy architecture, update schedule, and target-independent regularization fixed to show that learning from explicit credit targets substantially improves upon differentiating the clinical objective directly.

2. Controlled comparison with direct optimizers. We further compare PAC-VMAT with projected-gradient, Adam, and L-BFGS updates under the same initial plan, normalized credit targets, structural regularization, sequential Leaf–MU schedule, mechanical projection, and 70-step dose-evaluation budget.

Tab. R2: Direct Optimizer Comparison at the Same Target Coverage

MethodPTV D0.03cc ↓Lung–CTV mean ↓Spinal cord D0.03cc ↓Esophagus mean ↓
Projected gradient67.8410.8139.4216.08
L-BFGS66.4510.7338.3615.79
Adam65.8810.6437.8215.62
PAC-VMAT64.1210.3936.3715.16

Under the same credit targets, regularization, update schedule, and dose-evaluation budget, PAC-VMAT achieves the lowest target hotspot and organ doses across all four endpoints. The comparison isolates the benefit of learning structured updates through shared policy parameters beyond applying the same credit information with direct numerical optimizers.

Takeaway: PAC-VMAT's advantage persists after matching the gradients, update order, regularization, mechanical projection, and computational budget.

Reviewer_exAhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents PAC-VMAT, a physics-aligned optimization framework for volumetric modulated arc therapy planning. The method differentiates clinical objectives through a dose engine to construct explicit local credit targets for machine parameters, which are used to train patient-specific Leaf and MU policy agents. The framework combines structured policy updates with anatomy-guided initialization to generate mechanically constrained treatment trajectories. Experiments on lung and head-and-neck cohorts demonstrate substantial improvements over scalar-reward reinforcement learning and direct optimization baselines within the PyDoseRT environment, with reported dosimetric outcomes approaching those of commercial reference plans.

Critical strengths
  1. Clinically meaningful problem formulation with explicit physical feedback. The paper addresses an important challenge in automated radiotherapy planning: assigning local clinical objectives to thousands of coupled machine controls. Translating dose sensitivities into structured parameter updates is a well-motivated approach that connects optimization objectives with the underlying treatment delivery process.

  2. Substantial improvements within the evaluated optimization environment. PAC-VMAT demonstrates considerable improvements in target homogeneity and organ sparing compared with scalar-reward PPO, direct clinical-objective training, and reproduced automated planning baselines. The evaluation across two anatomical sites, together with the analysis of initialization and policy parameterization, provides useful evidence of the effectiveness of the optimization strategy within the chosen dose engine.

Critical weaknesses
  1. Serious dose-engine discrepancies undermine the clinical validity of the reported results. The same commercial reference plans yield substantially different dose distributions when evaluated using Eclipse and PyDoseRT. For example, the mean lung-cohort PTV D0.03cc changes from 63.38 Gy in Eclipse to 112.77 Gy after PyDoseRT recalculation (Appendix D.5). Although the authors acknowledge possible differences in tissue heterogeneity modeling and beam commissioning, such a large discrepancy raises fundamental concerns about the fidelity of the dose engine, particularly in heterogeneous lung tissue. Evaluating all methods within PyDoseRT provides an internally consistent comparison, but does not establish that the optimized plans would retain their apparent clinical advantages under independently validated dose calculations. Furthermore, the reported head-and-neck results exceed several listed organ-at-risk constraints, weakening the claims of clinically comparable plan quality.

  2. Insufficient evidence that learned credit-based optimization is necessary over stronger gradient-based alternatives. The proposed credit targets are constructed directly from gradients of the clinical objective, making the method closely related to structured gradient-guided optimization. While the paper compares against direct constrained optimization, the baseline uses joint machine-variable updates, whereas PAC-VMAT employs sequential Leaf–MU updates with intermediate gradient recomputation and structural regularization. Consequently, the improvements cannot be attributed solely to learned credit assignment. Stronger comparisons with alternating projected-gradient, Adam, or quasi-Newton optimization under matched regularization and dose-engine evaluation budgets are needed. Moreover, the policy agents are trained separately for each patient, leaving the advantages of learning over patient-specific numerical optimization insufficiently established.

Questions for authors
  1. How can the authors explain the substantial discrepancy between Eclipse and PyDoseRT for identical reference deliveries, particularly the increase in lung PTV D0.03cc from 63.38 to 112.77 Gy? Can they provide independent validation using a clinically commissioned dose engine or Monte Carlo calculations for representative PAC-VMAT plans? This evidence is essential for assessing the reliability of the paper's clinical conclusions.

  2. In Table 2, the reported mean larynx and brachial plexus doses exceed the listed clinical limits. What proportion of generated plans satisfy all relevant clinical constraints, and how do these rates compare with the commercial reference plans? Additionally, what validation beyond the configured machine-parameter constraints supports the claim of executable treatment delivery?

  3. Can the authors compare PAC-VMAT with stronger non-learning optimizers using the same sequential Leaf–MU update structure, geometric regularization, initialization, and dose-engine evaluation budget? This would help determine whether the performance gains arise from learned policy parameterization or from better-structured gradient optimization.

  4. The reported 4.90-minute runtime is estimated for approximately 20 refinement steps, whereas the final optimization comparison uses 70 steps. Can the authors report measured end-to-end optimization times and matched clinical outcomes at both 20 and 70 steps?

8 comments
AuthorsOpenReview ↗

Author Response (W1, Q1), Part 1: Dose-Engine Discrepancy and Independent Validation

We thank the reviewer for the detailed assessment and respond to each point below.

W1 & Q1: Dose-Engine Discrepancy and Independent Validation

1. Cross-engine transfer and coverage rescaling. The two values come from different evaluation conditions. The Eclipse row in Table 1 evaluates the GDP-HMM reference plans in their original commercial environment [1], whereas the Proprietary plans row recalculates the same Leaf and MU trajectories in PyDoseRT without re-optimization. This cross-engine transfer lowers unscaled target coverage, and rescaling the dose by 60/D9560/D_{95} to match target coverage amplifies the hotspot to 112.77 Gy. PAC-VMAT and all primary baselines are optimized and evaluated within PyDoseRT, a commissioned dose engine validated against physical measurements and 181 clinical VMAT plans [2].

2. Independent dose recalculation. We further recalculate the GDP-HMM and PAC-VMAT trajectories with three independent open-source dose engines, using the same CT volumes, structure masks, and machine parameters, following independent calculation-based dose verification practice for IMRT and VMAT [3].

Two engines use Monte Carlo transport, the reference approach for heterogeneous anatomy such as the lung [4]: ompMC, an OpenMP-parallel re-implementation of EGSnrc coupled photon–electron transport [5], and EGSnrc DOSXYZnrc [6, 7], with each arc simulated through its synchronized phase-space source [8]. The third engine, pyRadPlan, is the Python successor of the open-source planning toolkit matRad [9] and computes photon dose with the singular-value-decomposition pencil-beam algorithm [10]. Tab. R1 reports unscaled target coverage alongside dosimetric endpoints evaluated at matched target coverage.

Tab. R1: Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(a) GDP-HMM reference delivery

MetricLimitEclipsePyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.10 ± 0.2334.24 ± 6.4346.75 ± 1.0946.26 ± 1.5447.60 ± 1.37
PTV D95 after scaling≥ 6060.00 ± 0.2360.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7263.38 ± 0.99112.77 ± 19.0867.15 ± 1.2268.51 ± 2.7281.24 ± 6.81
Lung–CTV V20 ↓≤ 3720.64 ± 10.6426.48 ± 15.2420.16 ± 10.6220.31 ± 10.5921.33 ± 11.12
Lung–CTV mean ↓≤ 2011.79 ± 5.0114.79 ± 7.3611.53 ± 5.0511.59 ± 5.0312.21 ± 5.41
Spinal cord D0.03cc ↓≤ 50.541.61 ± 26.0937.08 ± 20.6327.88 ± 12.6628.31 ± 12.6730.01 ± 13.19
Esophagus mean ↓< 3419.73 ± 4.2022.42 ± 14.9314.67 ± 8.9214.85 ± 8.9915.52 ± 9.25
Heart D66.7% ↓< 452.44 ± 2.752.81 ± 2.742.27 ± 2.702.22 ± 2.712.69 ± 2.95
AuthorsOpenReview ↗

Author Response (W1, Q1), Part 2: Dose-Engine Discrepancy and Independent Validation

W1 & Q1 (continued): Dose-Engine Discrepancy and Independent Validation

Tab. R1 (continued): Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(b) PAC-VMAT delivery

MetricLimitPyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.00 ± 0.0052.93 ± 4.0252.49 ± 4.6155.64 ± 3.42
PTV D95 after scaling≥ 6060.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7264.12 ± 2.7372.42 ± 3.6373.35 ± 4.3475.00 ± 5.56
Lung–CTV V20 ↓≤ 3719.20 ± 9.0520.19 ± 8.6920.27 ± 8.6620.21 ± 8.70
Lung–CTV mean ↓≤ 2010.39 ± 4.6312.05 ± 5.2512.07 ± 5.2012.27 ± 5.51
Spinal cord D0.03cc ↓≤ 50.536.37 ± 14.1344.81 ± 19.2845.30 ± 19.3344.32 ± 18.95
Esophagus mean ↓< 3415.16 ± 9.4915.53 ± 9.8915.69 ± 9.9815.93 ± 10.00
Heart D66.7% ↓< 451.79 ± 2.812.26 ± 3.902.18 ± 3.862.64 ± 4.03

Independent recalculation with ompMC, EGSnrc, and pyRadPlan shows that PAC-VMAT maintains plan quality close to the commercial reference across dose engines, with lung, esophageal, heart, and spinal-cord endpoints remaining aligned with the Eclipse reference.

Takeaway: The 112.77 Gy value reflects cross-engine transfer combined with coverage rescaling, whereas independent recalculations in ompMC, EGSnrc, and pyRadPlan confirm that PAC-VMAT preserves competitive target homogeneity and organ sparing across dose engines.

[1] Gao et al., “Automating RT Planning at Scale: High Quality Data for AI Training,” 2025.
[2] Simkó et al., “A Physics-Informed, Plug-and-Play Dose Engine for Gradient-Based Radiotherapy Treatment Planning,” 2025.
[3] Zhu et al., “Report of AAPM Task Group 219 on Independent Calculation-Based Dose/MU Verification for IMRT,” 2021.
[4] Chetty et al., “Report of the AAPM Task Group No. 105: Issues Associated with Clinical Implementation of Monte Carlo-Based Photon and Electron External Beam Treatment Planning,” 2007.
[5] Doerner and Caprile, “Technical Note: Parallel Implementation of the EGSnrc Monte Carlo Simulation of Ionizing Radiation Transport Using OpenMP,” 2017.
[6] Kawrakow, “Accurate Condensed History Monte Carlo Simulation of Electron Transport. I. EGSnrc, the New EGS4 Version,” 2000.
[7] Kawrakow and Walters, “Efficient Photon Beam Dose Calculations Using DOSXYZnrc with BEAMnrc,” 2006.
[8] Lobo and Popescu, “Two New DOSXYZnrc Sources for 4D Monte Carlo Simulations of Continuously Variable Beam Configurations, with Applications to RapidArc, VMAT, TomoTherapy and CyberKnife,” 2010.
[9] Wieser et al., “Development of the Open-Source Dose Calculation and Optimization Toolkit matRad,” 2017.
[10] Bortfeld et al., “Decomposition of Pencil Beam Kernels for Fast Dose Calculations in Three-Dimensional Treatment Planning,” 1993.

AuthorsOpenReview ↗

Author Response (W1, Q2): Clinical and Mechanical Validation

W1 & Q2: Clinical and Mechanical Validation

1. Clinical criterion satisfaction. Tab. R2 separates plan-level clinical criterion satisfaction from mechanical feasibility. In the head-and-neck cohort, proximity and overlap between PTV70 and the larynx or brachial plexus prevent both PAC-VMAT and the commercial Eclipse reference from satisfying every single maximum-dose criterion simultaneously, even though spinal-cord sparing is consistently maintained.

Tab. R2: Clinical and Mechanical Constraint Satisfaction

CohortNEclipse ClinicalPAC-VMAT ClinicalPAC-VMAT Mechanical
Lung VMAT7067.1% (47/70)94.3% (66/70)100% (70/70)
Head-and-neck670% (0/67)0% (0/67)100% (67/67)

2. Mechanical feasibility. All PAC-VMAT trajectories satisfy the complete set of physical travel, jaw containment, leaf separation, adjacent motion, and meterset constraints in Table 13 and Appendix F.3 across both cohorts, establishing 100% mechanical feasibility under the configured machine limits.

Takeaway: Clinical criterion satisfaction and mechanical deliverability capture distinct requirements. Neither Eclipse nor PAC-VMAT satisfies every head-and-neck maximum-dose limit simultaneously, while all generated PAC-VMAT trajectories satisfy the configured mechanical delivery constraints.

Reviewer_evm1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PAC-VMAT, a method for automated radiotherapy planning. The authors focus on the credit assignment problem in existing reinforcement learning approaches, where a global dose-based reward makes it difficult to identify the contribution of individual leaf or MU adjustments. To address this, the method uses a differentiable dose engine to calculate gradients of the clinical objective with respect to machine parameters. These gradients are then used as learning targets for two policy networks that optimize leaf positions and MU separately. The method also includes anatomy-guided initialization and mechanical constraints. Experiments on 70 lung cases and 67 head-and-neck cases show improvements over the compared reinforcement learning baselines and direct optimization in several dosimetric metrics, with relatively short optimization times.

Critical strengths

1.I think the most interesting part of this work is using gradients of clinical objectives to directly guide machine parameter optimization. Compared with using a single global reward, this provides more specific information for individual leaf and MU adjustments. The authors also use policy networks and structural regularization instead of directly applying gradient updates. This is a reasonable design considering the dependencies between machine parameters in VMAT planning.

2.The experiments cover both lung and head-and-neck cases, which involve different anatomical structures and dose constraints. In Table 3, PAC-VMAT achieves better results than scalar-reward and direct clinical-objective training on several organ-at-risk metrics. Figure 5 also suggests that the policy-based method can reach good results with fewer optimization steps. These results provide encouraging evidence for the effectiveness of the method within the current dose calculation environment.

Critical weaknesses
  1. Table 1 and Appendix D.5 show a large difference in target hotspot doses between Eclipse and PyDoseRT for the same reference plans (63.38 Gy vs. 112.77 Gy). Although the authors use PyDoseRT consistently to evaluate different methods, this discrepancy raises concerns about the clinical comparability of the results. In particular, PAC-VMAT plans have not been independently evaluated using another dose engine. I think stronger validation is needed to support the claimed comparable performance with commercial systems.

  2. The proposed credit targets are derived from gradients of the clinical objective, and the policy networks are trained separately for each patient. Therefore, it is unclear whether the improvements mainly come from policy learning or structured gradient optimization. Although Section 4.4 includes a comparison with direct optimization, a more controlled baseline using the same gradients, update order, and mechanical constraints would help clarify the contribution of the policy networks.

Questions for authors

Q1. Dose calculation consistency Could the authors explain the large dose discrepancy between Eclipse and PyDoseRT and provide independent dose recalculation results for PAC-VMAT plans?

Q2. Contribution of policy learning Could the authors compare PAC-VMAT with a regularized gradient-based baseline using the same gradients, update order, mechanical constraints, and computational budget?

Q3. Consistency of initialization results Could the authors clarify why the Conformal arc in Table 1 and the anatomy-guided initial plan in Table 4 have different dosimetric results?

5 comments
AuthorsOpenReview ↗

Author Response (W1, Q1), Part 1: Dose-Engine Discrepancy and Independent Recalculation

We thank the reviewer for the careful reading and respond to each point below.

W1 & Q1: Dose-Engine Discrepancy and Independent Recalculation

1. Cross-engine transfer and coverage rescaling. The two values come from different evaluation conditions. The Eclipse row in Table 1 evaluates the GDP-HMM reference plans in their original commercial environment [1], whereas the Proprietary plans row recalculates the same Leaf and MU trajectories in PyDoseRT without re-optimization. This cross-engine transfer lowers unscaled target coverage, and rescaling the dose by 60/D9560/D_{95} to match target coverage amplifies the hotspot to 112.77 Gy. PAC-VMAT and all primary baselines are optimized and evaluated within PyDoseRT, a commissioned dose engine validated against physical measurements and 181 clinical VMAT plans [2].

2. Independent dose recalculation. We further recalculate the GDP-HMM and PAC-VMAT trajectories with three independent open-source dose engines, using the same CT volumes, structure masks, and machine parameters, following independent calculation-based dose verification practice for IMRT and VMAT [3].

Two engines use Monte Carlo transport, the reference approach for heterogeneous anatomy such as the lung [4]: ompMC, an OpenMP-parallel re-implementation of EGSnrc coupled photon–electron transport [5], and EGSnrc DOSXYZnrc [6, 7], with each arc simulated through its synchronized phase-space source [8]. The third engine, pyRadPlan, is the Python successor of the open-source planning toolkit matRad [9] and computes photon dose with the singular-value-decomposition pencil-beam algorithm [10]. Tab. R1 reports unscaled target coverage alongside dosimetric endpoints evaluated at matched target coverage.

Tab. R1: Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(a) GDP-HMM reference delivery

MetricLimitEclipsePyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.10 ± 0.2334.24 ± 6.4346.75 ± 1.0946.26 ± 1.5447.60 ± 1.37
PTV D95 after scaling≥ 6060.00 ± 0.2360.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7263.38 ± 0.99112.77 ± 19.0867.15 ± 1.2268.51 ± 2.7281.24 ± 6.81
Lung–CTV V20 ↓≤ 3720.64 ± 10.6426.48 ± 15.2420.16 ± 10.6220.31 ± 10.5921.33 ± 11.12
Lung–CTV mean ↓≤ 2011.79 ± 5.0114.79 ± 7.3611.53 ± 5.0511.59 ± 5.0312.21 ± 5.41
Spinal cord D0.03cc ↓≤ 50.541.61 ± 26.0937.08 ± 20.6327.88 ± 12.6628.31 ± 12.6730.01 ± 13.19
Esophagus mean ↓< 3419.73 ± 4.2022.42 ± 14.9314.67 ± 8.9214.85 ± 8.9915.52 ± 9.25
Heart D66.7% ↓< 452.44 ± 2.752.81 ± 2.742.27 ± 2.702.22 ± 2.712.69 ± 2.95
AuthorsOpenReview ↗

Author Response (W1, Q1), Part 2: Dose-Engine Discrepancy and Independent Recalculation

W1 & Q1 (continued): Dose-Engine Discrepancy and Independent Recalculation

Tab. R1 (continued): Cross-Engine Recalculation of GDP-HMM and PAC-VMAT Deliveries Before and After Target-Coverage Scaling

(b) PAC-VMAT delivery

MetricLimitPyDoseRTompMCEGSnrcpyRadPlan
Raw PTV D95 before scaling—60.00 ± 0.0052.93 ± 4.0252.49 ± 4.6155.64 ± 3.42
PTV D95 after scaling≥ 6060.00 ± 0.0060.00 ± 0.0060.00 ± 0.0060.00 ± 0.00
PTV D0.03cc ↓≤ 7264.12 ± 2.7372.42 ± 3.6373.35 ± 4.3475.00 ± 5.56
Lung–CTV V20 ↓≤ 3719.20 ± 9.0520.19 ± 8.6920.27 ± 8.6620.21 ± 8.70
Lung–CTV mean ↓≤ 2010.39 ± 4.6312.05 ± 5.2512.07 ± 5.2012.27 ± 5.51
Spinal cord D0.03cc ↓≤ 50.536.37 ± 14.1344.81 ± 19.2845.30 ± 19.3344.32 ± 18.95
Esophagus mean ↓< 3415.16 ± 9.4915.53 ± 9.8915.69 ± 9.9815.93 ± 10.00
Heart D66.7% ↓< 451.79 ± 2.812.26 ± 3.902.18 ± 3.862.64 ± 4.03

Independent recalculation with ompMC, EGSnrc, and pyRadPlan shows that the 112.77 Gy hotspot is specific to transferring and rescaling the reference trajectories in PyDoseRT. PAC-VMAT maintains competitive target homogeneity and organ sparing across the independent dose calculations.

Takeaway: The Eclipse–PyDoseRT discrepancy reflects cross-engine transfer combined with coverage rescaling, while independent recalculations confirm that PAC-VMAT's plan quality is not specific to PyDoseRT.

[1] Gao et al., “Automating RT Planning at Scale: High Quality Data for AI Training,” 2025.
[2] Simkó et al., “A Physics-Informed, Plug-and-Play Dose Engine for Gradient-Based Radiotherapy Treatment Planning,” 2025.
[3] Zhu et al., “Report of AAPM Task Group 219 on Independent Calculation-Based Dose/MU Verification for IMRT,” 2021.
[4] Chetty et al., “Report of the AAPM Task Group No. 105: Issues Associated with Clinical Implementation of Monte Carlo-Based Photon and Electron External Beam Treatment Planning,” 2007.
[5] Doerner and Caprile, “Technical Note: Parallel Implementation of the EGSnrc Monte Carlo Simulation of Ionizing Radiation Transport Using OpenMP,” 2017.
[6] Kawrakow, “Accurate Condensed History Monte Carlo Simulation of Electron Transport. I. EGSnrc, the New EGS4 Version,” 2000.
[7] Kawrakow and Walters, “Efficient Photon Beam Dose Calculations Using DOSXYZnrc with BEAMnrc,” 2006.
[8] Lobo and Popescu, “Two New DOSXYZnrc Sources for 4D Monte Carlo Simulations of Continuously Variable Beam Configurations, with Applications to RapidArc, VMAT, TomoTherapy and CyberKnife,” 2010.
[9] Wieser et al., “Development of the Open-Source Dose Calculation and Optimization Toolkit matRad,” 2017.
[10] Bortfeld et al., “Decomposition of Pencil Beam Kernels for Fast Dose Calculations in Three-Dimensional Treatment Planning,” 1993.

AuthorsOpenReview ↗

Author Response (W2, Q2): Policy Learning versus Direct Gradient Optimizers

W2 & Q2: Policy Learning versus Direct Gradient Optimizers

1. Ablations on update schedules and credit targets. Tables 10 and 11 show that varying the Leaf–MU update order or removing the intermediate credit refresh changes dose outcomes only mildly, while Table 3 holds the policy architecture, update schedule, and target-independent regularization fixed to show that learning from explicit credit targets substantially improves upon differentiating the clinical objective directly.

2. Controlled comparison with direct optimizers. We further isolate the role of policy parameterization by replacing the Leaf and MU policy networks with projected-gradient, Adam, and L-BFGS updates while holding the initial plan, normalized credit targets, structural regularization, sequential Leaf–MU schedule, mechanical projection, and 70-step dose-evaluation budget fixed.

Tab. R2: Direct Optimizer Comparison at the Same Target Coverage

MethodPTV D0.03cc ↓Lung–CTV mean ↓Spinal cord D0.03cc ↓Esophagus mean ↓
Projected gradient67.8410.8139.4216.08
L-BFGS66.4510.7338.3615.79
Adam65.8810.6437.8215.62
PAC-VMAT64.1210.3936.3715.16

Under the same credit targets, regularization, update schedule, and dose-evaluation budget, PAC-VMAT achieves the lowest target hotspot and organ doses across all four endpoints. The gain therefore comes from learning structured updates through shared policy parameters, beyond applying the same credit information with direct numerical optimizers.

Takeaway: PAC-VMAT's advantage persists after matching the gradients, update order, regularization, mechanical projection, and computational budget, isolating the benefit of learned policy parameterization.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their comments. We have revised the paper to include the additional experiments and clarify the evaluation settings. Changes are marked in blue.

Main text. Section 4.1 clarifies dose normalization and how the RL baselines are trained. Section 4.2 summarizes the independent dose recalculations, cross-patient baseline results, and the fraction of plans meeting all listed clinical criteria. Section 4.4 adds a comparison with projected gradient, Adam, and L-BFGS (Table 4), along with measured runtimes for 20 and 70 refinement steps. Section 4.5 clarifies the difference between the conformal-arc baseline and the anatomy-guided initialization used in the ablation.

Appendices. We added the full results and supporting details to the existing appendices:

  • D.4, cross-patient baselines. Table 9 compares patient-specific and cross-patient training for both RL baselines. The text also reports the head-and-neck spinal-cord results.
  • D.6, independent dose recalculation. Table 11 reports results from ompMC, EGSnrc, and pyRadPlan, including target coverage before scaling and dose endpoints after scaling. We discuss the higher target hotspots observed under independent recalculation. Table 12 reports clinical criterion satisfaction separately from mechanical constraint satisfaction.
  • D.8, direct optimizer settings. We describe the shared initialization, credit targets, regularization, update order, and mechanical projection used for the 70-step comparison in Table 4.
  • E.5, runtime and plan quality. Table 17 reports measured runtimes and dose endpoints at 20 and 70 steps. The timings include patient initialization and dose-grid setup. The runtime entry in Appendix D.7 has also been updated.
AuthorsOpenReview ↗

Main Contributions and Summary of the Discussion

We thank all reviewers and the Area Chair for their time. We summarize what PAC-VMAT contributes to treatment planning and what the additional experiments establish about the optimizer, the resulting plans, and their validation.

Main contributions

  1. Translate patient anatomy and clinical goals directly into a treatment-machine trajectory. The clinical question is how to cover a patient's tumor while sparing nearby organs. PAC-VMAT plans the leaf openings and beam intensity at each angle under the machine's motion and intensity limits. The challenge is that the resulting dose depends on thousands of coupled decisions across the arc, so a single plan score gives little guidance about which control to change.

  2. Separate physical credit assignment from learning how to update the trajectory. A dose gradient tells us how sensitive the clinical objective is to a change in one machine control. Choosing how much to change thousands of controls at once is a different problem, because their effects on target coverage and organ dose accumulate across the arc. PAC-VMAT converts the local sensitivities into normalized credit targets and holds them fixed during each policy update. The leaf and intensity policies share parameters across control points and learn how to turn those targets into coordinated changes to the machine trajectory under its mechanical limits.

  3. Learning to coordinate the controls improves the clinical trade-off even when direct optimizers receive the same guidance. With identical credit targets, machine constraints, and a matched 70-step budget, PAC-VMAT achieved a lower target hotspot and lower doses to the lung, spinal cord, and esophagus than projected gradient, Adam, and L-BFGS at the same target coverage (Table 4). In the lung cohort, 66/70 PAC-VMAT plans met every listed dose criterion, compared with 47/70 native Eclipse reference plans. On the more challenging head-and-neck cohort, PAC-VMAT brought mean near-maximum spinal-cord dose to 30.94 Gy at matched target coverage, close to 29.91 Gy for the native Eclipse reference.

What the reviewer discussion clarified

  • Does policy learning add value beyond gradients and initialization? (evm1, exAh, uLbP) We compared the learned updates with projected gradient, Adam, and L-BFGS using the same initial plan, credit targets, regularization, update order, machine constraints, and 70-step budget. PAC-VMAT achieved the lowest target hotspot and organ doses on all four reported endpoints (Table 4). The starting-plan ablation also isolates what happens after initialization: at matched coverage, refinement reduced the mean lung target hotspot from 96.80 to 64.12 Gy (Section 4.5).
  • Do the findings hold with stronger baselines and measured runtimes? (uLbP, exAh) Cross-patient training improved both reproduced RL baselines. Even with that improvement, their mean head-and-neck spinal-cord doses were 64.35 and 70.22 Gy, compared with 30.94 Gy for PAC-VMAT (Appendix D.4). Measured lung planning times, including patient initialization and dose-grid setup, were 8.56 minutes for 20 steps and 16.16 minutes for 70 steps (Table 17). Reviewer exAh confirmed that the controlled optimizer comparisons and measured runtimes addressed these concerns.
  • How well does the dose transfer to independent engines? (evm1, exAh, uLbP) In all three independent lung recalculations, PAC-VMAT's mean target D95 before normalization was 6.18–8.04 Gy closer to the 60 Gy prescription than that of the recalculated Eclipse trajectories. After matching coverage, PAC-VMAT's mean hotspot exceeded the 72 Gy goal in all three engines. The reference had lower hotspots in ompMC and EGSnrc, while PAC-VMAT had the lower hotspot in pyRadPlan (Table 11).
  • What do clinical and mechanical feasibility mean in this study? (exAh) Neither method met every listed head-and-neck dose criterion in any of the 67 cases. Both exceeded the larynx and brachial-plexus goals on average. The individual endpoints show the gain hidden by that count: PAC-VMAT substantially reduced spinal-cord dose relative to both RL baselines while maintaining target coverage. Separately, all 137 generated trajectories satisfied the modeled machine limits (Table 12). Reviewer exAh maintained the original rating over clinical reliability. Independent recalculation and trajectory checks provide planning evidence. Verifying delivered dose on a treatment machine requires measurements.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.